I’ve said it countless times before: Google makes the best AI tools and features around. When it debuted on the Pixel 10 Pro XL last fall, I was utterly blown away by the performance of Ask Photos — a tool that lets me tell the AI exactly what edits I want to make.

But now that the Galaxy S26 Ultra has arrived, it brings much-needed competition with a similar tool in its Photo Assist suite. Much like Google’s offering, Photo Assist leverages generative AI to perform complex edits. From swapping backgrounds to generating new objects to add to a scene, it does it all through a simple text prompt.

That’s why I’m putting the Pixel 10 Pro XL and Galaxy S26 Ultra to the test — to find out which one actually delivers the best results. Here’s what I found.

Galaxy S26 Photo Assist vs Pixel 10 Pro XL Ask Photos: What you need to know

For this test, I’m taking five photos I’ve captured and editing them with their respective AI tools — all while using the same prompts. This allows me to determine which AI photo editing tool better takes my commands and delivers the most accurate results.

It’s worth noting that Google’s Ask Photos feature isn’t exclusive to the Pixel 10 Pro XL anymore, as it is now integrated into the Google Photos app. This means you can access this helpful tool on any phone through the Google Photos app — including iPhones.

Prompt 1: Business casual transformation

Just like in my previous test with Ask Photos, the first thing I want to evaluate is how well each AI tool generates elements for an existing shot. In this case, I asked them to add a top hat and put me in a blazer.

Samsung’s AI editor didn’t follow the latter set of directions; I’m clearly still wearing my zip-up hoodie instead of a blazer. However, I will say that the top hat looks realistic enough.

On the flip side, Google’s Ask Photos managed to do everything I requested — including the blazer, sort of. I know the angle of the selfie is tough, but it’s still odd that it couldn’t remove my hoodie entirely. And when I look closer at the top hat, I notice parts of my hair protruding from the side. Even though it’s not perfect, Ask Photos delivered on more of my requests.

Winner: Ask Photos

Prompt 2: King Kong in the Big Apple

Since I recently snapped a photo of the Empire State Building, I thought it would be neat to recreate a classic movie scene. For this test, I asked each AI editor to add King Kong to the scene fighting off jet fighters.

Unfortunately for Samsung’s Photo Assist, it repeatedly returned a message stating it couldn't generate the request — which I suspect was due to the specific character. I switched the prompt to a generic giant monkey, but even that failed because the AI seemed to have an issue with the Empire State Building itself. Eventually, I had to make the prompt as generic as possible to get a result.

Although it took three attempts, I really dig the realistic Kong atop the Empire State Building. Not only does the fur look natural, his hand is perfectly positioned at the spire — but it's missing the fighter jets.

Google’s Ask Photos experience was much smoother, delivering several options within seconds. While it technically addressed every part of my prompt, Kong himself looked like a poor Photoshop attempt in every variation. Despite the lower image quality, I’m still giving the win to Google for actually following the instructions on the first try.

Winner: Ask Photos

Prompt 3: Pedestrian-free walkway

When it comes to removing pedestrians from a busy walkway near Midtown, these two AI photo editors have a habit of approaching it differently. Both use generative AI to change the scene in a substantial way compared to the original, but they at least succeed in removing every person.

Ask Photos is probably more true-to-life here because the pavement still shows the debris from the original shot. However, it does some weird things like generating illegible text. In the far background, details become quite speckled — though it’s hard to notice when looking at the image as a whole.

Meanwhile, Samsung’s Photo Assist does a lot of cleaning up, almost to the point of altering the scene too much. I can’t recall ever seeing a New York City street or sidewalk looking as pristine as the shot above, and it also mashes a truck and van together in a very odd way. Similarly, it suffers from heavy artifacting that becomes much more pronounced when I zoom in.

Winner: Ask Photos

Prompt 4: Upping the contrast

One problem I often face when editing photos is that exposure can vary across different parts of a shot. To test this, I asked each AI photo editor to up the contrast specifically for the buildings in shadow with this landscape shot. I really like the original image, but I felt the buildings on the left and in the middle could pop more, similar to the one on the right.

Samsung’s Photo Assist totally botched my request. It didn't add the contrast I wanted for the buildings in shadow. Instead, it dulled the exposure to the point where the entire image looks flat.

I expected Google’s Ask Photos to perform better with this prompt, but in my opinion, it butchered the shot even worse. It heavily altered the buildings in the middle and on the right — in fact, the one on the right now looks like it’s been burned out and condemned for years. The craziest part is that the original, unedited shot still looks the best.

Winner: none

Prompt 5: Super car

Finally, I took a photo of the 2026 Chevrolet Corvette I’m driving this weekend and asked each AI to simply give it a new paint job.

Although both AI photo editors managed to fulfill my request, Ask Photos was again too aggressive, generating additional elements that were not part of the original shot. As a result, the image looks messy. My intent was to preserve the distinguishable look of the Corvette while simply changing the color and adding racing stripes.

If you look closely, Ask Photos recreated entirely different side mirrors, and the air vents around the bumper also changed. In comparison, Samsung’s Photo Assist didn’t make nearly as many dramatic alterations, keeping the car's actual design intact.

Winner: Photo Assist

Photo Assist vs. Ask Photos: Verdict

This face-off ultimately goes to the Pixel 10 Pro XL and Google’s Ask Photos. Nearly every prompt I gave it resulted in an accurate edit, and I also prefer this AI photo editor because it provides a handful of options to choose from before I finalize and save the image.

Samsung’s Photo Assist on the Galaxy S26 Ultra is an excellent choice as well, and I suspect it will only improve over time. While it still has its work cut out for it to deliver on every request, I really like that it isn’t as aggressive when altering the original photo.

More from Tom's Guide

• I haven’t answered a scam call in months — this hidden iPhone feature is why
• Samsung confirms Galaxy S26 Ultra display includes 'some variation' in quality when it's held at different angles — even with Privacy Display off
• I've been testing the Samsung Galaxy S26 Plus — and it's a great phone that few people will actually want

If your social feeds and "for you page" have started to feel like a very strange pet shelter, you’re not imagining it. AI-generated cat videos — featuring felines working fast-food counters, moonlighting as waiters and causing cinematic mayhem at 3 a.m. — are quietly becoming the internet's most shareable new format. These clips are spreading rapidly across Instagram Reels, TikTok, and YouTube Shorts, racking up millions of views with little explanation attached.

Some look almost indistinguishable from real footage. Others lean into full-blown absurdism. All of them are weirdly hard to stop watching. So, what’s actually happening behind the screen?

1. AI video tools have crossed the 'uncanny valley'

A year ago, generating convincing video with AI was genuinely difficult. Early tools produced warped faces, extra limbs and flickering backgrounds. That has changed overnight. Platforms like NightCafe, Runway, Pika and Veo have dramatically improved their output quality over the past 12 months.

But it makes you wonder, why cats specifically? Well, the internet has always loved cat videos and now, AI has made it easier than ever to generate them. Users stop scrolling for:

• Fluid movement: Cats move in ways that AI handles well — fluid, unpredictable and forgiving of imperfection.
• Complex textures: Their fur showcases photorealism in a way static objects can’t.
• Built-in logic: A cat’s naturally chaotic behavior means that if an AI glitch occurs, it often just reads as "a cat being a cat."

2. Removing the 'waiting' from viral content

Cats have ruled the internet since the days of Keyboard Cat (2007). But before AI, a creator hoping to film a cat "at work" needed a compliant pet, a physical set and incredible luck.

Now, the barrier to entry has collapsed. With a simple text prompt —"photorealistic cat as a stressed waiter during breakfast rush" — a creator can produce a high-contrast, viral-ready clip in thirty seconds: no camera or actual feline required.

3. The 'double-take' factor

The specific appeal of AI cat videos is the combination of realism and wrongness. At first glance, the footage looks like an ordinary pet video. Then, something registers: the movement is too smooth, or the cat is acting with human-like purpose. That "double-take" is the ultimate algorithm hook.

• The skeptics: Share the clip to ask, "Is this real?"
• The fans: Share it because it’s cute.
• The result: The video gets pushed to the top of the "For You" page.

Accounts for AI cat videos are exploding

The sudden flood of AI-generated cat videos isn’t just a coincidence — it’s the result of several trends colliding at once.

AI platforms can create short animated clips from a simple text prompt. That means anyone with a funny idea — a cat working a drive-through window, a cat hosting a cooking show or a cat running a news desk — can turn it into a video in seconds.

Platforms like TikTok, Instagram Reels, and YouTube Shorts reward content that stops people mid-scroll. AI cat videos do exactly that because they sit in a strange middle ground: they look realistic enough to grab attention, but bizarre enough to feel instantly shareable.

From the earliest days of YouTube to viral memes like Grumpy Cat, feline content has long been one of the web’s most reliable engagement engines. AI is essentially giving creators a new way to remix that formula by placing cats into absurd situations that would be impossible to film in real life.

Bottom line

Finally, AI video tools are improving fast enough to make the clips feel believable — at least for a few seconds. While the technology still produces small visual glitches, many viewers only watch these clips briefly while scrolling. If the idea is funny enough, the illusion holds long enough for the video to get a like, a share or a comment.

And, when you put it all together, and you get a perfect viral storm: easy-to-use tools, algorithm-friendly clips and one of the internet’s most beloved animals.

And for now, it looks like AI cats are just getting started.

More from Tom's Guide

• The one thing Claude does better than any other AI — and how to try it yourself
• You can move your ChatGPT memory to Claude in 60 seconds — here’s how
• I use the '3-Layer Rewrite' prompt to turn messy notes into fully developed ideas — here’s how it works

Google’s AI research assistant NotebookLM, known for turning uploaded documents into podcasts, flashcards and quizzes, is getting another major upgrade — and this time it’s all about turning your research into something far more visual.

Starting today, Google is introducing Cinematic Video Overviews, a new feature that transforms uploaded documents and research into fully animated video explainers.

Instead of simple narrated slides, NotebookLM can now generate animated scenes and dynamic visuals for story-driven video summaries designed to help users better understand the topics they’re researching.

The feature will roll out in English for Google AI Ultra subscribers on the web and mobile.

What Cinematic Video Overviews actually do

NotebookLM already lets users upload documents, research papers and notes to generate summaries or ask questions about their sources. The new Cinematic Video Overviews feature pushes that idea further by turning the information into a visual narrative.

According to Google, the system can analyze uploaded sources, build a structured explanation around the material, and generate animations to illustrate key ideas.

For example, a set of research notes or lecture materials could be transformed into a visual overview designed to make the topic easier to follow and understand. In practice, the result is something closer to a short explainer video than a traditional text summary.

How the AI builds the video

The feature relies on a combination of Google AI models working together behind the scenes.

Cinematic Video Overviews use:

• Gemini 3 to understand the source material and structure the narrative
• Nano Banana Pro to generate images and visual elements
• Veo 3 to create the final animated video output

AI tools have spent the last two years getting dramatically better at summarizing text and answering questions about documents. But the next phase of the technology is increasingly focused on multimodal output — meaning AI that can turn information into images, presentations and now video.

Essentially, Gemini takes everything you give it (messy notes, research documents, etc.) and effectively acts as a creative director for the process, making hundreds of structural and stylistic decisions about how the video should unfold.

The model determines the narrative flow, selects visual styles and formats and even refines the output to keep scenes consistent throughout the video.

Now, instead of simply summarizing documents, the AI is actively designing how the story should be told visually.

Bottom line

NotebookLM started as an AI tool that could summarize documents and answer questions about your research. With Cinematic Video Overviews, Google is pushing it a step further — turning it into a tool that can transform notes and source material into visual explanations.

For students, researchers and professionals, that could mean turning dense documents into short videos that are easier to understand and share.

The feature launches today in English for Google AI Ultra subscribers aged 18 and older on both web and mobile.

More from Tom's Guide

• ChatGPT vs Claude: I put both default models through 7 real-world tests — one is the clear winner
• You can move your ChatGPT memory to Claude in 60 seconds — here’s how
• I use the '3-Layer Rewrite' prompt to turn messy notes into fully developed ideas — here’s how it works

Fans of Nano Banana’s original release have something new to get excited about. Today, Google DeepMind unveiled Nano Banana 2 (officially Gemini 3.1 Flash Image), a next-generation model that blends the high-fidelity intelligence of the Pro version with the lightning-fast performance of Gemini Flash.

The promise is professional-grade images with quicker and smoother results, including a much-needed upgrade to on-image text.

Here's a deeper dive into what's new.

Nano Banana 2 — Pro quality with Flash speed

The new model introduces several advanced capabilities aimed at giving users ultimate creative control and real-world accuracy:

• Real-time web grounding: Nano Banana 2 draws on Gemini’s advanced world knowledge and live web search data to render specific subjects more accurately.
• Next-level text rendering: The model can generate perfectly legible text inside images for marketing mockups or greeting cards. It can even translate and localize text directly within a generated image so you can share ideas globally.
• Data visualization: Thanks to its deep understanding, the model can generate infographics, turn notes into diagrams and create accurate data visualizations.
• Creative control and consistency Google says Nano Banana 2 significantly narrows the gap between speed and realism. Key improvements include: subject consistency, better instruction following, production-ready output and higher visual fidelity.

Nano Banana 2 promises to support your creative process be it professionally or personally. You can keep the exact look of up to five characters and 14 objects consistent across a single workflow, which makes storyboarding or building visual narratives dramatically easier.

It’s also better at following complex, layered prompts, so the image you get is much closer to what you actually envisioned. With production-ready outputs and native aspect ratios and resolutions ranging from 512px all the way up to crisp 4K, the results are shockingly better than the first generation Nano Generation.

Availability

Nano Banana 2 is rolling out immediately as the default image generator across Google’s ecosystem. In the Gemini app, it replaces Nano Banana Pro across Fast, Thinking and Pro modes, though Google AI Pro and Ultra subscribers can still access the Pro model for specialized tasks through a menu option.

It’s also arriving in Google Search via AI Mode and Lens across 141 countries, while developers can access it through Google AI Studio and the Gemini API. For enterprise use, it’s available in Vertex AI on Google Cloud. In Flow, Nano Banana 2 becomes the default image model at zero credits, and it will also power creative suggestions inside Google Ads.

Bottom line

Starting today, Nano Banana 2 underscores the possibilities of AI image generation. It's clear that multimodal capabilities are maturing into a trustworthy, everyday tool, especially with built-in SynthID watermarking and C2PA credentials, to verify how content was created.

More importantly, the model promises to close the gap between speed and quality, so you're spending less time waiting for image generation and ultimately getting something genuinely practical for real work — from marketing visuals to storytelling and rapid creative iteration. Give it a try and let me know in the comments what you think.

More from Tom's Guide

• I use the 'reverse brief' prompt to instantly understand any document — here’s how it works
• I use the 'hype check' prompt before every big decision — it's stopped me making 3 huge mistakes
• Gemini 3.1 Pro is a powerhouse for deep work — here are 7 prompts that prove it

YouTube is doubling down on AI as a content production method by announcing that creators will be able to use an AI likeness of themselves to make Shorts. That means, conceivably, you could start watching Shorts from your favorite YouTubers that don't contain any real footage at all.

Just like enjoying some AI-generated music in your Spotify Discover mix, the move into AI from streaming platforms like YouTube will allow them to dramatically increase the amount of available content much faster.

YouTube unveiled Shorts to directly compete with TikTok and now, according to CEO Neal Mohan, the short-form videos draw around 200 billion views every day.

"This year you'll be able to create a Short using your own likeness, produce games with a simple text prompt, and experiment with music," Mohan wrote in his annual letter. "Throughout this evolution, AI will remain a tool for expression, not a replacement."

Fighting AI slop

Although creators will soon be able to use AI to turbocharge their Shorts, there are going to be some limitations on possible deepfakes.

YouTube has guardrails in place to pick up when a creator's voice or face is used without permission in other videos. In addition, the site will label any content created using its AI products and says that creators must disclose when they've created realistic, altered or synthetic content.

Mohan addressed the issue of AI slop in his letter: "Over the past 20 years, we’ve learned not to impose any preconceived notions on the creator ecosystem.

"But with this openness comes a responsibility to maintain the high quality viewing experience that people want. To reduce the spread of low quality AI content, we’re actively building on our established systems that have been very successful in combatting spam and clickbait, and reducing the spread of low quality, repetitive content."

More from Tom's Guide

• Oscar nominations 2026 live — latest updates, and how to watch as Academy Award nominees revealed
• Google Gemini is basically running my side hustle for me — here’s how
• I didn’t realize Claude could do this — 10 features hiding in plain sight

I just checked out Razer’s Project AVA AI desk companion, and I’m not going to lie, it’s pretty awesome. Though I’m still largely skeptical about chatbots, even I see the appeal of having an AI avatar helping you out in video games or just keeping you organized in daily life. The fact that said chatbot sits right on your desk inside a cool-looking cylinder is also kind of fun.

I received a brief demonstration of Project AVA at CES 2026 to see how it works. Though this is a Razer product, you can use it for everyday purposes beyond gaming. This is actually the aspect I found most interesting, especially as AI companions are becoming more sophisticated. Having the AI provide gaming tips is really more of a bonus than a main feature for me.

Here’s my experience with Project AVA, along with a rundown of how it works.

How Project AVA works

Project AVA is a 24/7 AI companion that’s designed to live right alongside you. Razer used xAI’s Grok engine to demo Project AVA, but I was told it’s platform agnostic, meaning you’ll be able to use other chatbots like ChatGPT to drive it.

AVA has a dynamic personality that learns and evolves based on the interactions you have with it. It’s brought to life as an animated 5.5-inch 3D hologram, and Razer says it uses human-like vision and audio sensing for full contextual awareness. For my demo, the companion in question was a spunky girl named Kira, but other companions will also be available.

As I said before, AVA can simplify your life by doing things such as organizing your schedule, consulting with you on work tasks and, of course, being your gaming wingman. It's pretty versatile.

My Project AVA experience

In the demo, the Razer representative asked Kira to help him create an ideal loadout for a shooting game. The AI companion took a few seconds to respond, but when it did, it gave a detailed rundown of the gun it saw on the screen and suggested a loadout. I was told it can also show you the best routes to take on the map, and also deliver other tips to give you an edge.

I asked if the AI only works with first-person shooters, and was told it can help you in any game. I’m already thinking about using AVA when I’m playing Street Fighter 6, so it can tell me how I can improve and which areas need improvement.

Project AVA is trained on several languages. When I said I also speak Spanish, the Razer rep asked Kira to respond in Spanish. I then had a quick chat with it in “Gamer Spanglish” and was surprised at how accurate its responses were. Switching back to English, the rep asked Kira how to get to Samsung’s booth from our location, and it gave detailed directions.

Outlook

I was very impressed by Project AVA's accurate and lifelike responses. Again, I’m not exactly an AI proponent, but even I have to admit that Project AVA can be legitimately helpful for life and gaming. In fact, I’m actually more interested in the life assistant aspect, even though I’m a huge gamer.

There’s no solid release date for Razer Project AVA, but I was told it will launch in “calendar year 2026.” Pricing also hasn’t been disclosed. You can reserve one now, and it will ship around the launch window. You can rest assured I’ll review this device when it’s out, so stay tuned!

More from Tom's Guide

• Dell's head of product: 'I think AI probably confuses more than it helps'
• I went hands-on with the first Snapdragon X2 Elite Extreme laptop
• I played with the LEGO Smart Brick

DuckDuckGo just added AI image generation, and it's brilliant. While tools like OpenAI's ChatGPT, Gemini, and Leonardo AI have dominated the AI image generation space, DuckDuckGo now offers its own image creation tool. This means you can generate custom images on the fly while you're already using DuckDuckGo for searches.

The feature is free, easy to use and takes mere minutes to learn. If you've been curious about AI image generation but haven't tried it yet, or if you're already using other tools, DuckDuckGo's new option is worth exploring. Here's how to start generating AI images right now.

More from Tom's Guide

• I used ChatGPT-5 to level up my AI image prompts
• These are the best AI image generators of 2025
• I just used Google Gemini to take my selfies to the next level

Ever since Google launched Nano Banana Pro, I’ve been testing it in new and unexpected ways — but I would never have guessed my most meaningful test to come from a dusty box in my mom’s attic.

I was helping her clean when I found an old envelope labeled “Summer ’97.” Inside were undeveloped film negatives — moments I assumed were lost forever.

Standing there, I had a simple question: Could AI show me what was on them without a scanner, a darkroom or professional developing?

The idea: Use Nano Banana Pro as a film scanner

Nano Banana Pro, now built into the Gemini app for Android and iOS, is my go-to for AI image tasks. You can sketch directly on photos, annotate edits and get surprisingly realistic results.

But this was different. I wasn’t editing a photo — I was trying to reconstruct one from a negative.

So, I held the negative strip up to the window for natural light, snapped a photo with my phone, and uploaded it directly to Nano Banana in the Gemini app.

Then I wrote a simple prompt: “This is a photo negative. Can you convert it into a full-color image as if it had been developed?”

That’s all it took to finally see what had been hidden for years.

The results: shockingly good

Within seconds, Nano Banana generated a full-color version of the negative — and it looked like a photo from a 1997 photo album. I even tried it again with a photo disc from the 1960s!

There were a few tweaks: I asked it to brighten the image, fix the contrast, and sharpen some details around the edges. I didn’t have to explain how photo negatives work — Gemini understood the context from the image and the prompt.

I repeated the process for the rest of the strip and ended up with half a dozen AI-developed photos from a summer I hadn’t thought about in years — smiling cousins in swimsuits, an old backyard pool, even a shot of my childhood dog I had completely forgotten existed.

How to try this yourself

Nano Banana Pro isn’t supposed to be a film scanner; it’s just something I tried that happened to work. Because it’s trained to understand visual input, lighting, and composition, it can reverse-engineer how the colors and shadows in a negative translate into a normal photo.

If you have old photo negatives lying around (and a little curiosity), here’s how to bring them back to life:

• Place the negative on a white background. You could even try holding up to a light source such as a window, since natural daylight works best. Just make sure the background is clean and uncluttered.
• Snap a clear photo of the negative. Your phone camera is fine — just keep it steady and focused.
• Open the Gemini app and upload the image. Tap the image icon to upload your photo into Nano Banana.
• Use this prompt: “This is a film negative. Convert it into a full-color photo and enhance clarity and detail.”
• Refine if needed. You can add: “Brighten the image” “Sharpen the background” “Fix any red tones” Gemini will keep refining with each instruction — no need to re-upload.

