One of my favorite uses of AI is to generate images. Whether that means creating robot diaromas or katana-wielding samurai, I find the creative use cases to be endless. A popular AI image generator is Recraft, which brands itself as an AI image generation tool for designers. In principle, it works much like other AI image generators on the market, but I was curious whether its extra features could benefit more than just casual hobbyists.

Founded in 2022, people have used Recraft to generate over 400 million images. The platform includes comprehensive instructions for users at any skill level. It allows you to export your designs in popular formats, including PNG, JPG, and PDF. Despite pitching itself as a service to designers, Recraft says anyone can use it to create or edit their own AI photos. This all sounded promising, so I thought I’d try Recraft myself. Here's how you can us Recraft to create your own AI images.

Final thoughts

Recraft was fun and easy to play around with, but with plenty of other AI image generators to choose from, it’s wise to shop around to determine which one offers the features that suit your needs the most.

While Recraft did not make the cut for being one of the best AI image generators of 2025, these are the ones that did. If you’re looking for an AI image generator that offers a dedicated environment for creating and editing AI images and icons, Recraft offers exactly that.

While the free tier will give you some room to experiment, you’ll chew through your free credits quickly. To get consistent, high quality results, you’ll need to unlock the more powerful models, which will set you back some bucks each month.

More from Tom's Guide

• How to stop Meta accessing your camera roll without permission
• This hot new AI job requires zero experience — and anyone can do it
• ChatGPT-5 claims to be a superior writing assistant — I tested it to find out

Google Vids, the Workspace app designed to make video creation faster and easier, is getting a major update today (August 27). The company is rolling out new AI features powered by its Veo 3 video model, bringing its AI avatars and transcript tools to general availability, and introducing a no-cost consumer offering for the first time.

Alongside the product updates, Google is also launching a new “Vids on Vids” instructional series, designed to help anyone, from individual creators to business teams, get the most out of the platform.

New AI features in Google Vids

• Image to Video: Using Veo 3, users can now transform any static product shot or branded image into a short, animated video clip with motion and sound. A simple text prompt can turn stock photos into polished, eight-second videos.
• AI avatars: First previewed at Google I/O, AI avatars are now widely available. With a simple script, you can generate a consistent, professional presenter to deliver training, demos or internal announcements; no camera required.
• Automatic transcript trim: This feature automatically removes filler words and awkward pauses from recordings, streamlining editing and making presenters sound more confident.

Google says additional tools are on the way, including noise cancellation, filters and new video sizing formats for portrait, landscape and square.

Free consumer version

For the first time, Google is making Vids accessible to everyone with a new basic editor, rolling out globally today. This version excludes Gemini-powered AI features but includes core editing tools and templates, making it useful for tutorials, event invites or quick social clips.

To help users get started, Google is releasing a new instructional series on YouTube. The first episode covers the basics of the recording studio and how to plan a video story using AI. Future episodes will expand into animations, transitions and other advanced features.

Bottom line

More from Tom's Guide

• I Asked ChatGPT vs Gemini vs Claude to Write a Resume — Here’s the One That Got It Right
• I Built 5 Websites in Under an Hour with ChatGPT-5 — Here’s How it’s Possible
• These 9 Vibe Coding Prompts Are Pure Genius — And So Easy To Use

Google’s Gemini app is getting a powerful new boost: the "nano banana” image editing model offers advanced AI image editing capabilities that keep people (and pets) looking like themselves, even when you completely change their style, background or setting.

The update, rolling out today for both free and paid users, integrates a new DeepMind model that’s already earning top marks in early testing. Now users can preview their living room with a new color of paint, drop themselves and their pets into a fun scene or get a look at their future self.

These latest tools give users much more control over the images they bring to life and the stories they tell with each picture.

Editing that keeps 'you' consistent

One of the biggest challenges in AI photo editing is maintaining likeness. The small inconsistencies can add up, making an image feel “off.” Google says this upgrade focuses on keeping people’s identities and features consistent across edits. This means that users can have more fun with their images without sacrificing their "selfie." For example, users can try on new haircuts, give their pets silly clothing or even swap outfits without losing details that make images unrecognizable.

Gemini also supports uploading multiple photos so you can combine them. For example, blend your selfie with a shot of your cat to generate a new portrait of you both on a road trip.

Multi-turn editing and style mixing

Users can now make step-by-step changes to the same photo. Start with an empty room, wallpaper the walls, add furniture, decor and paint until it feels complete. Gemini allows for “multi-turn editing,” meaning you can tweak specific parts of an image while leaving everything else intact.

There’s also a new design mixing feature that lets you apply the style of one image to an object in another. Think flower petal textures on rain boots, or a dress patterned after butterfly wings.

Global rollout today

Both free and premium Gemini users worldwide can start experimenting with these new tools today. Every generated image will include a visible watermark as well as Google’s SynthID digital watermark to mark it as AI-made.

With this update, Gemini is raising the bar for how consistent and flexible AI-powered editing can be.

More from Tom's Guide

• Anyone Can Become a Vibe Coder — Here’s How AI Makes It Easy
• ChatGPT Is Stealing Wikipedia’s Readers — What It Means For You
• Vibe Coding is Blowing Up — 9 Genius Prompts To Try First

OpenAI is set to release a new version of its flagship AI video model Sora sometime this quarter. While revolutionary at launch, Sora has since lost ground to competitors, with Google's Veo 3 now setting the gold standard for AI video generation.

I expect Sora 2 will arrive in the coming weeks, maybe months given the rapid release of GPT-5. Like GPT-4o, GPT-5 is natively multimodal, handling any input or output type (including video) while performing complex reasoning tasks similar to the "o" series models.

Sora remains a solid platform. Its Storyboard feature breaks new ground, and ChatGPT Pro subscribers can generate clips up to 20 seconds long. But the underlying model is showing its age. The output still suffers from motion control issues, lacks sound generation, and struggles with complex physics rendering — unlike Veo 3, Kling 2.1, or MiniMax 2.

Even in the social video space, OpenAI now faces competition from virtually every AI platform, including Meta, Grok, and Midjourney. Yet OpenAI remains the world's largest AI lab with significant resources and — despite Meta's recent talent raids — a formidable engineering team. Don't count them out just yet.

What OpenAI Needs to Make Sora Competitive

To compete with Google's video model or the emerging Chinese competitors, OpenAI must leverage its multimodal capabilities while expanding Sora's feature set. Tighter ChatGPT integration wouldn't hurt either. Here are five essential improvements for Sora 2:

1. Native sound generation is non-negotiable

If OpenAI wants to compete with Veo 3, Sora 2 must handle both video and audio natively. Any model without sound generation starts at a disadvantage.

Currently, Sora produces only silent clips — a significant weakness when Veo 3 generates sound effects, ambient noise, and even dialogue as core features. This isn't just about tacking on audio as an afterthought; it's about true integration.

Veo 3 can produce lip-synced character speech in multiple languages. Sora 2 needs that same built-in audio capability — from atmospheric soundscapes to spoken dialogue.