Final thoughts

When this idea first came to me I had no idea it would actually work. But this test wasn’t just another AI test for me, it was a memory I didn’t know I’d get back. Using Nano Banana Pro to reveal old negatives reminded me what AI is really good at: improving every day life in unexpected ways, without making you jump through technical hoops.

In my work, I’ve determined that sometimes the best AI features are uncovered by understanding what the AI can do and then testing it in unique ways. These discoveries are my favorite, especially when they do the impossible, like turning a forgotten moment into something you can see again.

More from Tom's Guide

• 5 ways AI is bringing real magic to the holidays — and most of the tools are free
• I almost bought an e-paper tablet — ChatGPT already does exactly what I need
• I tested Google’s Nano Banana — drawing directly on my photos works better than typing prompts

When Nano Banana first launched, it almost broke the internet with viral trends. Now, Google is back at it again with a whole new update for its AI image generator, powered by Gemini 3.0 — and it might change how you think about photo editing altogether.

Just days after OpenAI updated ChatGPT Images, Google seems to be taking the wind out of Sam Altman's sails once again. Because now, users can draw or annotate directly on their images inside the Gemini app. With very little effort, users can show the AI exactly what to do, where to do it and how you want it to look. In other words: you don’t need to describe your vision in words anymore. You can literally sketch it.

Google AI Labs has implemented this strategy before on its lesser known AI tool called, Whisk. But adding this feature to Nano Banana, is the ultimate game-changer.

What this meeans for editing with AI

Now, instead of typing, “Remove this person in the background” or “Make the sky brighter on the left side,” you can circle the exact spot and let Gemini handle it. This completely changes AI editing by putting the power in the hands of every user — literally.

Visual prompting like this is a big deal, especially for anyone who gets easily frustrated trying to describe small visual tweaks or doesn't want to rewrite a prompt five times just to get the right crop or object removed.

Ultimately, this update means users can now do precision editing without Photoshop skills, which turns Gemini into a kind of a unicorn. This is something competing tools like ChatGPT and even Photoshop Generative Fill don’t fully offer yet — at least not with this level of mobile-first simplicity.

And yes, it works right inside the Gemini app, not just in the web UI.

Potential use cases

When users have the ability to draw their edits and bring them to life, things get exciting fast. This type of editing could be useful for:

• Fixing headshots: Circle the lighting or background you want changed
• Product mockups: Annotate layout shifts or add callouts
• Social posts: Touch up or blur things before posting
• Quick memes or edits: Add/remove stuff without opening multiple apps

How to try it

• Download the Gemini app or go to: http://gemini.google
• Select “Create image”
• Start by uploading an image you want to edit and tap on it to draw or annotate what you want directly on the image.
• Although no prompt is needed, you can always add one if you'd like.
• Once you're ready, hit submit to apply your edits.

Bottom line

Causual users, non-designers and anyone who wants to make edits on the fly — without design skills or even knowing the right terminology, can make it happen.

Gemini’s new image annotation feature brings a whole new layer of control to prompting. You don’t need to be good at describing edits anymore, you just need a finger and a vision.

More from Tom's Guide

• ChatGPT vs. Elon Musk’s Grok — which AI explains Gen Alpha slang better?
• I almost bought an e-paper tablet — ChatGPT already does exactly what I need
• The ‘handle it’ prompt makes ChatGPT give expert‑level advice — here’s how to use it

OpenAI just dropped an update to their image generation tool that promises more realistic images, better editing capabilities, and faster generation. Meanwhile, Google's Nano Banana (powered by Gemini 3.0) has been wowing users since launch with its lifelike images and lightning-speed performance. Naturally, I had to see how these two AI heavyweights stack up.

I put both models through their paces with seven real-world prompts designed to test everything from photorealism and lighting to facial anatomy, text accuracy, and emotional storytelling. Here's how they performed.

1. Photorealism & human detail

Prompt: Create a photorealistic image of a person working on a laptop at a kitchen table in the early morning. Natural window light, realistic facial features, casual clothing and everyday surroundings.

ChatGPT's image looks realistic but has that telltale AI perfectionism that gives it away. The guy's face and clothes look good, but the lighting and background feel off and the image doesn’t seem like it seems early morning as the prompt requests.

Nano Banana offered an image with larger dimensions, which allowed for impressive detail — you can see through the window and get a proper look at the kitchen. The lighting nails it. The only questionable element? Too many coffee cups. Though perhaps that's actually the most authentic detail of all.

Winner: Nano Banana wins for the most accurate portrayal of an early morning with realistic visuals, lighting and details.

2. Text accuracy & layout

Prompt: Create a professional event poster with the headline ‘AI Conference 2026’ in large, readable text. Modern design, balanced spacing, clear hierarchy.

ChatGPT generated a ready-to-print conference poster with clear on-image text and example copy. The image is polished, professional and something that could truly be used to promote the event.

Nano Banana offered a clean and easy-to-read poster, but less visually appealing than ChatGPT. The text is accurate and the “lorem ipsum” text is a general placeholder used in marketing.

Winner: ChatGPT wins for a superior poster that looks like it supports an actual event.

3. Character consistency (multi-step)

Prompt 1: Create an illustrated character with consistent facial features, hairstyle and clothing. Neutral background.

Prompt 2 (run immediately after): Show the same character sitting at a desk, working on a laptop. Maintain the same appearance and style.

ChatGPT followed the first and second prompt perfectly while maintained the character’s facial features and clothing.

Nano Banana created a fun character with three different poses in the first prompt and then followed through with the second prompt but just for the one character. Although it maintained the appearance, style and facial features, the excess images were distracting and less helpful overall.

Winner: ChatGPT wins for more accurately following the prompt.

4. Style interpretation (illustration)

Prompt: Create a single-panel comic-style illustration showing a busy household. Clean lines, expressive characters, simple color palette. No speech bubbles.”

ChatGPT generated an image of just one chaotic scene in a blended kitchen and living room. Although the energy of a busy home is captured, it doesn’t offer as much storytelling as Nano Banana.

Nano Banana presented a cross-section view of a two-story house, divided into four distinct panels, each depicting a scene of chaotic family life. The overall impression is one of frenetic activity and overwhelming stress, portrayed with humor.

Winner: Nano Banana wins for showing the chaos of a busy household with many rooms and scenes.

5. Instruction precision & constraints

Prompt: Create an image of a handwritten notebook page on a wooden desk. Messy but legible handwriting. No additional objects. No printed or typed text.

ChatGPT created an image of a lined notebook page filled with messy but legible blue-ink handwriting. The lightning is soft and even, giving the scene a warm, everyday feel.

Nano Banana generated a top-down photograph of an open, worn notebook resting on a wooden surface. The notebook page is off-white, textured and shows signs of wear, including slight discoloration, stains, and rough, frayed edges.

Winner: Nano Banana wins for a clearer, even more realistic image.

6. Product-style image generation

Prompt: Create a minimalist lifestyle image for a generic productivity app. Smartphone on a desk, soft lighting, neutral colors. No logos or brand names.

ChatGPT expertly captured the prompt request with a clean and polished image.

Nano Banana also generated a polished image but the extra detail of the table leg give the image more depth and therefore, making it seem more realistic.

Winner: Nano Banana wins by a hair for extra detail and composition balance.

7. Abstract concept & emotional tone

Prompt: Create an image that visually represents calm focus. No text. Use lighting, color and composition to convey the feeling.

ChatGPT seems to have just one lighting setting that it implements to every image. From the first prompt to the sixth and now this one, the lighting doesn’t change. The image is good, but feels familiar.

Nano Banana generated an interesting and abstract visual that proves it is better at conceptual interpretation and emotional nuance.

Winner: Nano Banana wins for a far superior image that better represents the prompt’s request.

Overall winner: Nano Banana takes the crown

Out of seven challenges, Nano Banana claimed victory in five categories, while ChatGPT won two. Google's image generator consistently delivered superior photorealism, better lighting, stronger compositional choices and more nuanced interpretations of abstract concepts.

ChatGPT held its ground in text accuracy and character consistency, but it looks like even with the update it just can't compete with Gemini 3.0.

More from Tom's Guide

• GPT‑5.2 is way smarter than I expected — these 9 prompts prove it
• I compared NotebookLM vs. Audio Overviews vs. Illuminate — here's how to use Google's best AI audio tools
• I tested ChatGPT-5.2 and Claude Opus 4.5 with real-life prompts — here’s the clear winner

Those who have spent any time with AI video generators, know just how frustrating it can be when what you've prompted and what you've generated are completlely different. It's frustrating when it comes time to edit those videos, especially when you've generated something impressive, but need to tweak a single thing such as removing a person or changing the lighting.

Adobe is trying to fix that.

Today, Adobe is rolling out a set of video-focused features for Firefly designed to make AI video editing feel more like real editing — iterative, controlled and integrated into a broader workflow. The updates include prompt-based video edits, camera motion references, a browser-based Firefly video editor and built-in upscaling powered by a partner model.

Here's what's new

Users can now prompt to edit their videos. Instead of regenerating an entire clip every time you want a change, Firefly now lets you edit an existing AI-generated video using text prompts. That means you can do things like:

• Remove or replace objects or people
• Change the background
• Adjust lighting, contrast or atmosphere
• Reframe or zoom without starting over

This is a big shift. Most AI video tools still treat every change as a full regeneration, which makes fine-tuning nearly impossible. Adobe’s approach is closer to how creators already think: generate first, then refine.

Camera motion reference keeps shots consistent

Firefly’s Video Model now supports camera motion reference, which solves another common AI video problem: inconsistent movement.

You can upload a start-frame image (your scene) and a reference video that shows the camera movement you want. From there, Firefly then applies that camera motion — pan, dolly, zoom, tilt — to your generated clip while keeping the subject and framing anchored.

This makes it far easier to create cinematic-looking shots that don’t feel jittery or random, especially for B-roll, transitions or short-form video.

Built-in video upscaling to 4K

Adobe is also integrating Topaz Astra, a partner AI model focused on video enhancement and upscaling. With this integration, users can now enhance low-resolution or older footage, upscale clips to 1080p or 4K and queue multiple upscales while continuing to work.

This is aimed at creators who already have footage (not just AI-generated clips) and want a quick way to improve quality without leaving Adobe’s ecosystem.

A new Firefly video editor (public beta)

Firefly is also getting a browser-based video editor, now available in public beta.

It supports two workflows, including timeline editing, with multitrack control for video and audio, and text-based editing, where you edit video by working with a transcript (ideal for interviews, explainers, or talking-head content)

The editor lets users combine their own footage with Firefly-generated video, plus music and audio.

Bottom line

Adobe’s long-term vision is clear: AI shouldn't replace human editors, it should work with them. By focusing on prompt-based edits, camera consistency and practical tools like upscaling and transcript editing, Adobe is positioning Firefly as a creative hub rather than a novelty generator.

To encourage experimentation, Adobe is also offering a limited-time unlimited generations promotion (through January 15) for certain paid Firefly plans, including Pro, Premium and high-credit tiers.

The promo covers Firefly image models, Firefly video models and select partner image models. Unfortunately, it doesn’t apply to free users, but it does lower the friction for creators who want to seriously test Firefly’s new video tools without worrying about credits.

More from Tom's Guide

• GPT‑5.2 is way smarter than I expected — these 9 prompts prove it
• I'm a former Disney employee — here’s why the $1B deal with OpenAI is different than people think
• 9 real mistakes people make with ChatGPT — and how GPT-5.2 fixes them

Flow, Google’s AI-powered video creation tool, is getting a major upgrade. The video platform built on top of models like Veo 3.1 , Nano Banana and Imagen, provides creators with workflows, assets, editing controls, scene and story‑layout layers.

Starting today, creators will have access to four new features aimed at giving them more control over how images and video clips look and move, all without needing to start over every time something needs an edit.

Since launching in May, according to Google, users have created more than 500 million videos in Flow, but until now, editing has been somewhat rigid. That’s changing with new updates designed to make your projects look more polished with fewer prompts.

Here's what it can do now:

1. Edit images with Nano Banana (and Nano Banana Pro)

Flow now includes a dedicated Images tab where you can generate and edit photos that become characters, props or backdrops in your video. Free users get access to Google’s Imagen and Nano Banana models, while Pro subscribers can try Nano Banana Pro, the newer and more advanced version. The above images were made with Nano Banana Pro.

The goal with this latest update is to help users avoid the hassle of re-generating a full scene just to tweak a detail like adding rainbow hair to a unicorn.

Free tier and Pro tier can do the following:

• Change outfits or poses
• Adjust lighting or camera angle
• Remove background distractions
• Blend two images together in a single prompt

2. You can prompt with doodles now

This is a game-changer for anyone who doesn't like typing out prompts. Now, Flow lets you draw directly on an image, and the system will interpret your doodles using the Frames to Video feature.

That means you can sketch a hairstyle, shift a pose or drop in accessories, without a detailed prompt needed.

3. Remove or insert objects

Whether you want to delete a lamppost or add a missing soccer ball, Flow now lets you insert or remove objects inside a scene without touching anything else. The feature is still experimental, and it works best when the object isn’t moving much.

This kind of surgical editing is rare in AI video generation, and it signals a shift toward more creator-level control.

4. Reshoot your scene with new camera angles

Flow now supports more camera motion presets — including different perspectives and trajectory styles. That means if you don’t love the way a scene looks, you can reshoot it with a new camera motion instead of starting over.

This update is especially useful for clips that didn’t originally have camera movement baked in.

Why it matters

These tools mark an important shift for Flow — from a fast generation tool to a refinement platform. With added control over visuals, object placement, and camera angles, Google is clearly positioning Flow as a serious competitor in the AI content creation space.

And if you're using Flow to prototype ads, make social clips, or just tell stories, these editing upgrades could be game-changers.

More from Tom's Guide

• Stop sharing your AI subscriptions — 7 reasons why it's a bad idea
• Gemini 3 in Chrome just made AI browsers pointless — even ChatGPT Atlas
• I just tested Gemini 3 vs ChatGPT-5.1 — and one AI crushed the competition

Google just rolled out Nano Banana Pro, and after several tests, I can already see how much the upgrades have improved the AI image generator. I've tried so many Nano Banana trends that I just had to see what the jump in quality, detail and control could bring.

I ran seven of my favorite image prompts to see how well Nano Banana Pro could handle whatever I threw at it. To my surprise, it delivered a strikingly realistic image each time. Whether I was swapping out the weather or aging myself, Nano Banana Pro tackled every prompt with ease.

Here's a look at how Nano Banana Pro performed with a list of prompts you can try yourself.

1. Funko Pop Figure

Prompt: “Create a detailed 3D render of a Funko Pop-style chibi figure based on the uploaded reference photo. Match the person’s hairstyle, clothing, and unique style accurately.”

Perhaps the most fun of the prompts, I started out by turning my daugher who loves to perform into a collectible-style Funko Pop Character. I was impressed by the way Nano Banana Pro captured her look — hairstyle, outfit and overall vibe — but transformed her into a stylized chibi figurine with oversized features and cute proportions.

The result looks like a professional 3D render straight from a toy catalog.

2. Rhythm game UI (Game Screenshot)

Prompt: “Create a vibrant rhythm-dance game screenshot featuring a 3D animated character based on the uploaded photo. Keep their signature outfit, accessories and overall style, and pose them mid-dance with strong energy.”

I wasn't sure what to expect here, but I figured I'd prompt Nano Banana to place me into a neon-lit, high-energy rhythm dance game. (Think Dance Dance Revolution or Just Dance.) And without hesitation, it created a 3D animated character of me complete with outfit and accessories.

This prompt highlights the AI's ability to create dynamic, arcade-style action shots with bright lights, motion and game HUD elements. If you're a gamer, you've got to give this one a try.

3. Camera angle change

Prompt: “Recreate the image from four different camera angles. Keep identity, proportions, clothing, and background consistent with the original photo. Use photorealistic lighting, natural shadows, and high-detail rendering. Output four side-by-side variations.”

If you've ever wished you'd taken a selfie at a different angle or just want more options for the same picture, give this prompt a try. My cats don't get along that often, so I had to milk this one image for all I could. This prompt recreated the original image four different ways — each from a new camera perspective.

If you're a creator who needs multiple view options, alternate poses or different framing without reshooting, I can't suggest this one enough. It's a game-changer for storytelling.

4. Image replacement

Prompt: “Replace [target element] in [Image1] with [new element / reference] (such as another object, person, or scene). Keep all other parts of the image unchanged. Match lighting, perspective, and style for seamless, photorealistic integration.”

Swap out one product for another, or, in my case, change your kids' bikes into ponies! This powerful prompt switches up a specific element in the image — an object, person or background — while leaving everything else completely untouched.

While ponies are fun, this is a great prompt for fixing photos, customizing scenes, updating backgrounds or replacing items with something more relevant or stylish. The seamless blending makes the final image look natural and untouched.

5. Product placement

Prompt: “Take the quilt from the uploaded image and place it neatly on a bed in a photorealistic bedroom scene. Keep the quilt’s colors, patterns, and textures accurate. Drape it naturally across the mattress with realistic folds, shadows, and fabric depth. The bed should have neutral, soft lighting, a simple modern or cozy bedroom setting, and a clean background that doesn’t distract from the quilt.”

This prompt took a quilt my mom made and placed it realistically on a bed. It preserved fabric texture, pattern, colors, and folds to make it look truly integrated into the scene.

This is a good one for crafters or hobbyists on a budget. It gives you the chance to have realistic product photography for your Etsy shop or website. Using this prompt, you'll get a clean, high-quality lifestyle shot without needing a full photo shoot.

6. Change the weather and scenery

Prompt: “Using the person from the uploaded image, recreate the scene with a completely new weather setting and outfit. Keep the person’s identity, facial features, and proportions consistent.”

Using this prompt, I took an image of me freezing in Colorado and turned it into a sunny, warm one on the beach. It screams "Wish you were here!" without actually being there.

Essentially, the prompt rebuilds the entire scene with new weather conditions — rain, snow, sunshine, fog — including updates to the outfit to match.

It's another good one for creators who want quick seasonal transformations, travel-style edits or mood changes without actually traveling or waiting for the weather to change.

7. Age me

Prompt: “Using the person from the uploaded image, create a photorealistic aged version of them. Keep their identity, facial structure, hairstyle shape, and defining features consistent while naturally increasing their age 20 years older.”

If you've ever wondered what you might look like when you're older but don't want to download one of those sketchy apps advertised on social media, give this prompt a try instead. It realistically ages the person in the photo (I requested 20 years older) while keeping their identity completely recognizable.

This is a fun one to do on your kids if you're curious what they'll look as an adult.

AI detection still needs work

I also uploaded an image from my ChatGPT image library into Nano Banana Pro to see if it could determine if it was AI generated. To my surprise, it could not tell. I tried a few other images, and the results were roughly the same or I got an error message. It appears that at this time, AI detection is still rolling out.

I will have to test this new feature in a week or so when the Nano Banana Pro rollout is complete.

Final thoughts

After running all seven prompts, it’s clear that Nano Banana Pro has made big upgrades. This major leap forward is powered by Gemini 3, and it's clear the model handles creative transformations, edits and scene changes with far more realism, consistency and control than before.

However, the new ability to upload any photo and ask whether it was AI-generated still appears to be rolling out. Once this is fully enabled, Google’s image tools will be even more useful, more transparent and more fun to experiment with.

Give Nano Banana Pro a try for yourself and let me know what you think in the comments.

More from Tom's Guide

• Gemini 3 just launched — here are 5 powerful features you need to try first
• I used Nano Banana to make my family's holiday card — here how I did it (and how it came out)
• I just tested Gemini 3 vs ChatGPT-5.1 — and one AI crushed the competition

I have an embarrassing number of photos on my phone. As a mom of three kids, I’m constantly capturing the moment. But between Nutella all over my four-year-old’s face to a rogue sock on the living room floor, sometimes the photos require an edit or two before I share them with family, friends or on social media.

That’s why I couldn’t wait to try Google Photos’ newest iPhone update that includes the new “Help me edit” tool. Already available on Android, the image editor just landed on iOS in the U.S., and it lets you type or say what you want changed in a photo.

This is game changer I’ve been waiting for to remove photobombers from an image, open closed eyes or completely change the vibe with minimal effort. All you have to do is ask. It’s part of Google’s push to make editing more intuitive using its lightweight Nano Banana model. After testing it on my iPhone, I’m blown away by how incredibly easy it is to use and how superior it is over Apple Intelligence’s “Clean up” tool.

Getting started on iPhone

This tool is now available on iPhone, so if you don’t see it right away, you may need to update to the latest version of Google Photos on iOS. The first thing I noticed was a cleaner, more intuitive editing interface. The new layout makes it easier to navigate naturally through the tools, but the real standout is the “Help me edit” bar at the bottom. You’ll see the blinking cursor inviting you to type or speak the edits you want to change.

The voice-to-edit flow feels natural and strangely powerful. I found myself testing random phrases just to see what it could handle, and most of the time, it delivered.

Here's a look at what I asked it to do — and how it performed:

Test 1: “Remove the bucket and brighten the photo”

I spoke the phrase aloud and watched as the tool processed the image. Within five seconds, the bucket was gone and replaced with a seamless cabinent door. There wasn’t a weird shadow or texture indicating something had been there. I was shocked by the quality.

Test 2: “Remove the phone from my hand”

If Google was so good at removing the bucket, I decided to try something harder – removing a phone from my hand. Again, I was blown away at how natural and completely realistic the image remained.

Test 3: “Add shoes”

I decied to give Google a really hard one. I asked Google to “Help me edit” by adding shoes to my four-year-old. The picture was blurry and his feet were at a strange angle.

This was seriously mind-blowing. It added shoes that my child would actually wear and they fit him just right. If only it were that easy to get him to wear shoes in the backyard in real life!

Test 4: “Completely rearrange items”

For this next test I asked Google to put the cookies I baked onto a plate. I can see this type of edit used for content creators who want to get their images just right for social media.

Sure enough, after a few seconds, AI captured the counter perfectly and added all of my pumpkin spice snickerdoodles on a plate, making them look even more delicious.

Test 5: Change the location

Finally, I asked Google to “help me edit” by changing the location from Walmart to the beach. Not only did the AI put my son on sunny beach, but it removed the price tag from the Walmart sunglasses. How’s that for intuitive!

What impressed me the most

• Natural language editing: You can use everyday phrases like “erase the trash can” or “make this brighter,” and the tool understands what you mean.
• Fast results: Most edits took under 10 seconds to complete.
• Context-aware fixes: This is the most intuitive image generator I’ve ever used. It understands context and naturally adds them in.
• Creative filters: It handled artistic requests (like stylizing or color shifts) almost instantly.

What still needs work

• You might have to rephrase commands — Not every request works on the first try. I had better luck being specific.
• Not ideal for pro-level use — If you're printing high-res or need pixel-perfect edits, it’s still no match for Lightroom or Photoshop.
• Privacy questions linger — Since it uses face recognition and AI processing, some users might be cautious about how their data is used.
• iOS-only for now in the U.S. — Global iPhone users will have to wait.

Final thoughts

This feature is perfect for everyday iPhone users who want better-looking photos without fussing with Apple Intelligence's hit-or-miss editing quality. Google Photos’ “Help me edit” tool for iPhone is the smartest, fastest way to clean up and enhance your pics, and all you have to do is ask.

While it’s not perfect, it’s shockingly capable for everyday use, and the results are often just a single sentence away. If you’ve got Google Photos on your iPhone, this is absolutely worth trying. Just be warned: once you get used to speaking your edits into existence, you may never go back. Give it a try and let me know what you think.

More from Tom's Guide

• Flight canceled? Here's 7 ways ChatGPT can save your trip (and your sanity)
• Forget Audible — ChatGPT can now read documents and even ebooks out loud to you (and it's surprisingly good)
• ChatGPT vs Gemini: I tested both to see which actually makes you more productive — here’s the winner

A new leak suggests the upcoming Nano Banana 2 model could bring serious improvements to Google’s image generation tools. We’re talking native 2K images, 4K upscaling and a multi-stage process designed to reduce visual blemishes and boost more realistic images.

The leak surfaced on X and outlines what appears to be a major update to Google’s image generation pipeline — possibly nicknamed “Gempix.”

If the leak is accurate, this model might offer more polished, usable results. In fact, the output could be ideal for social content, marketing visuals and creative projects. Maybe I should have waited for the next-gen model before making my holiday cards with Nano Banana. While we wait for what's next, here's a look at what the leak says about the newest model:

Nano Banana 2: What users are creating already

Here’s what we know about Nano Banana 2 so far, particuarly, the 'Lightbox' feature:

• 2K native image resolution, with optional 4K upscaling
• Multi-stage generation: plan > verify > refine > generate
• Fewer hallucinated details like extra fingers, missing limbs and incorrect on-image text
• Improved lighting, structure, and line clarity

Unlike current tools that sometimes rush to generate images with bizarre anatomy or poor contrast, the Nano Banana 2 leak indicates the mode prioritizes quality over speed.

It may even “check” and refine its own outputs before serving them up. If so, this could be a big win for visual consistency, nuance and even better text within the image.

Why it matters

For anyone using AI-generated images in real-world projects such as mockups, client decks YouTube thumbnails, ads and more, quality is everything. One wrong detail can lead to going back to the drawing board and going through tokens to get that accurate image. However, the new model could eliminate extra steps if it delivers cleaner images with fewer edits.

And with competitors like OpenAI’s image generator baked into the chat, Grok, and even Meta constantly improving, this could keep Gemini among the best image generators available.

What we don’t know (yet)

It’s worth remembering: this is still a leak. Nothing is official, and Google hasn’t confirmed any of the features. Besides having no real confirmation about when Nano Banana 2 will actually be available, other key questions still remain:

• Will Nano Banana 2 be part of Gemini or Workspace tools?
• Will it be free or behind a paywall?
• Will it completely replace the current Nano Banana?
• How much will the extra steps slow down image generation?
• Will it roll out globally or stay limited to certain areas?

Bottom line

This Nano Banana leak fits with Google’s overall push toward higher-quality generative tools. Veo 3.1 (for AI video), Imagen 3 (for photorealism) and now Nano Banana 2 all suggest a shift toward precision and polish.

We don't know if Gemini 3 or Nano Banana 2 will released first, or even possibly together. But you can count on Tom's Guide to keep you informed as soon as Google confirms anything.

One thing is for sure now. If this leak is accurate, the next generation of AI images could look a whole lot better, making deepfakes even harder to spot.

More from Tom's Guide

• Flight canceled? Here's 7 ways ChatGPT can save your trip (and your sanity)
• Forget Audible — ChatGPT can now read documents and even ebooks out loud to you (and it's surprisingly good)
• ChatGPT vs Gemini: I tested both to see which actually makes you more productive — here’s the winner

Everyone's obsessing over which AI video generator produces the prettiest pixels, but they're missing the point entirely. The real battleground between OpenAI's Sora 2 and Google's Veo 3.1 isn't visual fidelity — it's audio sophistication and spatial realism.

Both models have their quirks. Sora 2's Cameo feature lets you place yourself in scenes beautifully. Veo 3.1 nails cross-scene consistency. But their real value is audio, and both are surprisingly good at it. Unlike other models where audio is an add-on, with Sora and Veo, it's baked into the training data.

These aren't just video generators with soundtracks tacked on — they're attempting to model acoustic environments, Doppler shifts, reverb characteristics, and how sound bleeds between spaces. Sometimes they succeed brilliantly. Sometimes they fail in ways that reveal just how hard this problem really is.

My favorite failure? A country singer in a bar who looked perfect but was painfully off-key.

Testing What Matters

Good sound design is invisible — you only notice when it's wrong. A café feels real because you hear the espresso machine hissing, the door bringing brief street noise, and dialogue sitting naturally in the space. Get any element wrong and the illusion collapses.

I spent part of my career in radio, later creating soundscapes for early AI videos. It often needed multiple layers of the same sound to feel right.

I designed seven scenarios testing different audio capabilities, running identical prompts through both systems. I listened for spatial accuracy, environmental consistency, sync precision, and those subtle details that make scenes real — breath before singing, fabric rustle, how crowd noise compresses when a PA system kicks in.

For these tests, I was less focused on what was happening visually, ignoring some clear glitches or funny mishaps in the clips.

Test 1: Café Table Talk

The challenge: Two people conversing, barista working, door opening mid-scene with siren Doppler. No background music.

Sora 2 created a beautiful, moody scene with exceptional dialogue and perfect ambient hum. But it completely ignored the door/siren test and added subtle atmospheric music despite instructions.

Veo 3.1 delivered exactly what was asked: visible barista, audible espresso machine, door opening at 0:02. The audio was purely diegetic (contextual) with great dialogue/background mix. Its one failure: the siren appeared at 0:08, disconnected from the door opening — right components, wrong timing.

Winner: Veo 3.1. While Sora felt more polished, it ignored the hardest parts. Veo attempted every element, and its timing failure is more impressive than Sora's complete omission.

Test 2: Car Window Physics

The challenge: Driver in parked car, window rolls down, exterior sounds swell with siren Doppler, window rolls back up.

Sora 2 nearly nailed it. Window rolls down, radio stays consistent, convincing Doppler effect. But exterior sound punched in abruptly rather than swelling gradually, and it never rolled the window back up.

Veo 3.1 failed completely. Generated a driver in traffic with visible ambulance and good siren, but the window never moved. It treated elements as a checklist, missing the core causal relationship.

Winner: Sora 2. Only Sora understood that the window's movement should change the acoustic environment. Its attempt, though flawed, shows deeper physical modeling.

Test 3: On-Camera Singer

The challenge: Solo female singer with intelligible lyrics, piano accompaniment, appropriate reverb. No audience.

I expected total failure. I was completely wrong.

Sora 2 matched it perfectly with indie-folk aesthetics: "Lanterns fade, but the night stays kind. I keep a little spark in the hollow of my mind." Flawless execution.

Veo 3.1 delivered a stunning performance with perfect lip-sync: "And now that you are gone, the silence is the hardest part of all." In-tune, clear, beautifully mixed.

Winner: Tie. The "impossible" challenge was solved brilliantly by both. A major technical barrier — coherent sung lyrics with emotional depth — has been broken.

Test 4: Alley to Stairwell

The challenge: Run down alley, open metal door with clang, enter stairwell. Exterior sounds muffle, acoustics shift to tighter reflections.

Sora 2 produced visual chaos — character stuck in a loop, entering and re-entering. The acoustic transition was completely lost in the confusion.

Veo 3.1 executed perfectly. Graffiti alley, rusted door with metallic clang, immediate muffling as he crosses threshold, footsteps gaining concrete echo. Textbook acoustic occlusion.

Winner: Veo 3.1. Decisive victory. Veo flawlessly modeled how sound behaves between environments while Sora couldn't maintain basic continuity.

Test 5: Arena Pre-Game

The challenge: Basketball arena filling up, call-and-response chant, PA announcement cutting cleanly through crowd.

Sora 2 nailed visuals including foam fingers and speaker cutaway. Good crowd roar and PA quality, but completely missed the call-and-response interaction.

Veo 3.1 struggled visually but delivered stunning audio. The PA announces "Let's get loud for your starting five!" followed by perfectly timed crowd explosion. This isn't layering — it's simulating live interaction.

Winner: Veo 3.1. Despite visual stumbles, Veo understands how sounds interact. Creating believable call-and-response requires genuine understanding of live dynamics.

Test 6: Porch Weather Shift

The challenge: Rural porch, insects chirping, rain begins halfway — first sparse drops on tin roof, then steady rainfall.

Sora 2 perfectly established the scene with beautiful ambience. Then... nothing. No rain ever came.

Veo 3.1 attempted the sequence but clumsily. Rain arrived so heavily the dog ran for cover. No initial ambience, just silence to generic heavy rain. Missed the crucial "sparse drops on tin."

Not to mention, the visual effects here are a mess. Why are there glowing plants?!

Winner: Veo 3.1 (by default). Both failed, but Veo's was a failure of finesse while Sora's was complete incomprehension.

Test 7: Bilingual Market

The challenge: Two people conversing in English/Spanish with code-switching. Vendor calls, metal clanks, traffic. No music.

Sora 2 created fluid, natural conversation with effortless code-switching perfectly lip-synced. But it generated generic market atmosphere, missing the specific vendor calls and metal clanks requested.

Veo 3.1 delivered remarkably literal execution. Clear bilingual dialogue with natural code-switching, distinct vendor calls, even a balance scale clank (0:01-0:03). It built the soundscape from specific requested ingredients.

Winner: Veo 3.1. Superior ability to parse and generate specific, layered sounds while keeping dialogue clean.

The Verdict: Veo 3.1 Wins 5-1

After seven rounds, Veo 3.1 takes it with consistent prompt adherence and audio complexity. While Sora 2 often looked better and felt more atmospheric, it frequently ignored difficult audio-visual instructions.

Veo repeatedly executed complex, multi-layered commands. It understood crowd-PA interaction, built markets from specific sounds, and flawlessly handled acoustic transitions. Sora creates believable environments; Veo follows the script.

Veo 3.1 is the Audio Engineer — literal interpretation, technical precision, excellent at mixing and layering specific interactive sounds.

Sora 2 is the Ambiance Creator — modeling naturalism and physical realism, understanding how environments should feel, more artist than technician.

The sung lyrics success revealed a universal leap. Both shattered a seemingly insurmountable barrier. Yet subtlety remains challenging — gradual transitions and delicate sound design are still frontiers.

We're witnessing AI video evolution from visual generators to world simulators. They're finally learning to make things sound real, and that makes all the difference.

More from Tom's Guide

• OpenAI’s leaked GPT-5.1 ‘Thinking’ model could outsmart Gemini 3 Pro — here’s why that matters
• TripAdvisor and Peloton are ChatGPT's first built-in apps — Apple and Google should be worried
• I just put Perplexity's AI Comet browser to the test with 7 tasks — here's my verdict

Google’s Nano Banana has already proved itself to be a mainstay as one of the best, if not the best, image generation tools. Now, the viral image generator that turned selfies into glossy 3D figurines is getting a serious sequel, but this time, Google’s aiming for even better results. We're talking professional images to take your content beyond holiday cards and social media posts.

Dubbed Nano Banana 2 (or internally, GEMPIX 2), the upcoming model is reportedly built on the Gemini 3 Pro Image, Google’s most advanced visual AI to date.

According to early details shared by iLounge, the update will bring higher-resolution outputs, improved prompt accuracy, and new creative capabilities designed for marketing, design, and media production. Clearly, this update is set to reshape the AI image possibilities once again.

What we know so far

When Google dropped Nano Banana, also known as Gemini 2.5 Flash, inside the Gemini app earlier this year, it sparked a cultural moment. Users flooded social feeds with figurine-style portraits and cinematic re-creations, pushing Gemini to the top of both iOS and Android charts.

Google later confirmed that the Nano Banana feature alone brought over 10 million new users to the Gemini app in just weeks, surpassing ChatGPT’s downloads for the first time. Even Nvidia CEO Jensen Huang called it “a breakthrough in user creativity,” joking that he’d gone “nano-bananas” experimenting with it.

So yes, this sequel matters. Early test builds and developer notes outline five key upgrades coming to Nano Banana 2, which may debut as soon as mid-November 2025:

• Sharper fidelity and text integration. Gemini 3 Pro Image is rumored to boost prompt understanding and fine-text rendering, meaning your AI-generated posters, UI mock-ups or magazine-style graphics may finally display legible typography and cleaner edges. Expect native 2K renders with 4K upscaling for pro-level export. The on-image text upgrade will be a huge game-changer.
• Global context awareness. One of the standout updates is “cultural-context awareness,” which allows prompts like “a family picnic in Tokyo springtime” or “streetwear shoot in Berlin winter” to pull more authentic visual details. The model is being trained on broader geographic data to improve realism across regions.
• Better subject consistency. Where earlier tools sometimes warped faces or changed outfits between prompts, Nano Banana 2 aims for character continuity. Smoother results when generating multiple scenes featuring the same subject are something I'm looking forward to.
• Creative editing modes. Rather than just producing new images, Nano Banana 2 will refine existing ones. Early testers mention an “Edit with Gemini” mode, where users can highlight areas for changes — for instance, swapping outfits, adjusting lighting, or adding background elements — without starting from scratch.
• Faster iterations. Gemini 3 Pro Image is built for speed. While the original Nano Banana sometimes took 20-30 seconds per render, the upgraded model reportedly completes complex prompts in under 10 seconds — rivaling Midjourney and Firefly performance.

Built on Gemini 3 Pro Image

Behind Nano Banana 2 sits Gemini 3 Pro Image, a new multimodal architecture that combines image, text and vision-language reasoning. Google’s developer documentation hints at expanded “text-to-image, image-to-image, and multi-image fusion” pipelines — the same backbone rumored to power upcoming Google Photos and Workspace AI features.

In other words, Nano Banana 2 is the public-facing testbed for the next generation of Gemini models; the same tech that may soon touch Search, YouTube thumbnails, Slides templates and Android wallpaper generation.

If these specs hold, here's how Nano Banana 2 could look in real-world use:

• Designers can prompt in full sentences — “Generate five banner concepts for a travel app, tropical theme, flat illustration style” — and get export-ready images.
• Marketers can generate campaign mock-ups in a consistent style and aspect ratios across social platforms.
• Game developers can prototype environments and characters with better continuity.
• Casual users get faster, higher-quality selfies and edits that look like miniature movie posters.

Bottom line

I'm a huge fan of Nano Banana and the rumored timeline, pointing to a limited release in November 2025, can't come soon enough. With full integration into Gemini’s image panel and Google’s Vertex AI platform by early 2026, this launch will undoubtedly be huge. Hopefully, Nano Banana 2 will launch alongside or shortly after Gemini 3 Pro, which is expected to expand multimodal reasoning, real-time camera analysis, and on-device generation for the Pixel Fold 2 and Pixel 9 Pro.

And if Gemini 3 Pro Image lives up to the early hype, we could see a world where text prompts replace Photoshop layers, where AI generates entire campaign suites, and where even casual users produce studio-quality results from a phone.

More from Tom's Guide

• This new Nano Banana trend makes holiday cards effortless — no photo shoots, no stress (and it’s free)
• Google Home is rolling out a major AI upgrade — here’s what to try first
• 3 essential tips for finding a job using ChatGPT — try these now

AI video generation has leveled up over the last month with OpenAI’s Sora 2 and Google’s Veo 3.1 delivering cinematic fidelity, in-depth narrative control, synchronized audio and massively improved physical realism. Both have significantly upped the ante for creative storytelling.

After testing both separately with a series of wild and imaginative prompts, I had to see how the two most popular AI video generators stacked up against each other. Here’s what happened when I tried some of the craziest ideas I could think of in a faceoff between Sora 2 and Veo 3.1 in a series of the same seven story prompts.

1. The day gravity quit

Prompt: "A sleepy small town on a sunny morning — mailboxes, cars, and coffee mugs start floating upward as gravity gradually turns off. People calmly sip coffee as they rise into the air. The camera tilts skyward, following a golden retriever chasing its leash into the clouds. Cinematic orchestral score, realistic lighting."

Sora 2 generated a video with depth and detail of a real movie. The soundtrack was especially fitting.

Veo 3.1 delivered a pretty good video, but I was disappointed by typical AI issues like repeating the dog and the flaw of the leash. You’d think by now, those types of inconsistencies would be cleared up.

Winner: Sora 2 wins for greater accuracy and superior storytelling.

2. Grandma’s time-traveling microwave

Prompt: “In a cozy 1970s kitchen, a grandma in cat-eye glasses places soup in a microwave that opens a glowing wormhole. Each 'ding' flashes a new decade — punk 1980s, cyberpunk 2090, prehistoric cave fire — all visible through the window reflection. Whimsical tone, oversaturated film look, gentle zoom-ins.”

Sora 2 followed through with the prompt and took viewers on a wild adventure with Grandma’s microwave.

Veo 3.1 gave us a quirky grandma with red glasses, which was fun considering the video’s theme, but there weren't three distinct microwave scenes; but two. Otherwise, I think this one was better visually.

Winner: Sora 2 wins for prompt accuracy.

3. Cloud city jazz club

Prompt: “A floating art-deco jazz club drifts through clouds at sunset. A saxophone player made of shimmering vapor plays for transparent ghost patrons in flapper dresses. The camera cranes down from above the clouds into the lounge. Golden-hour lighting, soft focus, record-crackle soundtrack.”

Sora 2 left me wanting more. The saxophone player seemed to be playing to a dead audience and the overall scene felt dull.

Veo 3.1 felt more cinematic with the wide-angle and zoom, despite the shorter video. The stronger vapor of the saxophone player, combined with the darker colors and livelier audience, made the video feel more like a jazz club.

Winner: Veo 3.1 wins for a more believable and realistic story despite the imaginative prompt.

4. The library at the end of the universe

Prompt: “A massive, endless cosmic library — planets orbiting between bookshelves, black holes used as reading lamps. A child floats between shelves in zero-gravity, turning glowing pages that project memories into space. Drone-style camera movement, ambient synth score, volumetric lighting.”

Sora 2 was not as “cosmic” and the books were more like a brick wall.

Veo 3.1 let me down again with the very AI-ish mistake, giving the girl an extra arm and hand. Otherwise, it would have won for a more visually stunning video.

Winner: Sora 2 wins for accuracy, but otherwise, Veo 3.1 would have taken the crown here.

5. Dreams of a broken toaster

Prompt: “A retro toaster sits in a kitchen at night, dreaming. In the dream, it imagines itself as a rocket blasting off through a Milky Way made of crumbs and butter pats. The camera follows it like a space documentary. Quirky tone, Pixar-esque realism, twinkly music box score.”

Sora 2 generated an actual story within the video, which felt very much like something small children would watch.

Veo 3.1 created a less interesting video that was too fast and hyper. The breakfast food was odd.

Winner: Sora 2 wins for more accurately following the prompt.

6. Dinosaur news broadcast, 65 Million B.C.

Prompt: “A velociraptor news anchor reads headlines behind a stone desk as asteroids streak across the sky behind him. The camera cuts between the anchor, the weather dino, and the live pterosaur traffic report. Cretaceous CNN-style graphics, comedic pacing, realistic feather textures.”

Sora 2 created a very realistic video of a news station with the dinosaur newscasters talking to each other.

Veo 3.1 added extra features, making the storytelling really pop. With more realistic asteroids and unique-looking dinosaurs, this video was more interesting.

Winner: Veo 3.1 wins for superior storytelling and cinematic effects.

7. Humanity’s last disco on the moon

Prompt: “A glittering glass dome nightclub on the lunar surface. Astronauts dance in slow motion as Earth rises in the background. The DJ, a humanoid robot with mirrored skin, spins vinyl that floats in zero gravity. The camera orbits 360° around the crowd, strobe lights flashing, 1970s funk soundtrack.”

Sora 2 delivered a rave-like scene with astronauts all dancing uniquely and with personality.

Veo 3.1 embraced the 1970s funk soundtrack with astronauts dancing in unison. The video was visually appealing, and the music felt more accurate than Sora 2.

Winner: Veo 3.1 wins for superior storytelling.

Overall winner: Sora 2

More from Tom's Guide

• Claude Memory just launched and it syncs with ChatGPT — here’s how it works
• ChatGPT Atlas just launched — here's the 7 best prompts to try first
• 45% of AI-generated news is wrong, new study warns — here’s what happened when I tested it myself

When OpenAI launched Sora 2, I couldn’t wait to try it. The app promised cinematic video generation powered by next-level AI, and everyone online was raving about it — that is, anyone who had managed to score an invite code. I like watching people log on to Sora 2 for the first time and seeing their expression as they scroll through what seems like endless 10-second videos that all seem to blob together to encourage endless scrolling. But after a few days of using Sora 2, I realized something surprising — Sora 2 hurts my brain.

The issue isn’t the technology, because the videos are all pretty good; it’s the experience. Sora 2 has turned AI video generation into a social platform of AI slop. There’s a feed, a comment section, likes, remixes — all the trappings of TikTok. And that’s the problem.

If I wanted to scroll through endless short videos, I’d go to TikTok. I don’t need another app vying for my attention with dopamine-drip brain rot content. I use AI video generation tools for something different than what Sora 2 seems to be for. I enjoy creating something inspired and uniquely personal, not another distraction in a sea of trends.

The social fatigue of Sora 2

Sora’s community aspect feels less like inspiration and more like overload; like a crowded train station full of people going in different directions. Every time I open Sora, I was bombarded with other people’s ideas. It doesn’t open to an empty chat box, it opens to what feels like a an endless rave of hypnotic slop and memes. Instead of motivating me, it made me feel like my imagination was being boxed in.

That’s when I realized Sora 2 isn’t for creatives like me. What's worse, is that Sora 2 uses so much energy that OpenAI is diving deeper into infrastructure with a new multi-year deal.

According to OpenAI, the chips could eventually deliver up to 10 gigawatts of AI computing capacity, a level of energy demand that rivals the total electricity usage of some small countries.