If OpenAI delivers fully multimodal generation (video + audio) while maintaining 20-second clips or longer, it won't just catch Veo 3 — it could leap ahead entirely.

2. Physics simulation must improve dramatically

Visual realism extends beyond resolution — it's fundamentally about physics. Current Sora outputs often display unnatural motion or distorted physics: water that defies gravity, objects that morph unpredictably, or movement that feels fundamentally wrong.

Google clearly prioritized real-world physics with Veo 3, and the results speak for themselves. Their videos excel at simulating realistic physics and dynamic motion with minimal glitches. Meanwhile, Sora's older model produces jittery movement and inconsistent object interactions that shatter immersion.

For Sora 2 to compete, its model must better understand real-world behavior — from natural human gaits to bouncing balls, from smoke dynamics to fluid mechanics. OpenAI essentially needs to integrate a physics engine into Sora. Believable motion and interactions (no more warping limbs or melting backgrounds) would close a critical gap with competitors.

3. Conversational prompting should be the standard

OpenAI's ace in the hole? ChatGPT has already trained millions to communicate conversationally with AI. Sora 2 should capitalize on this by making video creation feel like dialogue, not programming.

Rather than demanding perfect prompts or complex interface navigation, the system should support natural back-and-forth refinement. Google's already moving this direction — their Flow tool uses Gemini AI to enable intuitive, everyday language prompting.

Runway does this brilliantly with its chat mode, and now the new Aleph tool that allows Gen-4 to expertly refine any single element. Luma’s Dream Machine was built with this concept from the ground up.

Picture this workflow: Type "medieval knight on a mountain," receive a draft video, then simply say "make it sunrise and add a dragon" — and Sora instantly updates the scene. This conversational approach would lower barriers for newcomers while accelerating workflows for professionals.

The technology exists. ChatGPT already interprets follow-up requests and adjusts outputs dynamically (as demonstrated with GPT-4os native image integration). Sora 2, fully integrated with ChatGPT, should let us talk our way to great videos. That user experience would outshine the technical prompting most competitors still require.

It would also allow for native image generation first, then animation using Sora, similar to the way Google is working with Veo 3 in Gemini or the new Grok Imagine feature.

4. Character consistency and customization are essential

Character and scene consistency represents another crucial improvement area. Currently, generating two clips of "a girl in a red dress" might produce two completely different people. Sora's outputs drift in style and details between generations, making coherent multi-scene stories or recurring characters nearly impossible.

Sora 2 must enable consistent characters, objects, and art styles across longer videos or clip series. Competitors already deliver this — Kling 2.1 boasts "consistent characters and cinematic lighting straight from text prompts." Google's Flow goes further, allowing custom assets (character images, specific art styles) as "ingredients" across multiple scenes.

OpenAI should provide similar capabilities: reference image uploads, style fine-tuning, or character persistence across scenes. If Sora 2 can maintain consistent character appearance throughout a video, creators can actually tell stories instead of producing disconnected clips. Especially if it has native audio integration over 20 second clips.

Consistency and customization work together — whether you're an artist maintaining a signature style or a filmmaker needing character continuity, Sora 2 should deliver that control.

5. Deep ChatGPT integration and universal access

Finally, OpenAI should maximize its ecosystem advantage by deeply integrating Sora 2 into ChatGPT while ensuring broad accessibility. Google's Veo connects to a broader toolkit (Gemini integration, API access, Flow app), and Meta will inevitably embed AI video throughout their products.

OpenAI can differentiate by making Sora 2 a seamless ChatGPT feature. Applying this approach to Sora 2 would instantly give millions of ChatGPT users an AI video studio without switching apps. They could follow Google and have a low limit on videos per day with a premium plan for unlimited access — as they do now with ChatGPT Pro and Sora.

Mobile optimization is crucial. Today's creators shoot, edit, and post entirely from phones. If Sora 2 works within ChatGPT's mobile app (or a dedicated Sora app) with quick generation capabilities, it could capture the TikTok and Reels creator market. Imagine telling your phone, "Hey ChatGPT, make a 15-second video of me as a cartoon astronaut landing on Mars," and receiving share-ready content instantly.

By making Sora 2 ubiquitous — through ChatGPT, developer APIs, and mobile platforms — OpenAI can rapidly build its user base while gathering essential improvement feedback.

Platforms like Leonardo, Freepik and Higgsfield are already heavily using Google’s Veo 3 and Hailuo’s MiniMax 2 because they’re impressive, fast and available via API — OpenAI is falling behind in the creative AI space by not updating Sora.

The bottom line

OpenAI has a genuine opportunity to reclaim leadership by learning from competitors' successes. Google's Veo 3 currently sets the benchmark with native audio, realistic physics, and strong prompt adherence, while emerging models like Kling 2.1 and MiniMax 2 continue pushing boundaries.

Runway races ahead with new refinements to its Gen-4 model — which is of a similar quality in terms of physics to Sora but with more features, and others like Pika are focusing on the creator market, further pushing OpenAI out of the valuable space.

Sora 2 can't be just an incremental upgrade — it needs to amaze.

The encouraging news? OpenAI has the foundational elements: a powerful language model, a first-generation video model to build upon, and ChatGPT's massive user base. If OpenAI delivers native sound generation, realistic physics, conversational ease-of-use, character consistency, and seamless product integration, Sora 2 could very well beat Veo 3, Kling, and the entire field at their own game.

When it all comes together, don't be surprised if the next viral AI video in your feed was created with Sora 2.

More from Tom's Guide

• OpenAI just made GPT-5 'warmer and friendlier' with a big personality update — and there's already a backlash
• GPT‑5 feels confusing at first — these 5 expert tips make it feel smarter than ever
• I asked ChatGPT-5 to take me behind the scenes of its brain — this is what I learned

If your main priority is a functional design at a great price, simply keep your eyes on the deal sections on Tom’s Guide so you don’t miss out on news such as bargains from the likes of Carhartt, Columbia, and Lululemon.

On the other hand, if you’re feeling creative and want to create your own clothing designs, AI image generators can help you to instantly visualize your ideas. Experiment with different styles, patterns, and color palettes to create unique pieces of clothing that could potentially make the jump from your screen to your wardrobe.

I’m no fashionista, but I do appreciate an outfit that helps you look good without trying too hard. So I thought I’d ask two of my go-to AI image generators, Sora and Leonardo, to design five different t-shirts and see what they could come up with.

For the purpose of this article, Leonardo’s creations will appear on the left, while Sora’s designs will be shown on the right.

1. Logo t-shirt

I wanted to start things off easy with a proof-of-concept prompt. Could both Sora and Leonardo design t-shirts and generate decent images featuring their own designs? There was only one way to find out, so I asked both AI image generators to design t-shirts featuring their company logo that a model could then wear.

First up was Sora, which produced a clean image of a model wearing a black t-shirt. The OpenAI logo was very much front and centre of the design. Everything was very standard, which is the overall look I was going for with this first test prompt.

I was expecting to feel this standard vibe again once I’d seen Leonardo’s creation, but it must have gotten lost along the way. Instead of Leonardo’s logo I got a typo-riddled t-shirt.