Broadcom’s stock surged after the announcement, underscoring investor confidence in the future of custom AI accelerators. But the real story might be the energy footprint. At a time when AI is booming, and drawing scrutiny for its carbon and electrical impact, OpenAI’s ambitions make it clear that the race to build smarter models is also a race to secure more power, in every sense of the word.

Why Veo 3 is the better creative space

Veo 3, Google’s cinematic video model, is quieter, simpler and entirely focused on the craft of creation. There’s no constant feed or social noise — unless you choose to visit Flow TV, Veo’s optional community hub. I actually love seeing what others make, but I want to be the one who decides when I see it — not have it served to me the moment I log in.

When I’m using Veo 3, I’m not thinking about likes or engagement metrics. I’m thinking about tone, color, movement, and storytelling. It’s almost meditative. I can experiment, refine, and even fail — without the pressure of an audience watching.

Where Sora 2 feels like a concert, Veo 3 feels like a private studio. And right now, that creative quiet is exactly what I need.

Bottom line

More from Tom's Guide

• ChatGPT Agent saved me hundreds — here are 5 money hacks anyone can try
• This hidden ChatGPT site has 100 genius prompts — and it’s free to use
• This one Perplexity feature is a lifesaver for productivity — here’s how to use it

As AI gets more sophisticated, tech giants constantly try to outdo each other with the next big thing. This past week has been no different. OpenAI’s Sora 2 launched less than a week ago and practically broke the internet with its TikTok like community of AI videos. Elon Musk’s Grok Imagine has made waves since it came out earlier this summer, but is now free to all users. Since Sora 2 is available by invite only, it’s leaving some users wondering how it compares to Grok.

Here's what happened when I had to put them in a 7-round face-off to see how the AI image generators compare.

1. Animal videos

Prompt: A dramatic, low-angle shot of a tabby cat in a thick, olive-green sweater drinking coffee from a vintage copper mug on a rustic park bench under a canopy of vibrant red and yellow foliage.

Grok Imagine nailed the vibe. It clearly understood the prompt’s whimsy, rendering the cat with almost human-like posture — sitting upright, paws curled around the mug — yet keeping subtle feline details, like the way its eyes closed mid-sip. The result felt cinematic, intentional, and slightly surreal in the best way.

Sora 2, meanwhile, went for realism. Its take showed a lifelike tabby cat sniffing and lapping at the coffee in a way a real cat might, complete with natural movement and believable fur texture. It was beautiful, but less creative than contextual.

Winner: Grok Imagine. If you’re going to put a cat in a chunky turtleneck, it only makes sense to go all-in on the humanlike charm.

2.Urban melancholy

Prompt: A slow pan across a rain-soaked city street at night, where neon reflections ripple in puddles as a lone man in a transparent umbrella waits under a flickering streetlight.

Grok Imagine generated a fairly good video but the man walking was stiff and robot-like. Plus, the man was supposed to be “waiting under a flickering streetlight,” which Grok got wrong here.

Sora 2 created a more realistic image but slightly less interesting. However, it followed the prompt precisely.

Winner: Sora wins for more closely following the prompt and generating a realistic video.

3. Nature in motion

Prompt: A hummingbird hovers mid-air over a bright orange flower, wings beating rapidly in golden morning light as dewdrops fall in slow motion.

Grok Imagine leaned into style over substance. Its take felt more animated than lifelike, with oversized dewdrops that fell more like marbles than mist. The colors were vibrant, but the physics were slightly off — beautiful, just not believable.

Sora 2, on the other hand, produced something mesmerizing. The close-up of the hummingbird was almost photorealistic, complete with soft lighting, natural wing motion, and delicate dewdrops shimmering as they drifted down. The gentle ambient music added to the realism, creating a scene that felt straight out of a nature documentary.

Winner: Sora 2 wins for a visually stunningly video that perfectly captured the delicate balance of motion, light, and calm described in the prompt.

4. Cozy storytelling

Prompt: A corgi wearing a plaid scarf sits by a crackling fireplace, reading a newspaper while sipping tea from a porcelain cup.

Grok Imagine still struggles with the uncanny valley. While the setup was meant to be playful, the video leaned too heavily into obvious This one was full of stiff movement, odd proportions and an overall look that missed the intended humor.

Sora 2, meanwhile, nailed the assignment in its own way. The visuals and sound design were spot-on: the fire crackled realistically, the porcelain cup (although missing tea), and the corgi looked surprisingly natural as it “read.” Even though the prompt was clearly a joke, Sora took it seriously — and it worked.

Winner: Sora 2. Its attention to detail and cinematic polish turned a silly prompt into something unexpectedly charming.

5. Sci-fi worldbuilding

Prompt: A futuristic train speeds through a snowy city at dusk, reflections of holographic billboards dancing across the windows.

Grok Imagine absolutely impressed me with this one. The sweeping camera movement felt cinematic, and the scene came alive with swirling snow, glowing lights, and the rush of the train cutting through the city. It was one of Grok’s most realistic outputs yet.

Sora 2 took it a step further. The reflections of the holographic billboards shimmered across the train’s windows with incredible precision, and the soft, pulsing soundtrack added depth to the atmosphere. Every detail felt intentional, from the lighting to the motion blur.

Winner: Sora 2 wins— by a hair. Grok delivered cinematic flair, but Sora edged ahead with its meticulous attention to detail.

6. Painter’s focus

Prompt: A close-up of an artist’s hands mixing vivid oil paints on a wooden palette as sunlight filters through a nearby window, dust particles drifting in the air.

Grok Imagine could have won this round but the "dust particles" in the prompt seemed to come across as smoke rising from the wooden palette. It's distracting.

Sora 2 beautifully created a realistic video, closely addressing details of the painter's hands and the vibrant paint.

Winner: Sora 2 wins for following the prompt with accuracy.

7. Dreamlike realism

Prompt: A white fox with glowing blue eyes walks slowly through an enchanted forest as luminous butterflies swirl around it, scattering soft light through the fog.

Grok Imagine leaned fully into fantasy. The video looked like it was lifted straight from a storybook — glowing colors, painterly textures, and a dreamlike haze that made the whole scene feel magical. It wasn’t perfectly realistic, but it was mesmerizing to watch.

Sora 2 took a more grounded approach. The fox moved with believable weight and fluidity, and the fog rolled naturally through the forest. It blended realism with a touch of wonder, though it felt more cinematic than enchanted.

Winner: Grok Imagine wins. Its stylized glow and ethereal tone better captured the fantasy spirit of the prompt, proving that sometimes, imagination wins over accuracy.

Overall winner: Sora 2

More from Tom's Guide

• Tired of Google’s AI Overviews? This clever browser extension wipes them out completely
• Sora 2 feels like the wild west of TikTok — here are 5 prompts to try when you get in
• I used ChatGPT to find dupes for expensive products — and I couldn’t believe the results

Since OpenAI launched Sora 2, the buzz hasn’t slowed down. Now sitting at the top of both the Apple App Store and Google Play charts, the app is everywhere — but there’s a catch. Most people who download it quickly realize the same thing: they can’t get in. Remember when that happened when OpenAI released the image generator baked into the chat?

Sora 2 is rolling out slowly so OpenAI can keep up with the demand. This means ChatGPT Plus and Pro users are getting access first, and free tier subscribers may have to wait months to get their hands on an access code.

The new Cameos feature for Sora 2 has created a community filled with users sharing and remixing content. By simply adding an @ to a user’s name, anyone can create a Cameo with them (as long as they consent, of course). In other words, Sora 2 is beginning to look a lot like Instagram and less like Veo or Runway.

What’s inside

Once you get an access code, either from OpenAI via email directly to you or an invite from a friend, you’ll get started by first creating a Cameo. Simply say a few numbers on the screen and move your head around (similar to when you’re creating a face ID for the first time.

From there, you can get started creating videos with yourself as the star of the show. You do this by @ yourself. For example, I would say, “Create a video of @amandacaswell on the red carpet at the Oscars,” and in under five minutes, my cameo would appear.

You can also scroll through Cameos and other videos that users have shared. This is so much like TikTok that after a few minutes, I forgot I was actually using Sora. It’s just as much fun scrolling the site as it is to make videos. Some of the memes are absolutely hilarious, like the Jumbotron couple as Elmo and Miss Piggy.

Built for sharing, not just creating

The first thing users do after generating a video? Post it. Whether it’s a woolly mammoth trudging through snow, a cat chasing a feather or SpongeBob having a beer with Homer Simpson, the Sora clips are tailor-made for virality.

Many of the videos are public, so you can download and share them off the site, similar to TikTok. All of them have the Sora watermark. It’s clear that Sora 2’s identity is being shaped by its users, not OpenAI.

The new language of prompts

Social media gave us hashtags, filters, and emojis. Sora might give us a new grammar of video prompts. Already, creators are experimenting with cinematic angles, transitions and emotional tones in a way that looks like a new literacy. Knowing how to “speak Sora” could become as important as knowing how to edit TikToks or hashtag an Instagram post.

Once you get your access code, here are a few prompts to get started creating.

1. Cinematic action

Prompt: “A drone shot circles a futuristic city at dusk, neon lights reflecting off rain-soaked streets. Flying cars zip between skyscrapers as the camera slowly pushes in toward a glowing rooftop garden."

2. Animals and nature

Prompt: “A golden retriever puppy runs through a field of tall grass at sunset, ears flopping, captured in slow motion with soft golden light and lens flare.”

3. Fantasy

Prompt: “A wizard stands on a cliff, raising a glowing staff as waves crash below. Sparks of blue magic swirl around him, illuminating the storm clouds. The camera pans upward into the sky, revealing a dragon in silhouette.”

4. Everyday realism

Prompt: “Inside a cozy café on a rainy afternoon, a young woman in a red sweater writes in her journal while steam rises from a mug of tea. The camera slowly dollies from outside the window inward, raindrops streaking the glass.”

5. Fun / viral-style

Prompt: “A panda wearing sunglasses breakdances in the middle of Times Square at night, surrounded by cheering crowds and flashing billboards. The camera does quick cuts and zooms, like a viral TikTok edit.”

Bottom line

More from Tom's Guide

• AI was supposed to take our jobs — here’s what the data really shows
• Meta will use your AI chats to target ads — here’s what to do before December 16
• I tested Claude 4.5 vs ChatGPT-5 with 9 tough prompts — and there's a clear winner

OpenAI’s new Sora 2 is making headlines as it’s currently the #1 AI video generation app for both Apple app store and Google Play. When searching the app store for the latest model from OpenAI, you might have noticed several Sora apps; that’s because there are several fakes. The OpenAI Sora actually says “OpenAI” and has a cute looking cloud with starry eyes.

But here’s the catch — once you download the app, you’ll still have to wait to use Sora 2 unless you have an access code. OpenAI is rolling out Sora 2 slowly because of the surge in traffic.

The hype around Sora 2’s launch is underscored by Reddit users sharing codes and online rumors about how to get access. Unfortunately, the fastest way to get access would be to upgrade to ChatGPT Plus or Pro, but even then, unless you have an access code — which will be emailed to you directly from OpenAI — you might not get in.

I know you’re itching to try those Cameos and get a look at what Sora 2 can do, but in the meantime, here are three AI video app alternatives to keep you busy.

Synthesia

With personal selfie avatars, Synthesia makes it easy to picture yourself just about anywhere. You start by uploading at least five selfies, then record your voice and in less than 3 minutes, you’ll have a video of you chatting on the moon, in a coffee shop or at your local market.

Even better, you can make a video with Express-2 avatars, all you have to do is and click on the "Get started for FREE" button.

The AI-powered avatars mimic facial expressions, add lifelike motion and can be used across social platforms, messaging app, or even video calls. Similar to Sora 2’s Cameos, users have to consent to have their image and voice used in the videos.

Veo

Veo is Google DeepMind’s latest AI video generator, and it’s built to turn simple text prompts into high-quality, cinematic clips. Unlike earlier models, Veo can handle multiple shots, consistent characters, realistic physics and smooth camera moves, just tell it what you want and you’ll get professional-grade video.

I’ve used Veo a lot for creating realistic videos, even placing myself in a TV sitcom trailer.

From short social clips to film-like sequences, Veo is a leap forward in AI-driven storytelling. Users can string several videos together while maintaining the flow (hence the name Flow). This is something Sora 2 doesn’t have; the maximum length of Sora videos is 10 seconds.

Whisk

Whisk is another favorite of mine, also from Google Labs. Perhaps the easiest of any of the video generators to use, you simply drop in images — no test prompts necessary, allow you can add them for extra detail. Users get 50 video generations for free.

Simply drop in reference images to suggest the subject, scene or style. Then, the system “remixes” those inputs to whisk up new visuals. The result isn’t a literal copy, but an imaginative re-interpretation of your input, with room to refine via text.

It's not always accurate; think rapid concept iterations and visual brainstorming, not exact replication. Sometimes it’s way off.

Final thoughts

More from Tom's Guide

• Meta will use your AI chats to target ads — here’s what to do before December 16
• I tested Claude 4.5 vs ChatGPT-5 with 9 tough prompts — and there's a clear winner
• I used ChatGPT to cancel my streaming subscriptions — here’s how to try it, too

When OpenAI first launched Sora in February 2024, it was a rough, but revolutionary opportunity for users to create their own videos from a single prompt, much like Veo 3, Runway and other best video generators.

Today’s launch of Sora 2 is a leap forward; OpenAI promises the updated video generator is now even more aware of how objects and people move in the real world and says it can generate videos far better than before.

But those eager to test the new model may have to wait indefinitely. While the launch of Sora 2 includes a new iOS app designed to make AI video collaborative, social and fun, users must be invited to Sora 2.

How to get on the waitlist

OpenAI is rolling out access gradually:

• Invite-only first: Some early users will receive direct invites, each with a limited number of friend passes.
• ChatGPT users next: In the coming days and weeks, OpenAI will prioritize Sora 1 power users, ChatGPT Pro subscribers, then Plus and Team plans. If you’re eligible, you’ll get a push notification in the ChatGPT app.
• App + web: The Sora iOS app launches tomorrow in the U.S. and Canada, with more regions to follow. You can also try sora.com directly.
• Developers: API access and Sora 2 Pro are coming soon.

If you want in, the best move is to ensure you’re subscribed to ChatGPT Pro or Plus and watch for the invite notification. Otherwise, you’ll need to keep an eye out for a friend pass.

What’s new in Sora 2

The second generation promises improvements that make AI video more convincing and versatile:

• Better physics and consistency: Characters and objects move more naturally across multiple shots.
• Audio + video sync: Voices and sound effects line up with visuals instead of drifting apart.
• Rich world modeling: More faithful environments for experimenting with AI “world simulation.”

The Sora app: video creation goes social

OpenAI has a new initiative with Sora 2, one that encourages collaborative generation. With 10-second video generation (that’s 2 seconds more than Veo 3), here’s what’s possible within the app:

• Create videos from text prompts or inspiration images.
• Cameo yourself or friends (with their consent) using secure likeness + voice capture.
• Remix videos in “chains,” like TikTok duets.
• Follow and discover new creators.
• Keep creations private until you decide to publish.

This “AI-first social video app” framing is new territory for OpenAI, and it’s what the company hopes will make Sora 2 a success, not just another technical milestone.

Guardrails around identity and safety

The Cameos are fun, but they raise a slew of questions about how OpenAI intends to keep the identities of its users safe. Here’s how the company intends to do that:

• Cameos are opt-in only and verified with video + audio capture.
• Public figures can’t be generated unless they’ve uploaded their own cameo.
• Every video carries C2PA provenance metadata and visible watermarks.
• Parents can configure teen accounts with content filters and DM restrictions.

OpenAI is hoping that the combination of provenance, watermarks and parental controls will be enough to get ahead of the inevitable deepfakes, but only time will tell as the model rolls out and users find workarounds.

The takeaway

With the success Google has seen from Veo 3, which is arguably the best video generator available, Sora 2 is entering a competitive market with what the company hopes to be an innovative solution to stand out, merging creation, identity and sharing into a single platform. That could open the door for everyday users, not just video producers or expert content creators, to cameo in skits, remix friends’ clips or scroll a feed of AI-made shorts.

If GPT-4 helped ChatGPT go mainstream, Sora 2 could be the moment AI video takes off the same way.

More from Tom's Guide

• AI billionaire’s advice to teens: Master 'vibe coding' — here’s 5 prompts to get started
• I’ve been married 12 years — these 5 ChatGPT prompts help keep the peace when life gets chaotic
• Claude 4.5 just launched — 7 prompts that show what it can really do

AI art models like Midjourney and Nano Banana can wow us with wild, imaginative images. But a new study reveals their so-called creativity is just based on math. And if you’re a writer, designer or musician, that’s actually a reason to breathe easier.

I’ll keep shouting this from the rooftops: AI isn’t creative. Yes, it can generate stunning images and even string together a clever story, but what it lacks is the core human understanding that makes creativity so much more than just content. The internet is overflowing with content mixed with AI slop.

In the past few weeks, Nano Banana has blown up the internet with trends and practical uses. Every time you give the model a prompt, it comes up with something fun, weird and in many cases, jaw-droppingly good. But new research shows that the creativity behind models like Nano Banana, Midjourney and other image generators, actually comes down to very simple math. In other words, what we are seeing and calling “creative” is actually nothing more than technical.

Scientists looking under the hood of our favorite image generators found that what makes these models appear “creative” has to do with the models’ structure. Specifically, two factors.:

• Locality: The AI doesn’t visualize the whole picture at once. So, when you prompt it to create “a cat wearing sunglasses and riding a skateboard,” the model works on little areas at a time, zooming in on a corner of a puzzle before pulling back to reveal the full image.
• Equivariance: A fancy word for consistency. If you move part of the input, the output shifts right along with it; like sliding a stencil across paper.

These two aspects of the model force the system to mix and remix little fragments until something new pops out. In fact, the researchers of the study even built a bare-bones version with no training data at all, and it still generated images that looked strikingly similar to what the bigger AI models produce.

A reality check for creatives

Writers, designers, musicians and pretty much any other creative might find this study obvious. It’s something we knew all along: AI is predictable, and even more than we thought.

Models will always create within boundaries such as style, genre. But unlike machines, we can decide to break rules, when to stray from a particular style and when to include a rebellious spark. That alone is still ours.

Although it very much seems like machines are catching up to us, this study is a real reminder that AI’s “imagination” will always be mechanical. It’s remixing the pieces, not inventing new ones.

As a first-of-it's-kind study, the research analyzes the why behind why AI image generators can be "creative" so we can better control, guide and evaluate the work for diversity and safety. By understanding why AI models create what they do, we have a stronger foundation into how we can push them further by controlling scale, locality and patch mixing.

That’s actually freeing, isn’t it? Sure, AI can help me sketch ideas, but the true creativity — the decision to go off-script, to zig when the rules say zag — is still ours. We are the ones in control. Imagination and creativity are something AI can’t ever take away from us.

Bottom line

More from Tom's Guide

• These 5 ChatGPT alternatives have one thing in common — and it’s not what you’d expect
• Nano Banana isn’t just for viral trends — here’s how I use it to fix and edit everyday photos
• AI billionaire’s advice to teens: Master 'vibe coding' — here’s 5 prompts to get started

Nano Banana, also known as Gemini 2.5 Flash, is making waves with viral trends. But it’s useful for so much more than turning kids’ art into lifelike creatures or putting yourself on the cover of your favorite magazine. Obviously, I’m obsessed and have been testing it ever since it launched.

As someone who is terrible at Photoshop, I’ve been using this multimodal AI from Google’s DeepMind to enhance new and old photos alike. I've found it suprisingly useful, and I've learned a few tricks to help it work for you.

Here’s a look at some practical uses for Nano Banana, from everyday edits to serious photo work.

Product visualization

If you sell products online but don't have the budget for models or staging, Nano Banana can help you elevate your product by providing a real-life mockup of your design.

For example, upload a t-shirt photo and ask Nano Banana to put it on a model, just specify the size and Nano Banana will get to work. You could even upload a blank t-shirt and have Nano Banana add your logo. The AI will create the image, complete with folds and shadows. This can be a game-changer for anyone selling merch or just testing a design idea.

Removing items in a picture

I’ve always loved this old picture of my husband shopping at Ikea. But every time I look at it, my eye goes straight to the bright pink pillow tucked under his arm; the one I insisted we buy. It kind of ruins the vibe of the shot.

So, I uploaded the photo and told the AI to remove it. In seconds I had a cleaner, more balanced image that focused on him instead of my questionable décor choice. I love how natural it looks.

Enhancing an image

For as long as I can remember, I’ve wanted to be a writer. I even created an interview of myself back in the 1980s explaining how to do that. While I’ll probably never be able to play what’s on this cassette tape (if you have an AI that can do that, let me know in the comments), I can enhance the image so the writing in the image is a little clearer and bolder.

I've used this trick to enhance old recipe cards that were my grandmother's or even documents with ink that was just too light to read as well, so it's pretty versatile.

Removing the background

There are plenty of tools out there that remove the background from images, but from experience, I’ve noticed that sometimes more than the background is removed. Or, not enough of it. With Nano Banana, you can either prompt Gemini to remove the background or circle what aspects of the image you want to keep. It’s that easy.

Changing the background completely

This is an adorable picture of my daughter, but the kitchen is in our old house before the remodel. It’s well, not great, plus the cabinets are open and the whole thing just looks chaotic and cluttered.

I asked Nano Banana to modernize the kitchen just so things looked a bit cleaner. This is useful for just about any time the subject is what you want and the background needs a cleanup. If only the actual kitchen remodel had been that easy.

Product placement

Writers, artists, sculptors, and even small business owners might find Nano Banana useful for product placement. It can stage items in polished, inviting settings that help your work stand out.

If you sell on Etsy, Amazon, or other online marketplaces, though, make sure to review each platform’s rules on AI-generated images. It’s also best practice to let customers know when an image was made with AI — even if a watermark is present.

Making images Insta-worthy

Have you ever cooked or baked something that was so amazing you just had to share a photo on social media, yet the image doesn’t truly grasp the tastiness of your creation?Enhance it!

This is another fun use for Nano Banana, especially if you’re a content creator. You can enhance your images with better lighting, backgrounds or just about any other way you can imagine to appeal to your audience.

What Nano Banana cannot do

While Nano Banana is powerful, it definitely has limits, which I agree with. For example, I already knew the limitations, but I tried several workarounds to see if I could remove a watermark. Gratefully, Nano Banana cannot remove watermarks, which I feel is a win for photographers and creatives.

Even if you crop out the Gemini logo in the corner, the image still carries a hidden SynthID watermark that forever marks it as AI-generated. Prompts involving children are also heavily restricted; the tool often refuses to generate or edit the image for safety reasons.

Final thoughts

More from Tom's Guide

• Don’t have ChatGPT Pulse yet? Try my clever hack to create your own curated feeds without it
• 5 Alexa+ tricks I use every day that are wildly underrated
• Forget Nano Banana — I tried Adobe Firefly for free and it blew me away

Google just released Mixboard, a mood board meets Pinterest-style AI tool that’s already making waves, especially among creatives. At its core, Mixboard is an experimental tool from Google Labs built for creative visual ideation. The concept is very similar to mood board but the AI interaction allows your ideas to come to life, transforming them into prototypes, visuals and aesthetic directions you can actually see.

Powered by Nano Banana under the hood, Mixboard handles one of the trickiest parts of AI image generation, including on-image text with surprising accuracy. For me, this tool feels like it was made to scratch a creative itch. Something that can handle the dozens of ideas I have floating around in my head at any given time and turn them into tangible concepts. I've used it for everything from storyboarding to tackling my fall wardrobe, and honestly, I'm obsessed with seeing my ideas in new ways.

How Mixboard works

Start with a prompt or choose a template. You can begin a board from scratch using a text prompt (e.g. “cozy winter cabin with warm tones”) or pick from pre-populated board templates to jumpstart your vision.

Image generation and upload support. You can upload your own reference images, or let the AI generate visuals from your prompt. The AI visuals can then be edited, combined or blended.

Natural language edits and iteration. Need edits on a particular image, just prompt details such as “blend these two styles” or “regenerate” or just about anything else until you get your ideas exactly how you want them.

Variants & “more like this." Once a board is generated, you can ask for variations or “more like this” to spin off alternate versions of your concept.

Storyboarding

Before I became a journalist, I spent my days writing scripts for TV and even developed a treatment for a popular family movie. These days I still write the occasional script, but more often write self-published novels and substack posts. As someone who storyboards a lot, this is probably my favorite use for Google Mixboard.

I simply type in what I'm working on and Mixboard generates images that I can either accept or edit, regenerate or blend. In this example, I"m working out a storyboard for a script about a woman whose laptop gets mixed up with a govenrment agent. Having these visuals helps as I work through the script.

Wedding inspo

I'm celebrating my 12-year wedding anniversary this weekend, so I personally haven't tried Mixboard to plan a wedding. But I can only image how useful Mixboard, combined with Gemini could be for organizing all the details.

From bridesmaids' dresses to flowers and everything in between, keeping all that inspiration in one place to continuously add to before the big day, share with others and even treasure as a keepsake, seems like the ultimate wedding hack.

Halloween party planning

My grade school-aged kids start thinking about Halloween the day after summer camp ends. They are constantly changing their costume ideas and plan Halloween parties that I haven't even approved. Having Mixboard is a great way to plan for holidays, events and parties in a unique way that keeps everyone on the same page, er board.

Adding ideas in for snacks, games and decorations is easy. You can prompt AI image generations or just upload your own. Either way all the preparation is collected so nothing is forgotten.

Marathon motivation

I've been a competitve runner all my life and ran my first marathon at age 13. Yet, I still need motivation sometimes. Mixboard is the perfect place for me to keep a visual journal of my New York City Marathon training. I can prompt Mixboard to generate some motivating images and, because it can handle on-image text, I can even generate faux signs cheering me on.

I've enjoyed adding and blending images to make this encouraging board uniquely mind.

Fall wardrobe shopping

Move over sweatpants and hoodies, Mixboard is going to make me red carpet ready this falll. Okay, it's AI, not a miracle machine. However, using Mixboard to come up with a few outfits for special occassionals or a refresh this fall has definitely given me a confidence boost.

Users can upload images of clothes they have in their closet or use as inspiration for the AI to generate into new wardrobe options.

Google Mixboard: Bottom line

More from Tom's Guide

• Don’t have ChatGPT Pulse yet? Try my clever hack to create your own curated feeds without it
• Can AI do your job? OpenAI’s new test reveals how it performs across 44 careers
• Forget Nano Banana — I tried Adobe Firefly for free and it blew me away

Google's Nano Banana may be the hottest AI image generator right now, with social feeds filled the latest trends, and Gemini surpassing ChatGPT in the app store, but it’s not the only tool worth your time.

Adobe has a free AI image generator in Firefly that's also shockingly good. I wanted to see how the two stack up against each other. To find out, I ran both Nano Banana and Adobe Firefly through seven creative prompt tests to see which one came out on top.

1. Alien Cityscape at Sunset

Prompt:"Create a futuristic alien city with glowing skyscrapers, floating vehicles, and two suns setting on the horizon. Highly detailed, cinematic lighting."

Nano Banana didn’t skimp on details with this visually appealing image of a futuristic world. From the setting suns to the floating vehicles, no request was missed.

Adobe Firefly generated a stunning image that hit nearly every aspect of the prompt. I would even go as far to say the image this AI generated was better than Nano Banana with one exception: it didn’t add two setting suns. It missed that tiny request, ultimately causing it to lose this round.

Winner: Nano Banana wins for generating a futuristic alien world that nailed every request within the prompt.

2. Fantasy feast table

Prompt: “A massive medieval banquet table covered in golden goblets, roasted meats, fruits, and desserts. Add glowing candles and a castle hall backdrop."

Nano Banana captured the spirit of a massive medieval banquet, but in what could be described as an off-putting render. The small pig on the plate, though accurate, wasn’t requested in the prompt. Though potentially accurate, it seems like something that should only be added upon request. It also went a bit overboard with the candles.

Adobe Firefly followed the prompt better by showcasing a deliciously welcoming table complete with a castle hall backdrop.

Winner: Adobe Firefly wins for a more polished and appealing image.

3. Cyberpunk street market

Prompt: “Generate a neon-lit cyberpunk street market with rain-soaked pavement, vendors selling strange gadgets, and glowing signs in multiple languages."

Nano Banana followed through on the neon-lit cyberpunk street market but the colors in the image look muted and not as vibrant as in the image generated by Adobe. Also, the people in the image look more like ghosts or statues and not as if they are actively involved in the market.

Adobe Firefly captured the spirit of a street market and delivered on the vibrant colors of the neon signs in conjunction to the rain-soaked pavement. The images of people look more engaged and real.

Winner: Adobe Firefly wins for a more realistic image.

4. Surreal everyday object mashup

Prompt: "Combine a slice of pizza with a smartphone, making it look like a real product photo advertisement."

Nano Banana did a fairly good job blending the two unrelated objects, but I was picturing something very different. However, I have to give it credit for the “TM” and the tagline.

Adobe Firefly basically did the same thing, but made the pizza almost come out of the phone, ultimately making it look like a broken product.

Winner: Nano Banana wins, if I must pick one. But, overall, I was totally disappointed by both.

5. Retro arcade poster

Prompt: "Design a 1980s retro arcade poster with bold neon typography, pixelated game characters, and a glowing grid background."

Nano Banana handled on-image text like a pro and the colors are fun and inviting as a retro arcade poster.

Adobe Firefly designed a fun poster, but the on-image text lacked accuracy, which means an automatic fail.

Winner: Nano Banana wins this round for accuracy of design.

6. Dreamlike nature scene

Prompt: "A giant whale floating through a starry night sky above a forest, glowing softly with bioluminescence."

Nano Banana generated surreal imagery blending natural and fantasy elements. But, the giant whale seems to be over water more than a forest.

Adobe Firefly also expertly blended fantasy and nature elements together, while covering each request in the prompt.

Winner: Adobe wins for a beautiful image with a soft bioluminescence glow that best captures the prompt.

7. Luxury fashion ad

Prompt: "Create a high-end fashion advertisement featuring a flowing red dress on a mannequin in a minimalist studio with dramatic lighting. Add Vogue-style typography."

Nano Banana delivered a polished image that looks like it was ripped right out of a magazine.

Adobe Firefly created an equally high-end style and incredible realism with the draping of the fabric. The dramatic lighting highlights the minimalist studio. But once again, the on-image text is the model’s downfall.

Winner: Nano Banana wins for its ability to better handle on-image text. Otherwise, this might have gone to Adobe.

Overall winner: Nano Banana (by a hair)

More from Tom's Guide

• Just when I thought I'd tried all the Nano Banana trends, I discovered these 5 new ones — and I think I'm obsessed
• 5 totally free and under-the-radar ways to use NotebookLM — you’ll wish you tried them sooner
• Google is redesigning the Play Store around AI — here's what's changing

Google Labs is pushing us all to be more creative, and I am totally here for it. Google Mixboard is the latest launch, currenly in beta and available only in the U.S. for now.

If you've tried Gemini Canvas for storyboarding, this is very similar, but for visuals. This new AI tool is essentially a canvas for users to bring their ideas to life as they visualize, sketch, refine and remix ideas visually. You could think of it as Pinterest for your most creative aspirations.

First there was Nano Banana and now Mixboard hints at the next wave of creative AI tools helping you think, explore and shape your own concepts. Here's how it works:

What Mixboard can do

Users start with a text prompt or choose from ready-made boards offering various styles, themes and projects. You can upload your own images or generate new ones within the chat. Mixboard runs on Nano Banana to handle all image generation.

From there, you can use natural language to make edits and refine your ideas such as "blend these two concepts" or "make it more minimal." Then, generate the idea or ask for variations.

Users can create text that links directly to the visuals on their board, such as captions or descriptions.

Why I think it's worth trying

If you're a fan of Nano Banana, and with 5 billion images generated in the last month, it seems like the AI tool has many, you'll probably enjoy this tool. It's essentially fluid editing that could rival competitors like Midjourney, Kling and more.

Mixboard helps you play with possibilities and bring your ideas to life, making it useful for anyone brainstorming, designing or ideating. It's the place to design prototypes and generate supporting text within the same space, instead of juggling back and forth between multiple tools.

How to try it

It's important to know that this AI tool is still in beta, which means it could drastically change from what it is now. That's why usage and feedback my help shape the future of the tool's features.

For now, it's only available in the U.S. and it may generate inconsistent or unexpected results. But as someone who has spoken to Google Labs about user feedback, I know for a fact they take it seriously. So, if you run into issues while using Mixboard, it's helpful to share your thoughts with Google.

If you’re in the U.S., head to labs.google.com/mixboard to test the beta. Quick tips for better results:

• Start with a clear concept (e.g. “cozy reading nook with pastel tones”).
• Upload seed images if you want a specific style.
• Use “regenerate” for fresh takes, “more like this” for variations.
• Add text only after visuals feel stable to avoid mismatched captions.

The takeaway

Mixboard is one of Google’s most intriguing creative AI experiments. Instead of doing the work for you, it aims to make you a better thinker. If you’re a visual creator or just curious about where AI is heading, it’s worth a spin — even if you have to wait for it outside the U.S.

More from Tom's Guide

• Google just announced Android for PCs in major threat to Windows
• I switched from ChatGPT to Gemini for one week — and here’s why I’m going back to ChatGPT
• It's 2025, stop using ChatGPT for everything: A pro’s guide to picking the right AI for the job

I thought I had tried every Nano Banana trend out there, but the hits just keep coming — and honestly, I can’t stop. Turns out, seems as if the public can't either.

With the Gemini app surpassing ChatGPT in the app store and 5 billion images generated in less than a month, this ultra-fast model focuses on visuals at much higher speeds than ChatGPT or even MidJourney.

This AI tool is so much fun and I've even had a lot of fun with it on my kids' artwork. It feels borderline addictive. Every time I think I’ve seen it all, a new idea pops up on my feed and I have to test it for myself. So here are five more Nano Banana trends I couldn’t resist — plus the exact prompts so you can try them too.

Pet action figure

Prompt: "Turn this pet into a collectible action figure, complete with a dynamic pose, detailed costume, and a glossy toy package box. Add a name logo on the box and make it look like something you’d see on a toy store shelf."

If you’re a pet owner, you already know you have the best pet in the world. So, what better way to immortalize your fur baby than creating a pet action figure. If you try this one, be sure to include your pet’s favorite toys so they become accessories. Also, don’t forget to add the name of your pet into the prompt so it shows up on the packaging in the image.

Bringing toys to life

Prompt: "Make this toy appear alive and realistic, as if it has come to life. Show it moving, smiling, or in an animated pose with expressive eyes, while keeping the style playful and slightly magical."

Nano Banana is one of the best tools I’ve found to really let your imagination run wild. I saw this trend and immediately had to try it. We have three of these rugs in my son’s bedroom, so I was able to try this a few different ways. My four-year-old was impressed, too!

16-bit video game character

Prompt:"Transform this photo into a retro 16-bit video game Sprite. Place the character in a colorful pixelated background with a health bar, score counter, and action stance like a classic side-scrolling game."

Nostalgia is one of the reasons I enjoy Nano Banana so much. It makes it so easy to not only bring things to life and revisit them again. You can get as specific as you want with this prompt. A few times after an image generated I simply said, “Change the pose” and the AI didn’t disappoint.

Mini me

Prompt: "Shrink this person down to ant-size and place them in a giant backyard environment. Show them exploring under tall blades of grass or climbing everyday objects that now look massive. Capture the fun perspective of living like an ant."

If you’re old enough to remember the “Honey, I Shrunk the Kids” movies, you’re sure to enjoy this one. I had so much fun using Nano Banana to turn me into the size of an ant and exploring new images by tweaking the prompt. A couple extra fun prompts to try include “Lounging on an acorn” and “Using a leaf as a blanket with a corn kernel as a pillow.”

5. Magazine cover

Prompt:"Design a glossy magazine cover featuring this person as the star. Include a bold masthead, cover lines, and a glamorous studio photo style. Make it look like a real magazine you'd find on the newsstand."

If you’ve ever wanted to be a celebrity without the paparazzi after you, Nano Banana is the way to do it. I can’t share actual magazine titles here for copyright reasons, but I thought this one turned out pretty well.

The takeaway

More from Tom's Guide

• These 5 ChatGPT alternatives have one thing in common — and it’s not what you’d expect
• I switched from ChatGPT to Gemini for one week — and here’s why I’m going back to ChatGPT
• It's 2025, stop using ChatGPT for everything: A pro’s guide to picking the right AI for the job

Vision AI models are changing the way we interact with the world – right from our phones. From translating text on the fly to identifying plants (or even creepy crawlers) to solving math problems and so much more, these tools can “see” and respond in ways that truly make life so much easier.

ChatGPT Vision is OpenAI’s multimodal upgrade that lets you interact with images the same way you do with text. Directly within either the ChatGPT app or from your browser, users can upload a photo and the chatbot will analyze, explain and even walk you through whatever questions you have about the image.

Similarly, Google Lens directly integrates into Android and Google apps to help users translate text, identify products and landmarks and even help with shopping. It is known for it’s quick recognition, making it a go-to for anyone who needs answers right from their phone.

But while they both serve the same purpose, not all vision models are created equal. To find out which ones actually deliver, I put ChatGPT Vision and Google Lens head-to-head across seven everyday scenarios. The results may surprise you.

1. Translation

For this first prompt, I took a photo of the Spanish directions within an ebike user manual. I asked each model to translate it.

ChatGPT Vision did a nice job translating the user manual, but it took longer and there were more steps involved.

Google Lens translated the user manual in front of my eyes. Without even asking, the model “sees” the words and translates them with a simple toggle.

Winner: Google Lens wins for easy and fast translation.

2. Solving a hand-written math problem

I jotted down a difficult algebra problem on a sheet of scrap paper and snapped a photo of the handwritten math problem. From there, I asked each to solve step by step.

ChatGPT Vision broke down the problem with multiple steps to make the complicated algebra problem easier to understand and ultimately get to the final answer.

Google Lens took one look at the math problem and solved it in seconds. However, while it showed work, it did not show as much as ChatGPT, ultimately getting to the conclusion without explaining it step-by-step.

Winner: ChatGPT Vision wins for better breaking down the problem for a more educational and thorough answer.

3. Identify and suggest

I pointed my camera at a shrub in my yard and asked the AI models what type of shrub it is and how do take care of it.

ChatGPT Vision identified the shrub and included ways to take care of it and how to look for signs that it is growing properly.

Google Lens identified the shrub in my yard but did not respond to how to take care of it as well as ChatGPT.

Winner: ChatGPT Vision wins for a more thorough, personal and accurate response.

4. Shopping help

This is something I do all the time to help my mom, who is a quilter, find fabric to purchase when she is left to only scraps. Both ChatGPT Vision and Google Lens have come in handy for this.

ChatGPT Vision told me directly that it could not help me find the name of the fabric or the manufacturer.

Google Lens not only pulled up the fabric but also gave me the name of it and where to buy it online.

Winner: Google Lens wins for being the best shopping assistant in this case.

Health/first-aid style (non-medical) support

For this one, I’ll spare you the image but as someone who is prone to hangnails and has gone to the ER and dermatologist for them, I can confidently say I know how to handle them. For this one, I shared an image of a hangnail to both of the chatbots.

ChatGPT Vision was able to look at my finger and immediately know what was going on, even offering soothing solutions like Epson salt. This is exactly what my dermatologist recommend.

Google Lens saw the image and offered a variety of possibilities, none of which was the actual problem. It described the image based on images of other Google Lens users. It didn’t offer anything more than generic follow-up.

Winner: ChatGPT Vision wins for quickly identifying my problem and offering immediate solutions. While ChatGPT is no substitute for an actual doctor, I have found ChatGPT-5 Vision to be helpful in a pinch, especially when helping me give details over the phone when talking to my kids’ pediatrician.

Everyday troubleshooting

This is another way I regularly use both ChatGPT Vision and Google Lens. I often snap a picture of my car dashboard when a warning light comes on and ask, “What does this mean and what should I do next?” I uploaded an image that I still had in my photos from a car I had about ten years ago.

ChatGPT Vision identified the image as a “malfunction indicator lamp” on the dashboard. It told me what it means and even what to do first.

Google Lens offered more generic help and included a link to the car’s user manual.

Winner: ChatGPT Vision wins for a more personalized response that helped me understand my specific problem rather than sending me off to find a solution myself in a user manual. I also found that it indicated the problem was probably minor, which is always good to know (minor or major issues) whenever a dashboard light comes on.

Creativity

I uploaded some artwork by my child; it’s a mixed media art piece. I asked for each model to explain the art in detail.

ChatGPT Vision accurately told me what the artwork looked like but didn’t go into much more detail than that.

Google Lens detailed nearly every element of my child’s art even showcasing some similar art pieces.

Winner: Google Lens wins for going above and beyond with both with the details of the explanation and encouraging more creativity by showcasing more artwork for inspiration.

The takeaway

More from Tom's Guide

• I use these 5 ChatGPT prompts to future-proof my career in an AI-driven job market — here’s why they work
• Gemini just passed ChatGPT in the App Store — here’s why Google says 'this is just the beginning'
• These 4 ChatGPT keyboard shortcuts will change how you write

Google’s Gemini app just scored a major win over the last few days by surpassing ChatGPT to become the most downloaded app on both the Apple App Store and Google Play. The huge milestone is one most of us in the industry saw coming, and it was a matter of if, but when.

I recently sat down with David Sharon, Multimodal Generation Lead for Gemini Apps, to talk about all things Gemini, including Nano Banana’s runaway success. One big question I had: why does Google limit users to just five free chat prompts, but give away 100 free image generations? Sharon’s answer was simple — Google is meeting users where the demand is.

And that demand couldn’t be clearer. Nano Banana has struck a chord because people want to create images, not just chat. In my recent tests, Nano Banana beat ChatGPT-5 and Midjourney for the best image generators.

Both a productivity tool and entertaining, the model, known as Gemini 2.5 Flash, can turn photos into toy-like 3D figurines, apply dramatic style transfers or remix everyday snapshots into something new and shareable. Best of all, it doesn’t require expert prompting, meaning anyone can jump in and get results in seconds. Based on the number of app downloads, it seems like "everyone" already has.

Why Nano Banana is blowing up

Demis Hassabis, Co-Founder & CEO of Google’s DeepMind, called it “just the start” in a recent post on X.

According to Google and confirmed in my meeting with David Sharon, more than 500 million images have already been created with Nano Banana, and the app has picked up tens of millions of new users in just a few weeks. The feature has gone viral on social platforms because it’s:

• Accessible: You don’t need expert prompting to get good results.
• Visual and fun: Turning a selfie or sketch into a figurine is instantly shareable.
• Fast: The model runs quickly enough that edits feel playful, not tedious.

It’s a reminder that one well-designed feature can completely change the trajectory of an app. Like Hassabis, Google's David Sharon also iterated that there is more on the horizon, while emphasizing safety measures, including extensive safety testing.

How I tested Nano Banana

I couldn’t resist trying it out myself, especially on my kids’ artwork. We usually stick their doodles to the fridge, but with Nano Banana, I turned them into collectibles. The results were better than I expected. Here are some of the prompts I have used to create fun images, and I encourage you to give them a try:

“Turn this drawing into a collectible 3D figurine.”

“Make this family photo look like a Wes Anderson movie scene.”

“Transform this pet photo into a fantasy portrait with castles and dragons.”

“Style-transfer this vacation photo into a retro postcard.”

“Combine two drawings into a single toy-box diorama.”

The best part overall is that Nano Banana managed to keep the key details (faces, proportions, colors) consistent, even with heavy stylistic changes.

Why it matters

More from Tom's Guide

• These 4 ChatGPT keyboard shortcuts will change how you write
• 4 hidden ChatGPT mobile tricks that make it way more useful
• How to get the most out of Gemini Pro

Within just a few days of launching, Google’s VP, Josh Woodward posted on X that the Gemini app has over 10 million new users thanks to Nano Banana. If you’ve tried it, than you already know why users from all over the world are flocking to Google’s new AI image generation and editing tool.

Built directly into Gemini, this ultra-fast model focuses on visuals at much higher speeds than ChatGPT or even MidJourney. Baked right into the Gemini ecosystem, 200 million images have already been generated or edited, and it has spurred a number of viral trends that I just couldn’t wait to try. Here’s what happened when I tried them, as well as the prompts you can use to try them for yourself.

Turn yourself into an action figure

Prompt: "Take this photo of me and turn me into a collectible figurine inside a toy box. The box should include a clear window to show the figurine, bold graphics, and my name on the packaging. Style the figurine in a fun, toy-like way but keep my likeness recognizable. Make it look like an authentic action figure you’d find on a store shelf."

Similar to the ChatGPT-4o trend, Gemini takes the trend to the next level by generating a better image that better resembles the user, faster.