Admittedly, not a great start for one of my favorite AI image creators, but on the other hand, this was just the first trial. Time to see if Leonardo could catch up.

2. Vintage style

No t-shirt creating contest would be complete without a retro round, so Leonardo and Sora had to go back to work to create a t-shirt that had a vintage look.

Sticking to the same order as in the test run, Sora got to show off its creation first. I wasn’t expecting a t-shirt to literally feature the word “retro” in it, but that’s what Sora decided was needed for the job. Wonky text alignment aside, I got a pretty pleasant-looking tee with a red stripe running around the collar and sleeve hems.

Interestingly Leonardo also stuck to the red stripe concept, albeit limiting it to the collar. The result wasn’t too bad, but it felt like Sora invested a bit more effort into its creation.

3. Bold text

What if I asked the AI image generators to create a t-shirt that was all about bold text?

I didn’t want to create any restrictions around what specific text or slogan they should use, so I simply used the same prompt asking for a stylish t-shirt featuring unapologetically bold text.

I started sensing a pattern as Sora created its image of a t-shirt with the text “unapologetically bold” printed on it. In a second attempt, I specified that “the text should look unapologetically bold”. This resulted in Sora keeping things exactly the same, except this time, it inserted the word “females” in front of the other two.

One would think that this time the bar was low enough for Leonardo to clear with ease. I was rooting for it to make a comeback, but was once again left with a typo. In Italian and Spanish “ameno” refers to something being pleasant, which I’m guessing is what Leonardo was going for?

We could call this one a draw.

4. Classic look

Going back to the basics, I crafted a prompt for a “classic striped t-shirt”. Sora quickly crafted a basic t-shirt that would suit most wardrobes.

Leonardo pulled out all the stops and went for stripes in different shades, which actually turned out rather well. It wasn’t a perfect result as some finishing errors sneaked their way in, but I think it deserved a few extra points for creativity.

5. Freestyle

All constraints were removed for the final challenge. Could both Sora and Leonardo design a t-shirt from scratch based on any style of their choosing?

Their task was to create a t-shirt for a model that was unique but still stylish.

In seconds, Sora crafted a spiral motif in a bold red, paired with abstract flames with a golden tone emerging from the center.

While Sora’s design lacked true character, its design was more exciting than the muted creation by Leonardo. The latter opted for a grey crew neck dotted with faded designs that to me look half-floral half-fruit from a distance.

Final thoughts

Sora edged out a narrow victory, producing the designs I’d be most likely to wear. Leonardo did its best to keep up, but it just wasn’t its day.

While the two AI image generators failed to truly impress in autonomously creating their own concepts for t-shirts, they definitely could be useful tools to deploy to render specific designs you have in mind.

Just keep in mind you might get better results if you upload a reference photo and spell out what you’d like the final t-shirt to look like.

More From Tom's Guide

• You can reverse image search using ChatGPT — here's how
• How to spot AI writing — 5 telltale signs to look for
• I used Perplexity AI to plan my next trip — here's how you can too

You probably know Alexa as the voice that sets timers, dims the lights and plays music, but after testing Alexa+ for a few weeks, I’m discovering new features nearly every day.

Recently, I discovered that upgraded Alexa has a creative side and can now generate realistic images on the fly. In my testing, I was impressed by the speed and ease of generating the images, that I couldn’t help but do a side-by-side image generation comparison with ChatGPT.

Whether you ask for a watercolor of a cozy cabin, a photorealistic puppy or a design mockup, Alexa+ responds in seconds, then texts the image directly to your phone. It’s surprisingly intuitive, shockingly fast and potentially a game-changer for everyday users.

Here’s what happened when I tested Alexa+ versus ChatGPT and the images each bot generated.

1. Robot family having dinner

Prompt: “A friendly robot mom, dad and two robot kids sitting at a kitchen table eating spaghetti, surrounded by retro 1950s-style decor, with warm lighting and a dog robot under the table.”

ChatGPT automatically went with a cartoonish image of a robot family. It nailed almost everything except for the robot dog and opted for a realistic-looking dog instead.

Alexa+ presented a modern-looking image of a robot family also with a “real” dog. It missed aspects of the prompt such as “two robot kids” and the style of the kitchen.

Winner: ChatGPT wins for accurately following the prompt.

2. Sunday morning at the local farmers market

Prompt: “People browsing colorful fruit and vegetable stands at an outdoor farmers market. A woman holds a bouquet of sunflowers, a man samples cheese, and kids eat popsicles. Background includes food trucks and string lights.”

ChatGPT once again hit every detail of the prompt, even if it’s hard to tell the man is sampling cheese.

Alexa+ offered a more diverse and more realistic look at a farmer’s market, though it’s not completely obvious if the man in the image is sampling cheese.

Winner: Alexa+ wins for an image that better captures the energy and diversity of a farmer’s market.

3. Rainy school day

Prompt: “Cars lined up outside an elementary school on a gray rainy morning. Parents holding umbrellas hustle their kids to the entrance. Backpacks, rain boots, and puddles all around. View from a car window with raindrops.”

ChatGPT created an image that is storybook and highlights the mood of a rainy school day. The backpacks in the puddles is a glaringly unrealistic mistake.

Alexa+ delivered a photorealistic image that also captures the mood, but looks less like it is a view from a car and doesn’t capture the hustle that the prompt asks for.

Winner: ChatGPT wins for following the prompt and better storytelling with this image.

4. Dog and toddler in a cozy living room

Prompt: "A golden retriever sitting next to a toddler on a cozy living room rug. The toddler is offering a cracker, and the dog gently takes it. Toys and a sippy cup are scattered around. Natural window light, soft and heartwarming."

ChatGPT used soft, warm lighting in an image that evokes a heartwarming, almost storybook-like feel, which best fits in line with the prompt request “soft and heartwarming” descriptor.

Alexa+ made a great photo that looks very natural and realistic, but it’s missing details such as the sippy cup and the dog is not taking the cracker. While it may arguably be a better image, it loses because of missed aspects of the prompt.

Winner: ChatGPT wins for the best composition match, setting and details.

5. Busy crosswalk in the city

Prompt: "A diverse group of people crossing a city street in the early morning — a woman in heels with a coffee, a jogger with headphones, a dad pushing a stroller, and a teen on a scooter. Background includes traffic lights, brick buildings, and steam rising from a manhole. Overcast sky, everyday realism."

ChatGPT captured all of the details within the prompt but delivered a messy and cluttered image that does not look real and one I wouldn’t use for anything.

Alexa+ delivered a polished image that highlights the prompt, but misses several details.

Winner: Draw. ChatGPT hit every detail of the prompt, but at the cost of an unrealistic crosswalk situation. Alexa+ made an image that better captures the vibe of a busy city but misses out on key details of the prompt.

Overall winner: ChatGPT

After five rounds of head-to-head testing, one thing is clear: ChatGPT might take a more whimsical or storybook approach, but it consistently nails the specifics.

However, I was pleasantly surprised by Alexa+. It generated refined images that often looked more realistic than ChatGPT that are instantly shareable. And it was faster than ChatGPT, too. But in most cases, Alexa+ fell short when it came to following the actual prompt.

Whether it’s missing a sippy cup, skipping key characters, or glossing over a specific setting, the details matter, especially when you're generating visuals with a purpose.

So, while Alexa+ has potential and impressive speed, it still loses out to ChatGPT when prompt accuracy counts. My suggestion to users is to use Alexa+ for speed and realism, but be ready to tweak any resulting image with follow up prompts.

More from Tom's Guide

• AI’s energy problem is exploding — so why are tech giants turning to poop and underwater bunkers?
• I test ChatGPT for a living — use this ‘3-word rule’ to get smarter answers
• OpenAI CEO Sam Altman warns AI will wipe entire job categories off the map

I've spent hundreds of hours testing AI video generation tools. All of them have specific stand out features and different uses.

Most recently, I used Argil, the AI video generation tool that can take any text, PDF or link and turn it into a video in about five minutes. It's useful for anyone who wants to create multiple types of creative assets.

So if you've ever wished your written stories could reach people beyond the written word, this tool is probably for you.

Here's what happened when I turned this week's Tom's Guide AI newsletter into a video.

What is Argil?

Argil is an AI-powered tool that turns written content into short, engaging videos in just a few clicks.

Although this tool is $39/month for the classic tier and $149/month for the pro tier, users can try it for free. Those tokens go really fast, so you'll probably get more like a few hours than a full week's worth to try the tool.

Fairly intuitive, you simply upload your content or paste a link, and Argil uses generative AI to identify key points, summarize them visually and match them with relevant stock footage, animations or avatars.

You can customize the tone, pacing and style, then export the result in vertical or horizontal formats for platforms like TikTok, Instagram, YouTube Shorts or LinkedIn.

It's a game-changer for creators and casual users who want to repurpose content without spending hours editing video.

You upload a document or paste in your article, choose your style and format, and Argil handles:

• Script generation
• Voiceover narration
• Matching B-roll and animations
• Optional subtitles and custom branding

How to try it

We all know attention spans are shorter than ever. Some people don’t want to scroll through a 1,200-word piece, even if it’s packed with useful info.

Video offers another way to deliver the same value, especially on platforms like YouTube Shorts, Instagram Reels and TikTok.

Argil cuts the work down to just minutes.

For my test, I uploaded a PDF of this week's AI newsletter. It is first-person, structured and had a clear narrative — ideal for visual storytelling.

Once uploaded, I selected a professional, energetic tone with dynamic transitions and subtitles. The AI automatically trimmed the content and suggested visuals to match.

Within minutes, I had a 90-second video ready to go — no editing software required.

What worked (and what didn’t)

I'll start with the issues I had first. There weren't a lot, but they were enough to give users a heads up.

First, there is an option to use yourself as an avatar. However, the process was not as easy as what I've experienced with Synthesia.

Unlike Synthesia, a video made with my iPhone and uploaded was not the correct format. I tried several times until I finally uploaded the video to ChatGPT and asked it to format my video for Argil. That felt like a step that could have been avoided since most people are probably making videos and uploading them from their phone.

Second, if you decide to choose a different avatar once your video has been previewed, you have to manually go frame by frame to change the avatar. This was another time suck that I would have rather avoided.

The tool also blurs out the face of the avatar during the preview stage, which is annoying.

But where Argil impressed me, is how easy the tool was to use. It also kept the core message of my text intact while trimming away unnecessary fluff.

I especially appreciated the ability to export in multiple formats; portrait for Instagram Reels and landscape for YouTube, which made it easy to repurpose the video across platforms.

That said, not everything was perfect. Some scene transitions felt a bit generic, the tone and voice felt robotic and the body movements were extremely repetitive. Occasionally the AI missed subtle nuances or tone, requiring a few manual edits on my part if I had the opportunity (and tonkens!) to go back and edit.

Still, for something that took less than 10 minutes to produce, the final result was more than impressive.

Final thoughts

If you're a content creator looking to transform blog posts into video, a marketer aiming to boost SEO and reach on social media or a writer building a visual portfolio, Argil is a no-brainer.

It’s a fast, simple way to give your existing work new life.

While it may not be flawless, it’s absolutely good enough to start using today, especially if you want your work to show up in more places with less effort.

More from Tom's Guide

• The ChatGPT job hunt hack no one’s talking about — but it actually works
• These 7 ChatGPT prompts rewired how I think — now I use them every week
• 5 hidden ChatGPT tricks most people don’t use — but they’re actually game changers

As chatbots continue to get smarter and gain the ability to do more, AI video generators are evolving just as fast. And while many are still in beta or invite-only, one thing is clear: prompt quality makes a huge difference.

After testing several platforms (including Google Veo 3, the premium video generator currently available for $249/month through Google Ultra) I've learned that the way you write your prompts completely changes the results.

Here's how I leveled up my AI videos — and the tricks you can use to make your own clips look sharper, more cinematic and more coherent.

To keep things fair, all prompts were tested using Veo 2 — Google’s lower-tier video generator that doesn’t include audio. To try it yourself, you’ll need to join the waitlist for Google Labs’ VideoFX tool, where Veo 2 is currently available.

You can also access Veo 3 through Google AI Studio or Canva Pro. These prompts can be used on any of the best AI video generators including Midjourney, Firefly and Sora.

Start with structure, not style

A common mistake many users make is starting with adjectives like, “a stunning cinematic video of a beach at sunset.” Sounds nice, but vague.

Instead, I always find better success when I lead with structure: “wide establishing shot of a beach at golden hour, camera slowly panning left.” The more you write like a director, the more polished your results will look.

Think like a camera operator

Veo seems to respond well to cinematic language.

Prompts like “low angle tracking shot of a child running through tall grass” or “drone shot pulling away from a mountain cabin at sunrise” generated far more dynamic visuals than basic descriptions.

If you’re not familiar with filmmaking terms, think in shots: wide, medium, close-up, overhead. While I am not a camera operator, I have written enough screenplays and been on enough television and movie sets to have these memorized.

Here are a few basic filmmaking terms that I frequently use in my prompts.

Camera angles & movements

• Establishing shot – wide view to set the scene
• Close-up – focuses on a subject’s face or detail
• Wide shot / Long shot – shows subject fully in the environment
• Over-the-shoulder – often used for dialogue scenes
• Tracking shot / Dolly shot – camera follows the subject smoothly
• Crane shot – sweeping overhead movement
• Static shot – no camera movement
• Zoom in/out – camera pushes or pulls
• Handheld camera – shaky, realistic effect
• POV shot – from the character’s point of view
• Dutch angle – tilted horizon for tension

Lighting & atmosphere

• Low-key lighting – moody, dramatic shadows
• High-key lighting – bright, cheerful, minimal shadow
• Natural light – outdoors or simulated daylight
• Backlighting – light from behind the subject (silhouettes)
• Lens flare – light streaks across the lens
• Cinematic glow – soft, dreamy lighting
• Neon lighting – vibrant, futuristic tones
• Film grain – adds texture, retro feel

Break it into beats

Instead of asking for a whole story in one sentence, break your prompt into visual beats. For example:

Beat 1: Wide shot of a skateboarder riding downhill at sunset

Beat 2: Slow motion close-up of wheels hitting the pavement

Beat 3: Camera follows from behind as sparks fly from the board

Even if your preferred AI video generator doesn’t yet support full scene transitions, this approach helps guide the system toward more intentional storytelling.