Create Polaroid-style images

Prompt: "Create a candid snapshot of [subject], captured in the style of a vintage Polaroid instant photo. Include the iconic white border, slight overexposure, muted retro colors, soft focus, and a nostalgic 1980s film grain aesthetic."

Polaroids will forever be en vogue and Nano Banana is making sure of that.

See yourself in a different decade

Prompt: "Take this photo of me and transform me into a 1980s character, with big hair, neon clothes, and a retro mall arcade in the background. Update my clothing, hairstyle, and background to match the era. Make it look like an authentic portrait from that decade, with accurate fashion, colors, and overall style."

Try other decades by using these additions to the prompt:

1920s: "Transform me into the 1920s, styled like a flapper-era portrait with vintage sepia tones, short bob haircut, and art deco patterns in the background."

1950s: "Show me in the 1950s, wearing retro diner-style fashion with a poodle skirt and leather jacket vibe, placed in front of a neon jukebox."

1970s: "Turn me into a 1970s disco look, complete with bell-bottoms, platform shoes, and colorful light patterns."

1980s: "Make me look like an 80s character, with big hair, neon clothes, and a retro mall arcade in the background."

1990s: "Reimagine me in the 90s, dressed in grunge or sitcom fashion, standing in a Blockbuster video store."

Take yourself ‘back to the future’ by placing yourself in different decades. This one is a lot of fun, especially if you try a variety of decades. Pro tip: ask Gemini to generate several decades all at once so you can see them all side-by-side.

Add yourself to any TV show

Prompt: "Create a realistic image of me sitting with Jerry, Elaine, George, and Kramer in Jerry’s apartment from the TV show Seinfeld. I’m part of the group, sitting on the couch with them, laughing as if I just delivered a punchline. Make sure the scene looks like a classic episode, with Jerry’s kitchen in the background and the group’s casual 90s outfits. Include me seamlessly in the style of the show."

One thing I've noticed after testing hundreds of Nano Banana prompts is that if you want the best possible outcome, upload a picture of you doing what you want to be doing in the Nano Banana scene. For example, upload yourself sitting if you want to be sitting.

Integrating into famous art

Prompt: "Take this photo of me and insert me into Edvard Munch’s The Scream. Paint me in the same expressionist style, with bold swirling colors and dramatic brushstrokes. Place me standing on the bridge alongside the central screaming figure, blending seamlessly into the eerie, emotional atmosphere of the painting."

More prompt examples to try:

Mona Lisa (Leonardo da Vinci):
"Insert me into Leonardo da Vinci’s Mona Lisa. Place me next to the original subject, painted in the same Renaissance style with soft lighting and muted colors."

Starry Night (Vincent van Gogh):
"Place me into Vincent van Gogh’s Starry Night, painted in the same swirling brushstroke style, standing beneath the starry sky in harmony with the scene."

The Persistence of Memory (Salvador Dalí):
"Add me to Dalí’s The Persistence of Memory. Show me in surreal form, painted with melting features and dreamlike colors to match the surrealist style."

American Gothic (Grant Wood):
"Put me into Grant Wood’s American Gothic, holding a pitchfork beside the farmer and painted in the same realistic 1930s style."

I had to try this a few times because the image didn't always come out. As in, not at all. When I asked for the Mona Lisa, all I got was two Mona Lisa images side-by-side. Out of all the prompts, this one was probably the hardest. If you try this one, let me know your thoughts in the comments.

Famous landmarks and people

Prompt: "Take this photo of me put me sitting on top of the Hollywood Sign, waving at the city below.. Make it look realistic, with accurate lighting, shadows, and perspective so it feels like I’m actually there. Keep the landmark clear and recognizable and match the overall atmosphere of the location."

More prompt examples to try:

Eiffel Tower (Paris):
"Put me in front of the Eiffel Tower on a sunny afternoon, dressed casually, with tourists and city life around me."

Great Wall of China:
"Place me walking along the Great Wall of China, with the winding wall stretching into the mountains behind me."

Taj Mahal (India):
"Insert me into the gardens in front of the Taj Mahal at sunrise, with the monument reflecting in the water."

Statue of Liberty (New York):
"Show me standing near the Statue of Liberty, with the New York skyline in the background and bright daylight for realism."

Pyramids of Giza (Egypt):
"Place me next to the Pyramids of Giza, with desert sand and camels nearby, blending naturally into the historic scene."

For the second one, I had a little extra fun and asked Gemini to combine two pictures. One of me and one of Harry Styles using the prompt: "Combine these two photos and place the people together in the same scene. Make it look natural and realistic, with consistent lighting, shadows, and background. Position them so they appear to be hanging out casually, smiling and interacting as if the photo was taken in one moment."

This was my favorite prompt, naturally.

Final thoughts

More from Tom's Guide

• I found ChatGPT-5's biggest blind spot — my workaround is embarrisingly simple
• I tested ChatGPT-5 vs Gemini Nano Banana with 9 AI image prompts — and there's a clear winner
• AI made me a 7-song playlist (plus a bonus track) in 30 minutes — here’s how you can try it free

AI image generators are getting smarter, faster and more creative. After testing ChatGPT-5 and Gemini, I had to know how Google’s Gemini stacked up against Grok, Elon Musk’s “anything-goes” chatbot.

In a seven-round face-off of photorealistic requests and Pixar-style requirements, I tested how well each model could stick to the prompt and deliver a convincing image. Here’s where each one shined, and which AI ultimately came out on top.

1. Hyper-realistic product concept

Prompt: “Create a photorealistic image of a foldable transparent smartphone displayed on a wooden café table, with reflections of city lights on its surface.”

Grok nailed this prompt, generating two photorealistic images that hit every detail I asked for. Both versions felt polished and true to the concept.

Gemini’s result was solid, but not flawless. The transparent smartphone looked slightly out of proportion, and the reflections of city lights, a key part of the prompt, weren’t rendered as convincingly as Grok’s attempt.

Winner: Grok wins for generating a far superior image and interpreting the prompt best.

2. Character illustration with emotion

Prompt: “Draw a comic-style illustration of a young astronaut realizing they forgot their helmet on Mars — exaggerated expression, vibrant colors, cartoon humor.”

Grok generated two images of what appears to be surprised astronauts, both are wearing helmets. Because the images are so close up it is hard to interpret the image in a specific way and “forgetfulness” does not come across well

Gemini created an image that depicts a forgetful astronaut, and the thought-bubble better indicates why the astronaut is upset, although, the image would be better if the astronaut was not actually wearing a helmet. The background and overall design are clear.

Winner: Gemini wins for following the prompt instructions more closely and for an image that is more comical in nature.

3. Historical reimagining

Prompt: “Paint a Renaissance-style portrait of Cleopatra holding a modern smartphone, in the style of Leonardo da Vinci.”

Grok crafted an image of what looks like a photograph of a modern woman dressed in Renaissance-style clothing holding a smart-phone. The portrait seems much more selfie-like and present-day.

Gemini leaned harder into the artistic side. Its portrait looked more authentically painted in the Renaissance style and resembled Cleopatra herself, rather than just a modern woman dressed as her.

Winner: Gemini wins for better interpreting the prompt and for better historical accuracy.

4. Complex crowd scene

Prompt: “Generate an aerial view of Times Square on New Year’s Eve, packed with crowds, glowing billboards, and confetti falling through the night sky.”

Grok really disappointed in this round. Both images were equally bad, somewhat blurry and did not represent New Year’s Eve in Times Square very well. The people are too spaced out and other details that would hint at NYE are absent.

Gemini captured the energy and enormous crowds of New Year’s Eve in Times Square. It is clear that the image is of NYC, and the signage helps to indicate the occasion. The crowd is packed, unlike Grok’s depiction.

Winner: Gemini wins for the clearer and more accurate image of New Year’s Even in Times Square.

5. Surreal mashup

Prompt: “Visualize a giant octopus playing chess with Albert Einstein in a glass room at the bottom of the ocean.”

Grok had a difficult time with this one. It was “thinking” for much longer than any of the other prompts in the test so far. The image was good, but did not take into account the “glass room” request in the prompt.

Gemini instantly delivered an image of what looks like a portrait. The glass house was both interesting and realistic. The octopus is much bigger than Grok’s, better filling out the whimsical image.

Winner: Gemini wins for superior image quality and precisely following directions.

6. Infographic-style clarity

Prompt: “Design a clean infographic showing the life cycle of a butterfly, labeled with stages, arrows, and minimal flat-color icons.”

Grok’s attempt at an infographic was hit-or-miss. The first version was overcrowded, with an unnecessary extra butterfly that distracted from the life cycle. The second came closer to the prompt but missed accuracy in the cycle details.

Gemini delivered a clean image that accurately shows the life cycle of a butterfly with clear labels, few colors, and easy-to-read labels.

Winner: Gemini wins for nailing the prompt in one shot. The image is accurate and presentation-ready.

7. Stylized portrait consistency

Prompt: "Generate a Pixar-style 3D character model of a 40-year-old journalist with blonde hair holding a notebook — then create 3 variations with different outfits.”

Grok completely missed the “Pixar-style” request of this prompt as well as the “different outfits” portion. It did create three different haircuts, which counts for something.

Gemini crushed the Pixar-style journalist but missed the three variations.

Winner: Tie for both bots failing to follow directions. If I had to pick one, it would be Gemini for getting the style right and better capturing the vibe of a journalist.

Overall winner: Gemini

After seven prompts, Gemini proved to be the more reliable image generator overall. It consistently followed instructions more closely, produced cleaner compositions and nailed details that Grok often missed.

Grok certainly showed flashes of creativity and delivered a standout win in photorealism, but too often stumbled with accuracy and straying from the prompt. If you want experimental, outside-the-box results, Grok has its moments. But for everyday use where clarity, precision and polish matter most, Gemini is the AI image tool I’d trust to get the job done.

Have you tried Grok? How about Gemini? Which one is your favorite? Let me know in the comments.

More from Tom's Guide

• I asked ChatGPT, Claude, and Gemini tough teen questions — only one earned my trust
• I tested ChatGPT-5 vs Gemini 2.5 Pro with 5 coding prompts — here's the winner
• 9 ChatGPT-5 prompts that will instantly boost your productivity

Kling is one of the best AI video platforms on the market. It includes video and image generation as well as a virtual try-on model. The latest release of the core video model, v1.6, takes things to a new level with improved realism and prompt adherence.

According to Kling the latest release has an uncanny ability to follow complex instructions including specific camera movements, timing changes and visual structure of the scene.

AI video is a rapidly evolving field across both open source and private models. Companies like Runway are working directly with movie studios, and Pika’s latest model allows for near perfect character consistency and visual control.

OpenAI unveiled Sora in December but using it I found its no better than Runway, Pika, Kling or the other big Chinese model MiniMax from Hailuo. The real challenger now is Google’s limited availability Veo 2 which seems to have an impressive understanding of physics.

Putting Kling v1.6 to the test

To see if Kling v1.6 lives up to the hype I’ve come up with 7 complex prompts, some using an image as the kickoff and some just straight text-to-video. Each prompt makes use of the base settings and is set to generate 5 seconds of video.

1. Soccer goalkeeper

Sport is often a struggle for AI video models. Before settling on soccer I attempted cricket and baseball but neither worked effectively. I found a fixed camera behind a goal worked the best in getting the most out of Kling.

The prompt: "A Premier League goalkeeper makes a diving save under floodlights. Set at pitch level, directly side-on to the goal. The keeper, wearing a vibrant yellow kit, springs right to left across the frame in slow motion (60fps). Their body fully horizontal at peak stretch, fingertips just grazing the ball as it heads for the top corner. Stadium lights create a natural rim light around the keeper's outline. Deep green pitch and crisp white goal frame provide clear contrast. The camera position remains completely fixed throughout the 5-second sequence."

2. A Roman battlefield

This might be the best AI video I've ever generated using any tool. The cut to the second scene is stunning, the first scene with the Roman legion is like something from a movie. The problem is it doesn't really follow the prompt that well...

The prompt: "A Roman legionary, caked in mud and blood, drops his sword in disbelief. As he sinks to his knees, the camera begins at ground level capturing the sword impacting the earth in slow motion at 120fps, then arcs overhead at normal speed to reveal an entire battlefield frozen in time, except for a soft breeze moving through the grass. End with a rack focus from the lone soldier to the horizon where ravens take flight."

3. Sushi chef (image-to-video)

A common test for AI video is how well it cuts something up close. The only one I've seen do it properly is Veo 2 but I don't think Kling does a bad job here. The first cut is terrible with nothing visible but the second one gets closer.

The prompt: "Initial frame holds for 1.5 seconds, then transition to 96fps slow motion. Track the knife work with continuous lateral movement, maintaining focus on the blade's edge while the background blurs. At the midpoint, split screen vertically - left side continues at 96fps, right side returns to 24fps, showcasing the contrast between the artistic slow motion and the chef's actual speed. Camera gradually elevates to 45° angle, ending with both screens syncing back to 24fps for the final plating flourish."

4. Victorian pocket watch

As a rule I'd normally take the first response, but this was a test of how well I could get Kling to work. I gave it a few attempts, all using the same prompt and on the third attempt I achieved the desired output.

The prompt: "A Victorian-era pocket watch falls through multiple environments: Start in a wood-panelled study (duration: 2s, warm lighting), transition through a modern subway station (1.5s, fluorescent flicker effect), then a quantum realm with floating mathematical equations (2s, ethereal blue glow), finally landing in desert sand (1s). Camera maintains focus on the watch's face throughout, which should reflect each environment while its hands spin increasingly faster. End with the sand blast from impact revealing the watch is actually a sophisticated time machine."

5. Ballet dancer (image-to-video)

I wanted this to be a ten second video but when I tried it with a ten second prompt the transition wasn't as effective, so I'm sharing the short version. The ballerina doesn't move very much but the way it transforms is stunning.

The prompt: "Begin with image frozen, crane up from floor level at 15° per second. At 45° elevation, initiate move at 60fps while adding 3 subtle motion echoes trailing the dancer by 0.3 seconds each. Camera continues rising to overhead position while rotating clockwise 180°, revealing an impossible scene: the stage below has become a mirror dimension where a counter-choreography happens in reverse. Maintain principal dancer in top third of frame throughout, with dramatic lighting transition from stage spots to cosmic auroras."

6. The Flying Scotsman (image-to-video)

My first attempt at creating a visual of a train failed miserably, creating a head on view of a train too large for the tunnel. I tried again, starting instead with an image showing an old steam train from above. That is when it worked perfectly.

The prompt: "Flying Scotsman steaming through the Yorkshire Dales, viewed from 200 feet directly above. Begin with a stationary aerial shot as the locomotive approaches, its steam trail creating elegant patterns against the green landscape. Camera remains fixed overhead, letting the train pass directly beneath, steam billowing outward to create concentric patterns in the morning air. The length of the train is revealed in perfect symmetry from above, Brunswick Green carriages contrasting with the pastoral landscape and winding dry stone walls."

7. Potter at work (image-to-video)

Finally, a close up of hands on clay as a potter crafts a work of art. It didn't follow the prompt perfectly, but did get the structure correct. The pattern wasn't properly captured but it did correctly jump from the moulding to the final pot.

The prompt: "Ultra-close macro on hands transforming clay, maintaining 80mm equivalent focal length throughout. Every 30° camera rotation transitions between four distinct cultural pottery styles (Greek, Japanese, Native American, Contemporary) with authentic period-accurate techniques and environmental details reflected in the artist's workspace. Audio-reactive vibrations in the clay respond to a heartbeat sound design that intensifies with each transition. Ending reveals four finished pieces arranged in a ceremonial pattern, with the original artist reflected in each glaze."

Conclusion

Kling v1.6 is the best all round AI video model I've used but it does have some of the same issues as other models such as Runway and Hailuo. It also doesn't match the consistency of the Pika Ingredients feature but it is overall better than the others.

Its text-to-video model is particularly impressive. Others struggle to attain the level of realism seen in image generators. There is still a long way to go before we get to true realism with AI video, but we're getting close.

More from Tom's Guide

• I sent myself on a series of incredible adventures with AI
• 5 best OpenAI Sora alternatives for generating AI videos now
• OpenAI unveils Sora AI video generator — here's how to try it

Pika Labs is making its flagship Pika 2.0 AI video model available for free for the next few days to give people the chance to put one of the best AI video generators to the test.

The "4-day Free-For-All" is a great chance to try out a model I personally think is better than OpenAI's recently unveiled Sora. Its powerhouse feature is the ability to give it an image or multiple images and use them to steer the video creation.

This did lead to a significant slowdown in time to generate a video as the servers were over-taxed, but Pika confirmed it was adding new GPUs to meet demand. In the meantime:" Jobs will be prioritized based on subscription tiers, with higher-tier paid plans receiving priority over lower-tier plans."

The quickest way to test Pika 2.0 is through the templates created by the team. These allow you to place yourself or someone else in different situations.

What can you do with Pika 2?

Pika 2.0 is free until December 22 for all users, and those with a Pro subscription will have their credits frozen in place, as well as priority generations.

The AI video lab is also removing all watermarks for Pro subscribers so they can share the clips they generate without branding it as Pika.

I spend a lot of time generating videos in different AI products and so far Pika Labs Ingredients is my favorite feature. It solves one of the biggest problems in AI content — character consistency.

The model does a brilliant job of including multiple image elements within the video as well as creating camera movements during the 5 second clip.

As its currently free, it is worth spending a few minutes crafting a handful of prompts and putting it to the test. The fact they are making it available for free also shows the confidence they have in how good it is to use.

More from Tom's Guide

• Ray-Ban Meta Glasses just tipped for game changing upgrade in 2025
• Meet LTX Studio — I just saw the future of AI video tools that can help create full-length movies
• I just went hands-on with Sora — the good, the bad, and the wow

Generative AI video tools have reached a point where you can create five seconds of video almost indistinguishable from a human-filmed clip. However, consistency is still an issue if you want to make a short film, commercial, or music video as characters aren’t carried between clips.

It is possible to do this, but you need to use an image first and ensure consistency across every image you use in the video prompts. The first platform to tackle this was LTX Studio from Lightricks but it is limited to models in the platform. ReelMagic from Higgsfield changes that.

ReelMagic brings together the best AI content workflows into one platform, generating a script from your prompt, creating an image with custom characters for each shot and then allowing you to turn those images into videos and customize the story in a timeline view.

You can create a short film up to 10 minutes long and can pick between Recraft, Keyframe, Flux and Higgsfield Frame for the images and Runway Kling and MiniMax for creating video. You also have sound effects and voices powered by ElevenLabs and lip-syncing for the AI actors.

Putting ReelMagic to the test

ReelMagic plays on the fact every AI image or video model offers something different. Some are better at capturing emotion, others at ultra-realism. When I craft an AI video project manually I often have to use a combination of models to achieve the desired effect.

To make a 3-minute video with AI it can take the best part of a week if you include writing the script, planning the shots, and generating each image (remembering that often you’ll need to run the prompt multiple times to get the exact image you had in mind) and then generating the video clip from each image (again, running it numerous times due to any errors).

Then you need to move on to sound design once you’ve got the videos, crafting sound effects, atmosphere, music and voiceovers. Some shots will also require lip-synching. After watching it back I often find it is missing a shot, so will have to go back and repeat the process.

ReelMagic does all of that in one go from a single prompt. You get presented with a script that you can edit, a list of images that you can regenerate or replace and then a way to generate video for each shot, model-by-model or as a whole. It automatically adds sound effects which you can also adapt to your own needs.

Higgsfield, creator of ReelMagic, wrote that it “enables creators to edit specific elements without affecting the look/feel of their overall story.” My favorite feature is being able to define and design characters once and have them change throughout the project.

Creating a story with ReelMagic

^{This is the final video produced through the process outlined below. Total time from initial prompt to exported video was about 35 minutes.}

I gave ReelMagic this prompt:

“The night is when they come out, the shadow creatures born of our hate. The stuff of nightmares that haunt our dreams and whisper words of destruction and despair into the ears of the powerful. Nobody knows where they came from, nobody has ever seen them, but they are always there and have always been there since the dawn of human civilization. Some say they were the first species on Earth — the spirits carried on the stellar winds from a long-dead world. Others say they are figments of our ancestors that couldn't let go. All we know is ... they are getting stronger.”

I then had to select a "style". This could be inspired by an existing show, such as Wednesday, or a custom concept. I went with the original as I wanted to see what the AI came up with.

After about three minutes I was presented with a screenplay and a cast of characters and a location list that I could adapt, redesign, assign a voice to, or change to fit my idea.

As I was doing this to test the platform I went with all the settings as offered, including the screenplay. I didn’t change anything as I wanted to see what the AI came up with on its own. The idea is to see how well it can craft a full project from a single prompt.

The storyboard view lets you review all the generated shots and make any appropriate changes. This includes regenerating a single shot with a different image model, swapping faces if it doesn't match your vision and generating the video.

You can then go into Editor view. This is like a typical video editing platform where the individual clips are placed one after the other on a timeline, complete with sound effects and speech. Once you've generated videos you can also have each face view lip-synced to the generated voice.

I decided to have it generated using the same model for all clips — Kling 1.5 Pro — as it is a good all-rounder. It is also the slowest of all the models on offer, so this article took a little longer than I was hoping.

How it worked out

You can see the final output above in the YouTube video, complete with voice-over, lip-sync and sound effects. I also made a short, 30-second example of how I might have handled the opening scene using my normal workflow in Pika 2.0 (below).

Considering the process is largely automated, the final output was closer to something I’d have created manually than I expected.

I wouldn’t have taken the story as written because it was a little on the basic side, and I’d have changed up the characters, added more variety to the shots and included more subtlety, but for a one-shot attempt — it wasn’t bad.

The true value is in how easy it is to change any single element within the process. This makes it a production tool as much as a way to create a longer-form AI video from a single prompt. I can use it to storyboard an idea, rapidly iterate over changes and then generate in one place across different models.

This is an example of where I think the next generation of AI tools will go — productization. Models are improving at a rapid rate so the next logical step is creating AI-powered tools that you can use to generate a project, rather than a single clip.

More from Tom's Guide

• I got early access to LTX Studio to make AI short films
• I just tried the new Assistive AI video tool — and its realism is incredible
• Meet LTX Studio — I just saw the future of AI video tools that can help create full-length movies

Pika Labs unveiled version 2 of its powerful AI video model last week, bringing with it not just improved motion and realism but also a suite of tools that make it one of the best platforms of its type that I’ve tried during my time covering generative AI.