Add motion, mood and details — then, test and iterate

Movement is key. I started adding phrases like “camera tilts up,” “fog rolling in,” or “wind rustling trees.”

These made a surprising difference in realism and cinematic feel. So did mood cues like “soft lighting,” “overcast skies,” or “neon glow.”

AI video is still new, and results can be unpredictable. I ran each prompt multiple times to compare outputs, tweaking a word here and there.

Veo's consistency is noticeably better than other tools I’ve tried. It doesn’t always nail it, but when it does, the results can look like they came from a professional video shoot. The prompts below were used with Veo 3 so they also have sound.

Prompts to try

Landscape

Landscape prompt: Wide establishing shot of a mountain range at golden hour, camera slowly panning left. Fog rolls between the peaks, soft ambient lighting, cinematic depth of field.

Sports

Sports prompt: Slow-motion shot of a skateboarder jumping off a ramp at sunset. Camera follows from low angle as dust kicks up. Warm lighting, dynamic energy.

Food reel

Food reel prompt: Overhead shot of a steaming bowl of ramen being served in a Tokyo street market. Close-up of chopsticks lifting noodles, steam rising. Warm lighting, vibrant details.

Action POV

Action POV prompt: First-person POV of a cyclist riding down a forest trail. Leaves fly past, sun flickers through trees. GoPro-style realism, immersive movement.

The takeaway

If your AI videos look generic or "like AI," your prompt may be to blame. With a few smart tweaks; focusing on structure, camera angles and specific action, you can get more polished, more cinematic results.

Google Veo is one of the most capable tools I’ve tested, but like any AI, it’s only as good as what you feed it.

Want to try it yourself? Start small. Pick a scene, describe it like a shot list and watch what happens next. Let me know in the comments how everything came out!

More from Tom's Guide

• I just tested Google’s Doppl app to try on clothes virtually with AI — but it's got some wrinkles
• ChatGPT is reportedly building a Google Workspace and Microsoft Office rival — here's what we know
• The one word you should never say to your AI chatbot

Adobe just launched a dedicated mobile app for Firefly, its generative AI platform for image and video creation, and it could seriously change how we create on the go for both iOS and Android.

For the first time, you can generate high-quality visuals and short video clips right from your phone, no desktop required.

Whether you're sketching out a quick concept, remixing an idea or building social-ready content on the fly, Firefly is making it easier for users to stay in flow, wherever inspiration strikes.

What’s new (and why it matters)

iOS and Android support — For the first time, users can generate and edit AI-powered images and short video clips directly from their phones.

Text-to-image, text-to-video, and image-to-video tools — The app includes popular Photoshop features like Generative Fill and Expand, along with new video controls for motion, angle, and zoom.

Multi-model integration — In addition to Adobe’s own models, users can generate content using AI engines from Google Imagen, OpenAI, Luma AI, Runway, Pika, Ideogram, Flux, and others — all within one interface.

Creative Cloud sync — Projects created in the app sync automatically with Photoshop, Premiere Pro, and Firefly on the web — enabling seamless cross-device workflows.

What stands out

AI video generation is officially going mainstream, with Adobe becoming the first major player to bring text-to-video tools to mobile, letting creators generate short, editable clips right from their phones.

What really sets Firefly apart, though, is its focus on commercially safe content; its models are trained on licensed or public-domain material, which means creators don’t have to worry about copyright headaches.

And with multi-model support built in, Firefly is now more like a creative hub where users can experiment with different AI image generators without jumping between apps.

What this means for creators

On-the-go ideation: Users can now quickly sketch, remix or concept visuals wherever they are, right from their phone.

Faster visuals and b-roll: Firefly’s text-to-video tool creates five-second 1080p clips with camera control, no editing suite required. This makes it ideal for social-ready footage.

Creative freedom: With access to multiple AI models in one place, creators can experiment with aesthetics, remix outputs, and test ideas faster than ever.

What’s next

Adobe is rolling out both free and paid tiers for Firefly, with Creative Cloud subscribers getting unlimited access to basic generation.

For more advanced features and third-party video models, users can expect a credit system or in-app purchases.

Looking ahead, Adobe is also beta-testing Firefly Boards — a collaborative moodboarding tool designed for video projects — and plans to deepen integration across its full Creative Cloud suite, making it easier for creators to move between tools like Photoshop, Premiere Pro and Firefly without disrupting their workflow.

 Final take

With its mobile Firefly app, Adobe is signaling the future of AI-powered creativity: on-demand, on-device and deeply integrated into pro tools.

By merging powerful text-to-image and text-to-video tools with a growing ecosystem of third-party models, Adobe is positioning itself at the center of next-gen creative workflows right in your pocket.

More from Tom's Guide

• I use the 'blank line' prompt every day now in ChatGPT — here’s why
• ChatGPT just got way better at search — and Google should be worried
• I stopped using punctuation in my prompts — here’s why I swear by it now

Google’s Veo 3 made a splash at Google I/O 2025 as the latest leap forward in AI-powered video generation.

As a feature within the new Google AI Ultra subscription, the highest access to Google’s most advanced models and premium features, you can get Veo 3 and Flow, which strings videos together for a more robust cinematic feature.

Designed to turn simple text prompts into hyper-realistic clips — with audio, dialogue, and sound effects — it’s arguably one of the most advanced tools available to casual creators. And after testing it hands-on, I can say it delivers some truly jaw-dropping results. But it also comes with its share of hiccups.

Here’s what I loved about Veo 3 — and what left me frustrated.

What Veo 3 gets right

I’ve tested my fair share of AI video tools, including earlier versions of Veo, and this latest release is by far the most user-friendly when it comes to adding sound and dialogue.

The realism is genuinely impressive — especially for the fact that the 8-second clips can be generated in under two minutes on a computer without a full production crew.

Users across the internet are sharing clips that are nearly indistinguishable from human-made videos. It’s a glimpse at just how fast this tech is moving.

Where Veo 3 still struggles

For all its strengths, Veo 3 still has a ways to go before it’s seamless. Obviously, it's still in experimental mode, so Google is working out the kinks, but here are the biggest issues I ran into while testing:

1. Prompt interpretation feels hit-or-miss

Veo 3 sometimes struggles with spatial prompts, like when I asked for an overhead camera angle but got a slightly tilted side view instead. It seems to prioritize cinematic flair over strict prompt accuracy, which limits creative control.

2. Audio doesn’t always work — and it’s not obvious why

By default, Veo runs in Veo 2 mode, which does not include audio. I only realized this after a few silent clips and some digging. You’ll need to manually switch to "Experiential Mode" under the "Quality" tab to activate audio and dialogue.

Even then, lip-syncing is inconsistent, and dialogue sometimes drops out altogether, like a badly dubbed foreign film. Additionally, the subtitles are almost always wrong or misspelled.

3. Complex scenes throw it off

Veo 3 shines with single-subject clips, but longer or more intricate scenes can fall apart. The narrative gets muddy, and character interactions often feel stiff or repetitive. If you're aiming to create a multi-character, multi-scene story, temper your expectations.

4. The interface still needs polish

There were moments when the interface felt unintuitive or unstable. I experienced an unexpected session timeout that erased a generated video, and I couldn’t find a recovery option.

Additionally, when I prompted the model to add dialogue within the scene I got something that did not fit the scenario at all.

For a tool this powerful, the UX still feels a bit rough around the edges.

5. It raises some big ethical questions

Veo’s realism is incredible — and a little unsettling. There’s growing concern that ultra-realistic, AI-generated videos could blur the lines between fact and fiction, especially as this tech becomes more accessible. It also sparks new debates around authorship and originality in creative work.

Promising but pricey

Veo 3 is a huge step forward in AI video — especially for casual users who want fast, high-quality results. But at $249 per month (with a discounted rate for the first three months), the Google AI Ultra package is a steep price to pay for a tool that still has some notable bugs.

If you’re just looking to experiment with video generation or create basic promotional content, Veo 3 is exciting — but not yet essential. For professional creators, though, it’s worth watching closely. Just keep your prompts tight, your expectations realistic and your finger ready to re-render.

More from Tom's Guide

• This $12.99/month hack gives you access to Google’s Veo 3 AI video tool — here’s how
• Claude Opus 4 is here — and it might be the smartest AI assistant yet
• The only 5 prompt types you need to master ChatGPT (and any other chatbot)

Pika, the AI video platform known for making high-quality tools accessible to everyone and for letting you "meet your younger self for coffee," just dropped a new feature that's already turning heads.

I couldn't believe how ridiculously realistic the videos appear when I tried it. Meet PikaTwists.

Designed to help creators inject dramatic, funny or surreal twists into their videos, the feature gives users the power to transform any clip with just a few words.

The video that sold the feature for me was the one with a cat casually relaxing on a table next to a cup of coffee, then waking up and drinking it.

So, whether you want your feline to join you for a cup of tea, your car to levitate in the middle of a drive, or for your dog to suddenly start breakdancing, PikaTwists makes it happen — in minutes.

How to try it

All it takes is uploading a video and describing the twist. Simply prompt the AI to make the change and it will handle the rest, generating realistic, mind-bending effects without the need for complex editing software.

What makes PikaTwists so impressive is how well it preserves the original footage. We're talking lighting, camera angles, objects, and characters — all stay consistent while the twist blends seamlessly into the scene. It’s the kind of result that previously required a VFX studio but now anyone with a video, an imagination and a good prompt can make it happen.

Pushing boundries

"Our goal is to empower users to push the boundaries of their creativity," explains Pika's Founder and CEO Demi Guo. "We believe that by making advanced AI tools user-friendly and accessible, we can inspire a new wave of content creation that is both innovative and fun."

With a growing suite of AI features and a rapidly growing user base exceeding 12 million, the platform continues to push boundaries by putting creators in the driver seat and giving them editing skills to bring high-concept ideas to life without specialized skills.

From casual content creators to aspiring filmmakers, PikaTwists opens up new possibilities for storytelling — all powered by whatever prompt you can imagine.

Accessibility and outlook

PikaTwists is currently available to both free and paid users. Free (Basic) users can access the feature through Turbo mode. Pro users get full access across all paid subscription tiers.

This latest launch follows other Pika Labs tools such as Pikadditions and Pika 2 with sound effects. With a mission to put advanced video tools into the hands of everyday users, regardless of tech skills, this tool is no exception.

Want to try it out for yourself? Head to pika.art and start twisting your videos today. How'd your PikaTwists turn out? Let me know in the comments!

More from Tom's Guide

• I use ChatGPT every day — here's 10 prompts I can't live without
• ​I've tested countless AI tools — this one surpasses them all
• ChatGPT just got a huge memory upgrade — here's why it's a big deal

After initially rolling out ChatGPT's image generation feature for free, OpenAI has recently implemented temporary rate limits due to overwhelming demand straining their GPU resources.

I actually found a bug myself when trying out the new image generation feature. Despite using ChatGPT Plus, the images were generated with DALL-E, yet there was no indication that they were DALL-E images.

CEO Sam Altman jokingly remarked that the company’s GPUs are "melting" under the load. OpenAI is working to enhance efficiency to manage the surge in requests and hopes these limitations will be short-lived.

So, what’s a user to do if they want realistic images put don’t want to pay the subscription price?

Meet Playground. Not affiliated with Apple Intelligence’s Image Playground, this site offers a variety of AI image generation possibilities, many of which are free. While there is a paid “Pro” option, users can generate mind-blowing images for free.

What is Playground AI?

Playground is a platform available via browser and app that enables users to create custom designs and graphics.

Playground AI has developed advanced text-to-image generative models, notably Playground v2.5 and Playground v3. Playground v2.5 is a diffusion-based text-to-image model that excels in producing highly aesthetic images with resolutions up to 1024x1024 pixels. It supports various aspect ratios, including portrait and landscape orientations.

Playground v3, the site’s latest model, introduces an integration of Large Language Models (LLMs) to enhance text-to-image alignment. The AI's unique approach improves the model's ability to adhere to textual prompts, perform complex reasoning, and accurately render text within images.

Users can start generating images with a completely empty canvas or explore a range of templates for logos, art, stickers, wallpapers, and more.

1. Album art or book cover

Prompt idea: “Create art for my new album that’s all about cats.”

One of my favorite aspects of Playground is the on-image text. It’s clear and accurate. Combined with the fun fonts available, users can add personality to their creations.

In this album art the placement of the three cats at the bottom of the frame draws the eye upward and ties the image together nicely. Their realistic poses make them feel like friends and the overall vibe is cohesive.

2. Interesting art

Prompt idea: “A surreal painting of a young man writing with a quill inside an ornate gold frame, hanging in a library. Ink spills out of a pen into the real world.”

Although the AI made the “pen” in this image a pencil, overall, this image is incredible. The ornate gold frame is beautifully rendered and contrasts nicely with the muted tones of the background.

The shadows, lighting, and depth all contribute to a very believable and cinematic scene.

3. Image with fun font for a friend

Prompt idea: “Create a fun note with an ice cream sandwich and bold, vibrant colors for me to send to a friend.”

This image is a lot of fun. The hot pink milkshake with bright toppings stands out against the dotted background. Despite the message being a little “blah,” the rest of the image is playful and exciting.

The typography looks great, and it is cleverly incorporated within the image. The clean composition gives this image a polished look.

4. Children’s book illustrations

Prompt idea: “Create an image of a little girl walking her cat in the rain.”