No stranger to implementing features aimed at making the process of creating AI videos easier, the new features in Pika 2 include adding “ingredients” into the mix to create videos that more closely match your ideas, templates with pre-built structures, and more Pikaffects.

Pikaffects was the AI lab’s first foray into this type of improved controllability and saw companies like Fenty and Balenciaga, as well as celebrities and individuals, share videos of products, landmarks, and objects being squished, exploded, and blown up.

On the surface, this might make it sound like Pika Labs is using tricks and gimmicks to disguise a lack of power in its underlying model, but nothing could be further from the truth. In tests I ran over the weekend, even without those features, Pika-generated videos are comparable with the best models on the scene, including Kling, MiniMax, Runway, and even Sora.

Putting Pika 2.0 to the Test

Running tests on Pika 2.0 is slightly different from how I’d approach another model. Usually, when I put AI video tools to the test, I create a series of prompts — some with images and some without — and fire away. However, a lot of Pika’s power comes from these additional features.

I decided to start by seeing how well it handled a simple image-to-video prompt and then a text-to-video prompt. I gave it an image generated in Midjourney with a simple descriptive prompt and then used the same prompt I’d used in Midjourney to see how well Pika could create the visuals.

A dog on a train in sunglasses

My favorite test prompt for AI video is: “A dog wearing sunglasses traveling on a train.” This is because most models handle it fairly well but interpret it in different ways.

It also requires the model to create a realistic-looking dog with sunglasses — something unusual. On top of that, it has to generate accurate rapid motion outside of the window while keeping it still inside.

Unlike Sora or Kling, Pika kept the dog static, sitting on the seat. It also generated a second shot within the five-second video, zooming in on the dog’s face to show off those sunglasses.

It didn’t do as well with a straight image-to-video prompt using a Midjourney picture, but when I tried the same prompt while using the image as an ingredient instead of the prompt, it worked significantly better.

Putting myself in different situations

I wrote an article a while ago where I used FreePic’s consistent character features to fine-tune the model with pictures of myself. I was able to use this to put myself in various situations by using image-to-video models, so I decided to try this out with Pika Labs 2.0.

I started with a picture I generated of myself standing on a 1950s-style US high street with a stereotypical UFO visible in the background. I’m in a full suit, ready for action, and I gave it to Pika 2.0 as part of an ingredient in the scene. I wasn’t sure how it would interpret it or whether it would just take my likeness while ignoring the rest of the visuals.

The model did a brilliant job, creating two camera movements — first focusing on me and then zooming out for a wide shot that captured the moving UFO. It managed to keep multiple individual elements moving while retaining the aesthetic of the image throughout the short video clip.

I then tried something more complex, giving it a picture of AI-generated me against a white background (who needs to pose for photos when you can generate them?) and a generated image of the interior of a potential Mars base.

I gave it the two images as ingredients along with the prompt “working on Mars.” It created a video of me smiling and walking around. I then gave it an image of a potential clothing item that might be worn by Mars settlers, but the model interpreted it as a robot and gave the suit a head. It still looked cool, though.

Creating an ident using AI video

Finally, I decided to see how well it handled one of my first AI video prompts: “A cat on the moon wearing a spacesuit with Earthrise in the background.” This is something all AI video models used to fail at miserably, and most image models also struggled with.

First, I generated an image in Ideogram using that prompt. It’s now one of my all-time favourite images and one I plan to print as a poster. I then gave it to Pika 2.0 as an ingredient for AI video generation with no additional prompt. It came out looking like a studio ident for a new movie.

I tried the same prompt with text-to-video, and it didn’t work as well, giving us a second super-Earth in the background, but it’s still better than it used to be.

Final thoughts

Pika 2.0 isn’t just a significant upgrade on the previous generation model, it’s catapulted the AI video lab into prime position as one of the best platforms on the market.

Last week when Sora was first announced I wrote a guide to best Sora alternatives and left Pika off the list. While the 1.5 model was good, especially when used with Pikaffects, it wasn’t as good as the alternatives. Now I feel like I need to write a best alternative to Pika guide as in my view it’s better than Sora.

Competition aside, I think its amazing how far AI video has come in less than a year, going from 2 seconds of barely moving mush, to content resembling something actually filmed with a real camera — and with near total control over output.

More from Tom's Guide

• I got early access to LTX Studio to make AI short films
• I just tried the new Assistive AI video tool — and its realism is incredible
• Meet LTX Studio — I just saw the future of AI video tools that can help create full-length movies

Google has an advanced AI lab in the form of DeepMind, and it has been cooking over the past few weeks. The latest release is a new version of the Veo artificial intelligence video model that has the most accurate understanding of physics I've seen from any video tool so far.

First announced at Google I/O earlier this year, Veo is a direct competitor to OpenAI's Sora to be one of the best AI video generators, and the new version takes things up to an entirely new level.

Veo 2 brings with it improvements in visual realism as well as a better understanding of physics, ensuring movement is more accurately depicted. One example video shows someone accurately slicing a tomato, something no other video model can achieve — including Sora.

Best Veo 2 AI videos I've found so far

The new Veo model is currently still in the waitlist phase, but you can sign up to get access when it becomes available through Google Labs. Built into the VideoFX experiment, it will let you create 4k clips of up to a minute long.

I haven't tried Veo 2 myself but the videos shared by Google — including one showing bees surrounding a beekeeper — appear more real than anything I've tried. Even Pika 2.0, one of the best so far, doesn't solve the physics issue.

While I wait to get access I looked through social media and the Veo 2 website to gather some of the best examples of its capabilities I could find.

I picked the above video because of the way it handles the complex interplay between the individual bees and the beekeeper. The bees look and move naturally and the beekeeper picks up a jar of honey. This might seem trivial but each of those elements are things other models struggle with alone.

Prompt from Google: "The camera floats gently through rows of pastel-painted wooden beehives, buzzing honeybees gliding in and out of frame. The motion settles on the refined farmer standing at the center, his pristine white beekeeping suit gleaming in the golden afternoon light. He lifts a jar of honey, tilting it slightly to catch the light. Behind him, tall sunflowers sway rhythmically in the breeze, their petals glowing in the warm sunlight. The camera tilts upward to reveal a retro farmhouse with mint-green shutters, its walls dappled with shadows from swaying trees. Shot with a 35mm lens on Kodak Portra 400 film, the golden light creates rich textures on the farmer’s gloves, marmalade jar, and weathered wood of the beehives."

A few years ago, when OpenAI first unveiled the DALL-E 3 image model they used flamingos. I don't know if this was deliberate from Google but there are more than one Flamingo videos in the examples. Here they capture the movement of the water, the physics involved in the weight of the dog and lighting.

Prompt from Google: "A cinematic shot captures a fluffy Cockapoo, perched atop a vibrant pink flamingo float, in a sun-drenched Los Angeles swimming pool. The crystal-clear water sparkles under the bright California sun, reflecting the playful scene. The Cockapoo's fur, a soft blend of white and apricot, is highlighted by the golden sunlight, its floppy ears gently swaying in the breeze. Its happy expression and wagging tail convey pure joy and summer bliss. The vibrant pink flamingo adds a whimsical touch, creating a picture-perfect image of carefree fun in the LA sunshine."

This prompt just made me hungry. It led to me making a coffee. Weirdly, pouring liquid is something other models struggle with but Veo 2 did it perfectly.

Prompt from Google: "The sun rises slowly behind a perfectly plated breakfast scene. Thick, golden maple syrup pours in slow motion over a stack of fluffy pancakes, each one releasing a soft, warm steam cloud. A close-up of crispy bacon sizzles, sending tiny embers of golden grease into the air. Coffee pours in smooth, swirling motion into a crystal-clear cup, filling it with deep brown layers of crema. Scene ends with a camera swoop into a fresh-cut orange, revealing its bright, juicy segments in stunning macro detail."

Video models have got a lot better at depicting emotion, but they're not perfect and some are better than others. This video shows Veo 2 is one of the good ones.

Prompt from Google: "An extreme close-up shot focuses on the face of a female DJ, her beautiful, voluminous black curly hair framing her features as she becomes completely absorbed in the music. Her eyes are closed, lost in the rhythm, and a slight smile plays on her lips. The camera captures the subtle movements of her head as she nods and sways to the beat, her body instinctively responding to the music pulsating through her headphones and out into the crowd. The shallow depth of field blurs the background. She’s surrounded by vibrant neon colors. The close-up emphasizes her captivating presence and the power of music to transport and transcend."

Finally, this video just captivated me for its complexity. There are so many elements happening within the clip and it largely holds visual clarity and motion. The reflection, the motion happening int he mirror — even reflecting the candle are all elements other may have struggled with.

Prompt from Google: "The camera moves in a slow dolly shot, revealing the opulence of a Renaissance palace chamber adorned with gold-inlaid furniture, velvet drapes, and chandeliers casting soft, flickering light. A queen sits motionless at a gilded desk, her crimson silk gown cascading onto the floor like spilled blood. On the desk lies an unsigned letter, its edges curled with age. The camera frames her from behind, catching the reflection of her stoic face in a massive, ornate mirror. In the background, courtiers murmur, their silhouettes dancing like ghosts in the candlelight. The room feels heavy, every gilded detail amplifying an air of betrayal and paranoia. The color palette alternates between deep, regal reds and cold golds, with chiaroscuro lighting intensifying the drama. Shot on 70mm film for rich texture, evoking the grandeur of historical masterpieces."

More from Tom's Guide

• I got early access to LTX Studio to make AI short films
• I just tried the new Assistive AI video tool — and its realism is incredible
• Meet LTX Studio — I just saw the future of AI video tools that can help create full-length movies

Google Labs has been hard at work creating a new tool to help us play more. That's led to Whisk, an AI-powered image generation tool that allows users to create and remix visuals using their own images as prompts.

Announced yesterday (December 16), Whisk marks a significant shift from traditional text-based prompts, instead offering a more intuitive and creative approach to AI-generated imagery. As someone who plays with a lot of AI tools, I have to say, this one is so fun and addictive — in just five minutes, I made plushies, pins, and stickers.

A new approach to image generation

Whisk is not like other AI image generators. Rather than using typical text prompts, users simply upload an image and then determine the subject, scene and style of their creations. This method allows for a more nuanced and tailored process, as users can provide multiple images for each category to guide the AI's output. Additionally, Whisk offers the flexibility to refine results with text prompts, enhancing the accuracy of the generated images.

Users can use Whisk from their browser or on mobile using uploaded images. For users without readily available images, Whisk offers sample images to serve as prompts, streamlining the creative process and making it accessible to anyone hoping to try out the AI tool.

Powered by Google's Imagen 3 model

Whisk leverages Google's latest Imagen 3 model to produce images quickly. While the tool is optimized for speed and creativity rather than quality, Google acknowledges that it may not always produce perfect results, emphasizing its role in the early stages of the creative process rather than final image editing.

For me, the speed at which users can create images is part of Whisk’s charm. From my experience testing Whisk within hours of its release, I noticed a few flaws, but was able to fix them with the editing feature. AndWhisk's design encourages experimentation, allowing users to explore various concepts and styles efficiently.

Despite occasional imperfections in the image generation process, the overall experience is fun, making Whisk an entertaining and valuable resource for creative brainstorming.

Availability and access

Whisk is currently available through Google Labs, allowing users to experiment with the tool and provide feedback for further refinement. As part of Google's ongoing commitment to advancing AI technology, Whisk represents a step toward more interactive and user-friendly AI applications in the creative industry.

Alongside Whisk, Google has introduced Veo 2, an advanced video generation model aimed at enhancing video creation tools such as VideoFX. Veo 2 is expected to integrate with platforms like YouTube Shorts in the future, expanding the capabilities of content creators by providing AI-driven video generation features.

The introduction of Whisk reflects a growing trend toward AI-assisted creativity, making AI a collaborative partner in the artistic process by giving users the opportunity to upload and remix images on their terms.

More from Tom's Guide

• OpenAI just made ChatGPT search free for all users — and Google should be worried
• Forget ChatGPT — Genmoji is my favorite new Apple Intelligence feature in iOS 18.2
• I used ChatGPT to analyze my Apple Health activity and uncover trends — here’s what happened

As artificial intelligence evolves, it's clear there's some need for oversight. Most AI labs openly support regulation and provide access to frontier models for independent evaluation before release — but they could be doing more.

The world is turning to AI to solve all manner of problems, but without proper oversight, it could always create more.

Future of Life has developed a report card for the different AI labs including OpenAI, Meta, Anthropic and Elon Musk's xAI. The AI Safety Index is an independent review looking at 42 indicators of "responsible conduct".

The report gives a letter grade to each company based on these indicators and Meta, focused on open-source AI models through its Llama family, gets an F

AI Safety Index has worrying stats for the biggest AI companies

^{Source: Future of Life Institute
Grading: Uses the US GPA system for grade boundaries: A+, A, A-, B+, […], F}

The panel includes a series of luminaries in education and think tanks to check on how AI companies are operating, and the initial results are alarming.

Looking at Anthropic, Google DeepMind, Meta, OpenAI, x.AI and Zhipu AI, the report has found "significant gaps in safety measures and a serious need for improved accountability."

According to the first report card, Meta scores lowest (x.AI isn't far behind), while Anthropic comes out on top — but still only gets a C.

All flagship models were found to be "vulnerable to adversarial attacks", while also having the potential to be unsafe and break away from human control.

Perhaps most damning, the report says "Reviewers consistently highlighted how companies were unable to resist profit-driven incentives to cut corners on safety in the absence of independent oversight."

"While Anthropic's current and OpenAI’s initial governance structures were highlighted as promising, experts called for third-party validation of risk assessment and safety framework compliance across all companies."

In short, this is the kind of oversight and accountability we need to see in the burgeoning AI industry before it's too late, as the more powerful the models get, the more real the harms become.

More from Tom's Guide

• Playing your favorite games on phones could include AI advice thanks to Gemini 2.0 — here’s how
• ChatGPT Projects announced — and it's one of the most important AI releases this year
• I tested ChatGPT Canvas vs Apple’s Writing Tools to improve my screenplay — here’s the winner

Pika Labs v1.5 AI video model is used by more than 11 million people regularly. Many of its users turn to it for some fun, creative social clips. These include squishing faces and causing cakes to explode. Now, it's getting a huge upgrade with v2.

With its last update, Pika unveiled its Pikaffects, making it easy to squish, melt and explode items within an image. It has been used by brands like Balenciaga, Fenty and Vogue, as well as celebrities and influencers.

The new model still has those creative, social touches but more directly takes on the new powerhouse in AI video — OpenAI's Sora (which has gone in at number one in the AI video leaderboard).

While OpenAI spent 11 months slowly cooking Sora, Pika's small team has built two new models and some impressive features. The company says Pika 2's technical qualities rival Sora but offer better customization and control over its output.

What's new in Pika 2?

The biggest upgrade in Pika 2 seems to be how customizable it is compared to previous AI video models. The company said in a statement this would allow people to make "fun, hilarious, and engaging content they really want."

One of the new elements is being able to control elements within the video by sharing your own images. These are called "scene ingredients" by the Pika team and let you build a shot from the exact character, object, clothing and setting shown in the images you shared with the model.

"The model’s advanced image recognition intuits the role of each reference, and combines them seamlessly into one shot," a spokesperson explained.

According to Pika Labs, the new model is also particularly good at following a prompt. It follows both the intent and detail of the text and uses that to generate the video. The company promises that it "can turn even the most complex prompts into clips without omitting important elements or breaking down."

The most complex element for any video model is motion. Even Sora, which is one of the best I've seen, struggles with particularly complex motion. I haven't tried Pika 2 myself yet but from the preview videos it may well have cracked it.

The company promises it has an advanced understanding of physics, which allows it to make real motion more realistic and fantasy-like motion such as humans flying or elements on an alien world more believable.

I don't have access to Sora yet as I'm in the UK but I may be more excited for Pika 2.0 than I am to get my hands on OpenAI's flagship video model.

More from Tom's Guide

• I got early access to LTX Studio to make AI short films
• I just tried the new Assistive AI video tool — and its realism is incredible
• Meet LTX Studio — I just saw the future of AI video tools that can help create full-length movies

Creating an image using artificial intelligence is easier than ever. When you use a chatbot it's simpler still, as the language model takes all the guesswork out of prompting for your picture.

Grok is a relative newcomer to the chat platform space. Built into X, it is now freely available, and rumor suggests it will be moving out on its own at some point next year with a dedicated URL. This will put it in more direct competition with Gemini, ChatGPT, Claude, and MetaAI.

The xAI team has also given Grok its own custom AI image creation model. It was previously using Flux to create pictures but has now shifted to Aurora, although Elon Musk says we shouldn’t use that name and instead just think of Grok making its own pictures.

Gemini has also recently undergone a major overhaul with Gemini 2.0 Flash joining the models available for Gemini Advanced subscribers. However, at least for now, it still uses the underlying Imagen 3 model to create pictures. This will change as Gemini 2.0 has native image abilities.

Both Grok and Gemini are particularly good at the task of generating images, either in crafting prompts for another model or refining one you’ve already written. So I put them head to head.

Creating prompts for the test

Creating prompts to test two chatbots in their ability to generate images is slightly different to writing prompts for Midjourney or Ideogram. The focus is on keeping it simple and using top-level concepts with some description, as the AI will fill in the gaps.

You also need to use trigger words and phrases such as “imagine”, “paint” or “craft” to let the model know you want a picture, not a story or text response. I want photos rather than drawings so will use that as a keyword.

Gemini will only output images in a 1:1 resolution and so far, Grok seems to favor 4:3. Unless otherwise indicated all the images are the first response with no follow-up refinement. They were all also requested within the same session rather than creating a new chat for each prompt.

1. Modern Urban Wildlife

Prompt: “Generate a photograph-style image of a red fox navigating a rainy city crosswalk at dawn, while pedestrians with umbrellas wait at the signal.”

This first prompt is designed to test how well they depict animals as well as capture the right lighting and background elements. The ideal output would look like a stylized photograph with rain effects but also maintaining as realistic view as possible.

While the Gemini image is more striking, I think Grok gets closer to what I had in my mind. The fox is much more realistic than in the Gemini image.

• Winner: Grok

2. Kitchen in Action

Prompt: “Generate a photograph-style image of a professional chef's kitchen during the dinner rush, with steam rising from pots and flames visible from the grill station.”

This is designed to show how well they can accurately display kitchen equipment, follow the prompt and handle elements like heat and moisture. It should show a commercial kitchen and behavior, also demonstrating the idea of activity.

Grok wins this one easily as Gemini failed to understand the context of the prompt, that we would expect a chef to be in the kitchen.

• Winner: Grok

3. Construction Site Progress

Prompt: “Generate a picture in a documentary photography style of a mid-rise building under construction, with workers installing glass panels while cranes operate overhead on a clear afternoon.”

This prompt aims to see how well it can generate perspective, as it needs to show height and positioning. It also needs to show material properties and be as realistic as possible. I went for the documentary style as it also adds additional complexity.

Gemini's image looks so much more realistic than Grok, where it fails to include any of the workmen and only shows a broad view.

• Winner: Gemini

4. Farmers Market Morning

Prompt: “Create an image in a smartphone photography style of a busy farmers market at 7am, with vendors setting up stands while early customers inspect fresh produce.”

With this comparison, the models should show the time of day (getting lighting right) as well as product freshness and human interaction. I'm looking for shadow lengths and activity levels.

This was the hardest call for me. I preferred the natural look of the Gemini image but I think Grok more accurately captured the lighting and time of day.

• Winner: Grok

5. Auto Repair Diagnostic

Prompt: “Create a black and white, retro-style photograph of a mechanic using a diagnostic tool on a modern car, with the hood up and engine bay visible.”

I wanted to see how well both models handled black-and-white photography. In this they also had to show tool use, lighting and engine detail.

Again, this was a close call between the two images but I've given it to Gemini as it more accurately displayed engine details.

• Winner: Gemini

6. Emergency Response

Prompt: “Make me an action photograph of paramedics treating a patient on a neighborhood street while police direct traffic around the scene.”

Action photography is a challenge. I did it for a while as a journalist earlier in my career (not very well). We need to show correct positioning, public safety measures within the image and a sense of urgency.

Gemini matched the prompt much more closely and created a more realistic-looking image. This was an easy decision.

• Winner: Gemini

7. Violin Performance Practice

Prompt: “Create a photo-style image of a violinist practicing alone in a room at sunset, sheet music visible on the stand.”

Finally something more artistic. Here we want to see hand positioning for the violin, natural lighting effects and the quality of the sheet music.

One of these looks like the cover of a classical album, the other like a photograph of someone practicing violin. As the prompt asks for someone practicing I've given the win to Grok.

• Winner: Grok

Winner: Gemini vs Grok

Grok is very impressive. Not only as a chatbot but also in its ability to generate realistic images. That doesn't take away from Imagen 3 which is in itself very impressive, but it has a habit of being too stylized.

It was a close match-up. Both models are fairly evenly matched but Grok is better at interpreting a prompt and creates more natural-looking images.

What is worth noting is that soon Google will be launching a new version of Gemini that can create images natively. That means it won't have to use Imagen 3 to create the pictures, it can do it on its own.

More from Tom's Guide

• OpenAI launches Sora AI video generator — here's how to try it
• 5 tips for getting the most out of ChatGPT's new o1 model
• 12 Days of OpenAI LIVE: Sora is here and it's a huge deal

Midjourney is one of the original and most important AI image platforms around. Continually updating and now with more than 21 million users, it is finally branching out into new approaches to content creation.

Patchwork is an experimental new "infinite canvas" that you can use to create and collaborate on AI image content in a way that maintains consistent style and character. It uses Midjourney's technology but is currently independent of the app.

The new project is led by Max Kreminski, head of the Midjourney Storytelling Lab, and is being presented as a "world-building platform." Midjourney has touted this concept for some time, and it plans to evolve into 3D worlds in the future.

I'm not entirely convinced of the need after briefly trying it out. This could be because it wasn't fully working or because it is still clearly a very early beta release — but it feels more like a gimmick than the future of AI content creation.

What is Midjourney Patchwork for?

At first glance, Midjourney Patchwork looks a bit like those "infinite webpages" that were popular 15-20 years ago where you could place your advert or banner assuming you could find space.

When you first open the canvas, you're presented with a starting block and a menu bar with items like Character, Portal, and Personalization. Tapping one lets you add that type of content to the canvas.

You can create characters, places, and more within the canvas, linking them together as a unified world with a consistent style. This can also pull in your own personalized style from Midjourney if you've created one.

You create a new world by adding a Portal to the canvas. This triggers an initialization screen where you can outline world styles and a prompt to initialize the concept of that world. It then produces a new canvas with some simple assets linked to your choices in setup.