The idea of a cat in a tiny raincoat being walked by a child in matching rainy-day gear is adorable and that’s why this image is so much fun to me.

The AI really made this feel like a page out of a children’s storybook. The soft watercolor effect and loose linework give it a warm, storybook aesthetic.

5. Products

Prompt idea: “Minimalist coffee cup mockup on a beige background.”

I could see this image being used for a small business like a coffee shop or local café. It’s cute and realistic while adding a touch of whimsy. It’s almost hard to believe this was created with AI.

6. Sports

Prompt idea: "A flat lay of a white baseball jersey with black piping, paired with a tan leather baseball glove featuring red lacing, and a classic white baseball.”

Once again, the AI image generator hit it out of the ballpark. (Sorry, I couldn't resist). This looks like a very realistic image of a baseball uniform. I could see this type of prompt used by someone who has an Etsy shop or some kind of hobby with fashion or sports.

7. Memes

Prompt idea: “A cute gray tabby kitten sitting upright on a cozy, blue-patterned bed with soft lighting.”

What’s almost as good as cats? Memes about cats! Playground even has AI templates to help you get started making your own memes.

The realistic images rival subscription-level images of big tech chatbots.

Final thoughts

If you were disappointed by the news that OpenAI is not allowing its enhanced image generator within the free tier of ChatGPT-4o yet, that doesn’t mean you can’t generate great-looking AI images for free.

Whether starting from a template or using your own prompts, Playground might just be what you’ve been missing.

Another perk of the site is the community. Users can share their designs, prompts, and templates. But that also means that anything you make is free for others to use or alter. Some may see that as a downside to the site, however I see it as an opportunity to enjoy the endless possibilities of AI image generation.

More from Tom's Guide

• What happens when you feed Gemini 2.5 ChatGPT’s best prompts? I found out
• I used ChatGPT Vision and Deep Research to rid my kitchen of sugar ants for good — here’s how I did it
• I just tested ChatGPT-4o's enhanced image generator with 7 prompts — here's the results [Update]
• I've made Gemini my primary AI tool instead of ChatGPT — here's 5 reasons why

Social media trends come and go, but the recent phenomenon that has captured hearts of many is the “Met Myself for Coffee" movement. Originating on platforms like TikTok, this trend involves users crafting narratives about imaginary encounters with their younger selves over a cup of coffee.

These heartfelt stories often delve into themes of personal growth, self-reflection, and healing, resonating deeply with audiences worldwide. What would you say to your younger self if you had the opportunity? Well, Pika wants to help.

Building upon this introspective trend, Pika, the AI-driven video platform, has introduced its latest new feature titled "Selfie With Your Younger Self." This tool allows users to transform the written concept of meeting their past selves into a vivid visual experience.

How to try it

Anyone can try this tool for free, but keep in mind that users with a basic account (that's the free one) may wait longer to get their videos — maybe up to an hour! Pro subscriptions hop the line and get their videos in less than a minute.

The process is designed to be user-friendly and can be done via browser or the Pika app. Simply start by recording a video of yourself as you are today. Next, upload a photograph from a previous stage in your life.

If you don’t have one, not to worry, Pika's advanced AI technology can generate a “childhood photo” of you. Once you’ve uploaded your video and old photo (or told Pika to generate one for you), the AI gets to work, merging these inputs and generating a cohesive video where the two versions interact.

Internet celebs get in on the fun

To demonstrate the variety of possibilities with this AI tool, Pika has collaborated with several internet-famous personalities who rose to prominence as memes.

Figures such as Sammy Griner, known as "Success Kid," Zoë Roth, the face behind "Disaster Girl," Kyle Craven, famously recognized as "Bad Luck Brian," and Tay Zonday of "Chocolate Rain" fame have participated in creating videos that reunite them with their younger, meme-famous selves.

Expert tips

The process works best if there is nothing else in the photo. For instance, the first time I tried to generate a video, my “younger self” image was holding a doll. Unfortunately, the AI generated a video of me hanging out with the doll, not my younger self.

Also, try to not take up the entire space within the video frame, so the AI generated image has room, otherwise the younger version of yourself may get cut off.

Another tip is to make sure you look forward. I tried to make the video more realistic and it was a total fail as the AI me came in on the other side, not the side I thought.

Other uses for the feature

The versatility of the "Selfie With Your Younger Self" feature opens up a myriad of creative possibilities. Some users may choose to highlight their transformation over the years, showcasing changes in appearance, style, or demeanor.

Additionally, the ability to stitch together multiple generations allows users to craft narratives that honor their familial heritage, illustrating the interconnected stories that have shaped their identities.

For instance, beyond having coffee with my younger self, I would love to enjoy coffee as an adult with my grandma who has passed. Using an image of her, I could see us together today.

Maybe you’re missing a friend or relative and just want to be together for a moment. With Pika you can do that.

Looking ahead

With a user base exceeding 12 million, Pika consistently offers its users new tools. Pikadditions is another recent feature that lets users explore AI. The tool offers a video-to-video model that allows users to incorporate any person or object into their existing videos.

The introduction of the "Selfie With Your Younger Self" feature exemplifies Pika's commitment to providing tools that let users explore their creativity, reflect on their personal narratives, and share their stories in innovative and meaningful ways.

More from Tom's Guide

• I used ChatGPT Vision and Deep Research to rid my kitchen of sugar ants for good — here’s how I did it
• I just tested ChatGPT-4o's enhanced image generator with 7 prompts — here's the results [Update]
• I've made Gemini my primary AI tool instead of ChatGPT — here's 5 reasons why

OpenAI just enhanced its image generating capabilities with a new model integrated within ChatGPT-4o.

Promising more realistic, richly detailed images from the simplest prompts, the model sounded too good to be true, especially since I haven’t been all that impressed with ChatGPT’s image generating capabilities in the past. I have to admit I was skeptical.

OpenAI's new image generator is designed to be fast and intuitive, so I wanted to try it with 7 prompts. I tested everything from cinematic landscapes to mock product concepts to see how well it works and what sets it apart from other AI image generator tools.

Here's what happened when I put this ChatGPT-4o image generator to the test.

Update: After scratching my head wondering why my initial results weren't as good as everyone else's, I realize I encountered a bug—and i'm not the only one.
Sam Altman mentions in a tweet just after 12pm ET today that the model is experiencing difficulties.

Even after logging off, logging back in, and restarting my computer, I was somehow still stuck with the old model. OpenAI confirmed that I indeed had found a bug and showed me what the images should have looked like. For comparison, I have added them here under the ones I generated.

1. Structures in a setting

Prompt: Create a photograph of a suburban house in early spring.

These houses are beautiful, and the chatbot catches the season correctly, but these do not look like photographs. They are pretty close and, for a free chatbot, definitely eye-catching and clear. But they still give off that AI-glow.

I could see images like this being used for a brochure or travel company, but in terms of fooling someone that it’s a photograph of your house, nah. The driveway alone gives it away.

2. Design a logo

Prompt: Create a logo for an ice cream shop on the boardwalk.