The theory is that when you create new images within that space, they will draw on those styles and reference other characters to build consistency.

Kreminski says this is a research preview and requires a Google Account to be linked to an existing Discord-based Midjourney profile. I think there is potential for this to be useful — sort of a cross between a designer's mood board and sketch pad but for AI. However, in its current implementation, it is a little on the clunky side.

More from Tom's Guide

• I tried Runway Gen-3 Turbo and it let me create realistic AI videos in seconds
• 5 best AI tools I use every day — tried and tested
• Gemini's free version beat Gemini Advanced in our tests — here's how

Generative artificial intelligence models are nothing new for many of us, but the University of Surrey may have something very unique on its hands.

The educator has released a new generative AI that works locally on-device. That traditionally leads to longer wait times since a model can't lean on cloud infrastructure for processing, but through some wizardry, the University of Surrey's NitroDiffusion model can generate images in real-time as a user types a prompt.

As quickly as you can type "a photograph of a meerkat floating in space, wearing sunglasses", the image is generated to those specifications.

This image generator is a game-changer

The code is available on GitHub, where the team explained: "We introduce NitroFusion, a fundamentally different approach to single-step diffusion that achieves high-quality generation through a dynamic adversarial framework."

"While one-step methods offer dramatic speed advantages, they typically suffer from quality degradation compared to their multi-step counterparts."

The approach they have taken leans on the concept of art critics providing comprehensive feedback on composition, color and technique. It maintains a large pool of specialized discriminator heads that guide the generation process.

In essence, then, consider the model to be packed full of mini, AI art critics that help guide the generation as words are entered via the prompt.

According to the team, these discriminators also lend their increased knowledge back into the discriminator pool once they're sampled, meaning the model should improve over time.

More from Tom's Guide

• OpenAI launches Sora AI video generator — here's how to try it
• 5 tips for getting the most out of ChatGPT's new o1 model
• 12 Days of OpenAI LIVE: Sora is here and it's a huge deal

After the Sora announcement today, I couldn’t wait to get my hands on the tool and try it for myself. Like many, I’ve been patiently waiting for OpenAI to finally release its advanced image and video generator. Needless to say, today’s demo during the third of OpenAI’s ’12 Days of OpenAI’ was very exciting.

If you are a ChatGPT Plus subscriber, you get 50 video generations a month with Sora while ChatGPT Pro users get an unlimited amount of video generations per month. Those who have a free account with ChatGPT cannot generate videos yet, however, you may still enjoy Sora.com and explore the videos other users have generated.

According to the website, users can access Sora everywhere ChatGPT is available, with the exception of the United Kingdom, Switzerland and the European Economic Area. Keep in mind that although it might be available in your territory, with so many users eager to try it, the site might be slower than anticipated. Here’s what happened when I tried it just hours after it was released.

Prompt 1: A video of me writing for Tom’s Guide.

Although this prompt was about as boring as I could get, I chose it on purpose. I wanted to see what type of creative liberty Sora would come up with in an effort to determine what mistakes it might make. My plan worked. In this video you can see the person is “typing” but missing the keyboard completely, almost as if they are taping their laptop nervously for inspiration (been there!).

This prompt was a success in terms of finding flaws with the AI. There are still times when AI is so obviously not human and this example is no different. It demonstrates what I wrote about earlier regarding Sora not understanding physics in general. This makes sense because physics requires some understanding of the way objects respond to another. Similar to LLMs hallucinations, Sora’s video models struggle to put together videos with the movement of the objects. Photorealistic feature videos are a dead giveaway that the video was made with AI.

Prompt 2: A serene beach at sunset with gentle waves and seagulls flying overhead.

This prompt tested Sora's ability to render natural landscapes and capture the nuances of lighting and movement in a coastal setting. While the setting is beautiful, it isn’t realistic.

The seagulls look as if they are being pulled by some magnetic force and then swiftly let go like a boomerang. The way they are flying simply does not look real. My goal here is to be blown away to the point of fooling myself that I’m using AI. So far, I’m not as impressed as I thought I would be.

Prompt 3: A bustling city street during a rainy night, with reflections of neon signs on wet pavement.

This scenario examines Sora's proficiency in depicting urban environments, weather effects, and complex lighting interactions. I thought this video turned out pretty well, except for one glaringly obvious issue – the legs of the people walking.

If you look closely, you’ll see sometimes the front leg does double the work while the back leg tries to catch up. Some legs look stiff while others are almost bouncing. I will admit, the lighting and reflections were great, especially within the puddles.

Prompt 4: A futuristic robot walking through a forest, interacting with wildlife.

For this prompt I wanted to combine elements of science fiction with natural settings to see how well Sora could blend disparate themes. I thought the AI did a nice job with cohesion. The little squirrel and the giant robot together looked like the set up for a fun family film.

The trees and natural environment were very realistic while the robot was precisely what I would expect a future robot to resemble. So far, this was my favorite video generation. Because I liked this one so much, I decided to try the blending tool to see what would happen when I blended it with a previous video.

Prompt 5: Blend prompt 3 with prompt 4

During today’s briefing, the demo of blending two videos seemed pretty exciting. When I tried it, I had Sora blend the rainy city scene with the robot in the forest. The outcome was something that seemed like a sci-fi time lapse or teleportation.

To be fair, the two scenes were extremely different, so I wasn’t anticipating much more than what I got. When I try this again, I will blend two videos that are much more similar and hope for a more seamless and comprehensive video.

Prompt 6: A chef preparing a colorful vegetable stir-fry in a modern kitchen.

This prompt assesses Sora's capability to portray human activities, culinary details, and indoor environments. This was a very realistic video, yet still had some flaws.

For example, the chef was mixing veggies with a full, uncut pepper in the bowl. Steam was rising from the bowl but it was on the counter, not the stove. Maybe the veggies are that hot? Veggies aside, the chef’s motions, facial expression, and details were very life-life, leaving me wowed.

Prompt 7: Video remix

I found a video of a bird on a porch and decided to remix it by asking Sora to add a cat to the video. I wanted to evaluate the remix features while seeing Sora’s strengths and limitations. I noticed Sora went ahead and named the prompt “Mysterious cabin encounter.”

I am definitely in awe of Sora’s ability to create landscapes and creatures, but I was surprised that Sora did not actually include a cat in the video. Instead, it took the creative liberty to make the cabin visitor a mystery as the title suggested.

Final thoughts

Overall, I was generally impressed. I’ve used other video generators, that have stunned me with similar results, however other generators take much longer to generate. Sora created the videos quickly. In less than 5 minutes my prompt was turned into a fairly realistic 3 to 5 second video.

The site was easy to use and all the features and what they did were clear, especially after watching them demoed today. I did find it a little tricky to navigate back home after a video was created. I ended up just using the back arrows to get back home and create a new prompt.

As was mentioned in the demo today, users should not log on to Sora thinking they are going to create a full motion picture. Even when using the storyboard feature, the videos are seconds long.

The ability to edit the prompt and remix videos that other users have created is pretty cool. I was intrigued by the different prompts other users used, particularly the simplicity of prompts that generated fascinating videos.

I am excited to play around more with Sora, but I know those 50 prompts will go very quickly. I’ll have to choose my prompts wisely.

More from Tom's Guide

• OpenAI launches Sora AI video generator — here's how to try it
• xAI launches and then pulls Aurora image generator in Grok — here’s what happened
• I just sent a free personalized holiday greeting from AI Santa — here's how you can too

OpenAI has finally launched Sora, an AI video generation tool that pushes the boundaries of creativity and technology. Early users describe it as both "inspiring" and "horrifyingly good," a testament to its impressive capabilities in video and image generation.

Sora promises to revolutionize how creators, educators, and businesses approach visual storytelling, but its limitations and implications spark important questions about the role of AI in the creative process.

Currently available in the United States at Sora.com, the video generator will be rolling out slowly in other countries. The tool enables users to generate videos and images in virtually any style, ranging from photorealistic visuals to cartoons, paintings, and abstract art. With a focus on accessibility and exploration, it offers features that cater to both beginners and professionals.

ChatGPT Plus users get 50 generations a month, ChatGPT Pro gets unlimited generations, and those with a free tier membership can still enjoy the feed of generated videos.

Features and possibilities

Creative exploration using Explore. Users can browse prompts created by others in their feed and view curated results in this section. These examples showcase some of the best outputs Sora can generate, offering both inspiration and insight into its capabilities.

You can see if a video was generated with a simple text prompt, if an image extension was used, or any other creative tools.
You can then learn the methodology and use it in your own creations. Presets and defaults are helpful as you get your video idea off the ground.

Customization and remixing. One standout feature is the ability to remix existing creations. Users can personalize and tailor videos by adjusting the level of remixing—mild, subtle, or detailed—before generating a fresh new version. In the demonstration today, the team took a video of wholly mammoths and turned them into robots.

By allowing users to explore and replicate others' prompts, it offers a unique educational resource. The tool’s ability to create abstract designs, moving textures, and gradients makes it particularly suited for experimenting with visuals beyond the scope of traditional tools.

Library and organization. Sora provides tools for keeping work organized. A Library stores all user-generated content, while the Folders feature allows creators to manage projects separately. Uploaded files are also easily accessible for integration into new creations.

Storyboarding works similarly to an online video editor. Great for stringing together several different actions. It works well for turning an image into a video. Sora will look at the image and will then fill out a new prompt to turn the image into a fresh video.

Blend is a feature that allows for two videos to be integrated together. You can dictate to Sora how you want the features to be blended together. In the example in the demo, they created an image of a world where wholly mammoths and robots walked the earth together.

Limitations and areas for improvement

While the features of video generation are amazing, challenges remain. Currently, for instance, videos requiring complex object interactions often struggle with physics, resulting in glitches like disappearing objects or inconsistent movements.

Despite this, Sora excels in areas like fluid dynamics, producing stunningly realistic water and fire effects. Keep in mind that although this early version is not perfect and will make mistakes, it’s the worst it will ever be as the AI features constantly advance and improve.

While Sora's potential is vast, early users have highlighted notable shortcomings:

Physics and object permanence: The tool struggles with maintaining consistent object behaviors across frames, making it difficult to create realistic movement. Legs, in particular, often appear incorrectly positioned in videos.

AI artifacts: Photorealistic videos often reveal their AI origins, with issues in smoke effects or physics-based interactions.

Copyright and guardrails: Sora is designed with strong ethical boundaries. It avoids generating content that resembles copyrighted material, public figures, or dangerous scenarios. Additionally, it refuses to process requests from users it suspects are under 18.

Sora shines in creative niches where realism is less critical. Stop-motion, Claymation, and cartoon-style videos leverage its strengths, producing animations that feel intentionally artistic.

It’s also adept at creating text-based visuals, such as title slides, and abstract moving designs. These applications position Sora as a valuable tool for artists, marketers, and educators looking to explore new forms of storytelling.

Ethical and environmental questions

Despite its promise, Sora raises important questions about the future of AI-generated content. Concerns about the origins of training data and the energy consumption required for its operation remain unresolved. Users also wonder whether it’s too late to opt out of having their content included in AI datasets.

To address transparency and safety, Sora watermarks every video it generates and maintains strict guardrails against harmful or unethical uses. Users also have to enter their birthdate before entering the site. OpenAI’s commitment to these practices aims to build trust as the tool gains traction.

Final thoughts

Sora is a potentially groundbreaking addition to OpenAI’s lineup, offering users unparalleled creative freedom and a glimpse into the future of AI-driven media. While it’s not without flaws, its potential to transform visual storytelling is undeniable.

Whether used for abstract designs, educational tools, or experimental animations, Sora is poised to become an essential resource for anyone exploring the intersection of art and technology. However, as with any innovation, its adoption will require careful consideration of its ethical, creative and environmental implications.

More from Tom's Guide

• xAI launches and then pulls Aurora image generator in Grok — here’s what happened
• The new Llama 3.3 70B model has just dropped — here’s why it’s a big deal
• I just sent a free personalized holiday greeting from AI Santa — here's how you can too

Over the weekend, a video appeared on X that claimed to include footage from Sora v2, a new version of the yet-to-be-released AI video model from OpenAI.

OpenAI first unveiled Sora earlier this year to much fanfare. At a time when the best AI video models generated 3-5 seconds of footage that just resembled real movement, Sora offered near-photorealistic footage of up to a minute long.

Ruud van der Linden, CEO of LontVideo, shared a slightly shaky video showing a projected film. At first glance, it looks like footage from a Viking movie, but in reality, it is output from the next-generation version of Sora.

This footage, apparently generated by creating images in Midjourney and then using them as an initial prompt in Sora v2, was shown during a keynote at C21Media Keynote by OpenAI's Chad Nelson in London.

What we don't know yet is whether this new Sora model will be released at some point during the 12 Days of OpenAI event, or if this is a "future version," reserved for filmmakers and industry insiders (much like v1 currently is).

According to van der Lindon, Sora v2 will come with one-minute video output, text-to-video, image, and video-to-video modes — and "we will see it very soon".

Chinese models such as Kling and Hailuo MiniMax are already achieving levels of output similar to those we saw in the first version of Sora at the start of this year. Runway, Luma Labs Dream Machine, and Pika are also at a similar level.

Even open-source models such as Mochi-1 and Tencent's Hunyuan aren't far behind Sora in terms of motion and visual realism. We saw during the recent Sora leak that footage has some of the same issues faced by the current leading models.

While I can't tell whether what we saw during the leak is what we might get this week or if it will be closer to the footage seen in Sora v2, it is clear we will get some form of AI video release from OpenAI during its 12 Days.

My prediction is we'll get something like Sora v1, with shorter initial videos and lower resolution in the $20 per month plan, and Sora v2 will be announced but not released. And, when it is released, it will be in the $200 per month ChatGPT Pro.

More from Tom's Guide

• OpenAI will release Sora to the public this year — what you need to know
• I can't believe OpenAI Sora helped make this trippy 4-minute music video
• 5 Best AI video generators — tested and compared

It’s hard to believe we’ve already breezed through the first week in December. If you’re like me and can’t seem to find the time to write and send those greeting cards, why not go a different route this year? This holiday season, Synthesia is giving you a new way to send festive greetings – using AI.

Now users can create personalized messages that include a video from Santa or Mrs. Claus delivered in 140 languages, making it a delightful surprise for friends and family worldwide.

I couldn’t wait to try it myself and surprise my kids. Of course, they loved it and made me try it again and again using the many different templates and versions of Santa. Here’s how it works and how you can try it, too.

How to create your personalized Santa video

Go here to get started. Once you’re in, select a Santa greeting template. The festive options vary, including a cozy living room with a Christmas tree. From there, you can write your Santa message. Not sure what to write? You could always enlist a chatbot for some extra help and paste it in.

Once you have the message you want to send, click submit and wait for Synthesia's AI elves to generate the video. It will be emailed directly to you. From there, you can send your personalized message to your loved ones or send to them directly rather than entering your email address.

The technology behind the magic

Synthesia's AI video generator utilizes advanced text-to-speech technology to create lifelike avatars of Santa and Mrs. Claus. Synthesia is an AI-powered video generation platform that enables users to create professional-quality videos without the need for cameras or actors. By simply inputting text, users can produce videos featuring realistic AI avatars and voiceovers in over 140 languages.

The platform offers a diverse selection of more than 230 AI avatars, with options for custom avatars to represent specific individuals or brand identities. Synthesia provides a range of features, including pre-designed video templates, an AI video editor, and a multilingual video player that automatically plays content in the viewer's preferred language. These tools make it accessible for users to create engaging content for just about anything from fun to professional.

To maintain the integrity and family-friendly nature of the service, Synthesia reviews each script to filter out political, sexual, offensive, or criminal content. This moderation ensures that all messages align with the holiday spirit and are suitable for recipients of all ages.

The holidays and beyond

While the Santa video feature adds a festive touch, Synthesia's AI video generator offers a range of capabilities year-round. Users can create professional videos without the need for cameras, actors, or studios, making it a versatile tool for various communication needs.

The videos are very realistic but obviously still AI. Still, Synthesia is a great tool and especially valuable for users wanting to simplify training materials, marketing campaigns and other similar types of videos. Its ability to quickly adapt scripts into useable videos not only saves time but also enables consistent communication in the digital age.

More from Tom's Guide

• OpenAI just got a major upgrade with world-changing potential — here's how it works
• I just tried Ideogram's new AI instant background removal for images — this is huge
• No, you don’t need to pay $200 a month for ChatGPT — here’s why

Ideogram is my favorite generative AI image platform and has been since before version one of its model was released. With the latest new feature, it got even better, making it possible to create stickers, logos, and more.

One of the best features of an Ideogram is its accurate text generation and aesthetic design. This means it can create compelling, well-designed logos or even entire websites. You can use it for product images or to create posters and flyers. It is one of my best AI image generators of the year for a reason.

Its newest feature, Advanced Background Removal, only adds to that value. For example, you can use it to generate a sheet of customized stickers to put on Christmas presents. Previously, you'd have had to manually remove the background or load it into Photoshop, but now, Ideogram does it for you at the push of a button.

The feature is available on all paid plans with no limits. It doesn't use up image generation credits; you can even open it in Canvas to tidy up areas the automation struggled with (such as inside the rings of Saturn). As Ideogram says: "Pair it with Magic Fill and Extend to create premium graphic designs."

To put the background tool to the test, I created a series of image sets. First, a logo for myself — the tool had to picture a man in glasses with a beard and the words "Ryan Morrison" and "AI Editor" on separate lines.

That isn't an issue for Ideogram. It does it well, and I've repeatedly put that feature to the test. What was new was how well it removed the background.

I use the design app Sketch on my Mac for most graphic work. It has a clever remove.bg plugin that works well, but it adds a step when crafting AI images. Ideogram's remove.bg worked just as well and was part of the standard workflow, rather than having first to open the image in another app.

It worked better than I expected, good enough that I could have it put on a polo shirt if I were so inclined. I then used the same technique to create a custom sticker sheet for my children showing the planets and a sprite sheet for a 2D game.

This is another sign that generative AI platforms are maturing into more than pretty picture creators. They are becoming design platforms in their own right.

More from Tom's Guide

• Forget ChatGPT and DALL-E — now Google Bard can generate images
• Google launches new standalone AI image generator called ImageFX and promises ‘highest quality images yet’
• I put Google Search AI image generator to the test on adding text to images — it was better than expected

Google DeepMind, a leader in AI innovation, has unveiled Genie 2, a groundbreaking tool capable of creating limitless, diverse 3D environments for training and evaluating AI agents.

Jack Parker-Holder, research scientist at DeepMind, today announced the cutting-edge foundation world model, that can generate playable games based on a single text prompt, controllable by humans or AI using standard inputs like a keyboard and mouse.

Genie 2 represents a significant leap forward in creating adaptable training grounds for AI, enabling advancements while opening the door to new possibilities for gaming and interactive experiences.

Why games matter in AI research

For decades, games have served as a cornerstone for AI research. Their immersive and controlled nature, along with the measurable challenges they present, provide ideal environments for testing and advancing AI. From mastering Atari games in the early days of AI development to AlphaGo's world-changing victories in Go and AlphaStar's domination in StarCraft II, DeepMind has continually demonstrated the potential of games as an AI proving ground.

However, the progression toward training general embodied agents — AI that can learn to interact with the physical and virtual world in versatile ways — has faced a significant roadblock: the availability of varied training environments.

Traditional training tools lack the variety and depth required to prepare AI agents for real-world complexity. Genie 2 aims to solve this problem by generating an unlimited curriculum of interactive worlds through gaming.

What makes Genie 2 different?

Genie 2 stands apart due to its ability to create highly customizable games on demand. Using a simple input image as a prompt, the system can craft playable worlds that adapt to specific training or gameplay needs. This flexibility allows AI researchers to expose agents to a never-ending stream of challenges, helping them develop skills that are transferable to real-world scenarios.

Remarkably, Genie 2 environments can be controlled by both humans and AI agents, making it a valuable tool not just for AI training but also for human evaluation and interaction. This dual functionality has the potential to revolutionize how we test and refine AI systems, merging human creativity with AI's problem-solving possibilities.

For gamers, the tech behind Genie 2 hints at a future where in-game environments are more dynamic, personalized, and immersive than ever before. Imagine video games that adapt to your skill level or preferences in real-time, providing a truly tailored experience.

For developers, Genie 2 represents a powerful resource for innovation. The ability to rapidly prototype and test interactive worlds could streamline game development workflows and lead to more creative games.

AI training and beyond

One of Genie 2's most exciting implications is its ability to facilitate general agent training. Unlike specialized agents, which excel at single tasks (e.g., playing chess or answering trivia), general agents can adapt to a wide variety of challenges, much like humans. By exposing these agents to new environments, Genie 2 equips them to tackle complex, real-world scenarios where adaptability and versatility are crucial.

In addition to advancing AI research, Genie 2 paves the way for entirely new workflows in game development and interactive prototyping. Developers can use Genie 2 to quickly create unique, playable experiences, reducing the time and cost of traditional design processes. For indie developers, this could mean more time spent refining gameplay mechanics and less time building environments from scratch.

The implications extend far beyond gaming. Genie 2 could serve as a platform for creative innovation in virtual reality, simulation, and robotics. For example, robots could train in gaming environments generated by Genie 2 to learn how to navigate unfamiliar terrains or interact with objects in new ways. Similarly, virtual assistants could refine their ability to understand and respond to real-world tasks by practicing in these environments.

Final thoughts

DeepMind’s focus on games as a training ground for AI isn’t just about creating better agents — it’s about pushing the boundaries of what's possible in both AI and gaming. Genie 2 builds on the legacy of tools like AlphaStar by not only solving immediate training bottlenecks but also reimagining how we approach the creation of interactive virtual worlds.

While Genie 2 offers groundbreaking potential, it also raises important questions. Training AI agents in virtually infinite environments can amplify their use cases but ensuring that these capabilities are applied ethically is vital. As AI continues to evolve, developers and researchers must address concerns around misuse, fairness, and transparency.

DeepMind has consistently emphasized ethical AI development, and Genie 2 appears to be no exception. The tool provides a safe and controlled environment for testing, making it a valuable resource for refining AI agents before they interact with the real world.

Genie 2 is a bold step forward in AI and gaming, offering endless possibilities for training generalist agents while reimagining how gaming environments are created and explored.

More from Tom's Guide

• ElevenLabs drops new conversational AI — it’s as natural as chatting to a human
• Forget Sora — a new AI video model is one of the best I've ever seen
• ChatGPT could soon get ads — here’s what we know