Knowing that it is this easy to create a logo makes me wish I owned a business. With a simple prompt, I had several logos to choose from in seconds.

After the first few logos that had the same aesthetic, I asked the chatbot to switch the style. It went from whimsical to bland really fast, but with a little prompt tweaking, I’m sure the chatbot would come up with a usable.

I did notice a few misspellings and blurred letters, so be sure to do a spell check after you generate your logo.

3. Comics

Prompt: Create a cartoon with two cats chatting about the weather.

I can see myself playing around with these prompts a lot. Maybe Tom’s Guide needs a weekly AI comic strip?

When I first prompted the chatbot, it generated just one comic. From there, I asked for a comic strip.

I wasn’t incredibly impressed with the layout. I was expecting something more like what you find in the Sunday paper. I then prompted the chatbot to do a third comic strip in the style of Disney. Definitely not "The Aristocats," but I thought it was pretty close.

4. Package design

Prompt: Create the package renderings for a chocolate bar called Amanda.

This was pretty cool and the most realistic prompt yet. I thought it was interesting that the chatbot went a “fancy” direction for the chocolate bar, but that’s probably because I used my name. Had I used “Crazy 4 chocolate,” it probably would have done something simpler.

But what’s interesting is when I asked the chatbot to generate two more images, it gave me an error message:

“I wasn’t able to generate the two additional chocolate bar package designs due to a content policy restriction. It might’ve flagged something in the description, even though the request seems fine on the surface.

"If you’d like, I can rephrase the prompt and try again with slightly different wording — just let me know if you want to stick with a luxury theme or explore something more fun, playful, or eco-friendly!”

After that, I tried again, but the chatbot was done with that prompt.

5. Landscape

Prompt: Create a photograph of New York City in the summer.

The first image the model generated looked more like a painting than a photograph. I keep hoping ChatGPT will have the ability to create photorealistic images, but I’m finding it fails most of the time.

I asked the chatbot to be “more realistic,” but even the second image didn’t cut it for me. I also find it interesting that the chatbot focused on the greener aspects of NYC in the summer rather than generating an image of the busy sidewalks.

The images are good, but I’m disappointed.

6. Awards

Prompt: Create an image of an AI award.

Because I have AI Madness on the brain and OpenAI has said that the ChatGPT-4o image generator can create images with little prompting, I decided to try it here. The chatbot did a nice job coming up with some realistic-looking awards.

The first one is a little busy, but once I prompted the chatbot to go for a simpler style, the image was much cleaner.

This would be great for parents who want to give their kids awards for chores, office parties, and similar events for recognizing a job well done.

7. Various objects

Prompt: Create a realistic image of a dinosaur, a potato, and a cat.

The potato was the most impressive of the three images. The dinosaur looked like a generic image of a prehistoric creature and the cat was downright comical. I purposely did a mixed bag here to test the AI's ability to understand the word "realistic."

The AI seems to be somewhat unpredictable with it's style, so my biggest recommendation is to be as specific as possible when making your prompts. OpenAI says you don't need to, but I highly recommend it.

Final thoughts

Overall, the ChatGPT-4o image generator feels like an upgrade for several reasons. It is easy to use directly within the chat window, the chatbot pivots when prompted or given more information, and anyone can use it, even without an account and for free. That last point is a huge plus.

Although the chatbot tends to have trouble with on-image text and occasionally needs extra prompting, the enhanced image generating tool is something I plan to keep in my AI toolbox.

More from Tom's Guide

• Google unveils Gemini 2.5 — claims AI breakthrough with enhanced reasoning and multimodal power
• I just tested Grok vs. DeepSeek with 7 prompts — here's the winner
• I tried these 7 ChatGPT prompts to supercharge my Netflix viewing experience

OpenAI today announced the GPT-4o image generator, introducing advanced image generation capabilities integrated within the ChatGPT-4o language model.

The company says that GPT-4o represents a significant leap forward in image generation that should create images that are not only visually stunning but practically useful.

Practical visuals for every day use

The GPT-4o image model focuses on 'useful image generation,' which means users can now use the AI model for everyday needs such as logos, diagrams, and infographics.

Unlike previous generative models that often produce surreal but impractical visuals, GPT-4o was designed to deliver more contextually relevant and accurate imagery.

Superior text integration

Key features of GPT-4o include upgraded text rendering, allowing seamless integration of textual information into images. This capability supports visual communication, elevating the utility of generated images.

Additionally, GPT-4o supports multi-turn generation, enabling users to refine and adjust images through natural conversational interactions, maintaining consistency throughout iterative design processes.

Complex instruction handling

The image generation from GPT-4o is capable of managing complex prompts involving up to 20 distinct objects, which is an improvement over existing systems.

Through in-context learning, GPT-4o can analyze user-uploaded images, seamlessly incorporating these details into subsequent image generations, thus creating a more personalized and contextually informed visual output.

Comprehensive multimodal training

Built upon extensive multimodal training on vast online image and text datasets, GPT-4o has developed sophisticated visual fluency, allowing the model to produce images that are contextually aware, stylistically diverse, and photorealistically convincing.

Limitations and safety concerns

Despite its advanced capabilities, OpenAI acknowledges certain limitations, such as occasional cropping issues, hallucinated details, difficulties rendering dense information at small scales, and precision editing challenges. Multilingual text rendering, especially for complex non-Latin scripts, remains an area under active development.

Safety continues to be a paramount focus, with rigorous measures in place to block harmful content, including explicit materials or images that violate content policies. Provenance tools, such as C2PA metadata tagging and internal reverse search, ensure transparency and accountability in generated visuals.

Availability and future outlook

GPT-4o image generation is available now across all ChatGPT platforms, including Plus, Pro, Team, and Free tiers, with Enterprise and Education access anticipated soon.

Regardless of tier, users can specify detailed image requirements — from exact colors and aspect ratios to transparent backgrounds — making professional-quality image creation as straightforward as a simple chat interaction.

OpenAI's GPT-4o signifies a major advancement in AI-driven visual communication, turning generative image creation into an accessible, practical and powerful tool for everyday users and professionals alike.

More from Tom's Guide

• I just tested Grok vs. DeepSeek with 7 prompts — here's the winner
• Is ChatGPT making us lonely? MIT/OpenAI study reveals possible link
• I replaced Alexa with ChatGPT on my Amazon Echo — here's how you can do it too

Image generation is the veteran tech of AI. While the market is dominated by companies like Midjourney and Google, Ideogram has managed to keep pace in our list of the best AI image generators, and has just introduced the Ideogram 2.0a model — a faster, cheaper alternative to its original product.

I tested out the new model to see whether it was a worthy successor, and what I found was quite surprising in a number of ways.

The company originally made its name from producing high quality, interesting images which followed prompt instructions well, and delivered excellent text elements. So I was excited to see whether the latest update had improved on these qualities.

My goal was to produce five images for the test, with no specific science behind the prompts, just a range of creative demands which are typically challenging for image generators.

The focus was on five categories: people, place, nature, portrait and design.

Ideogram offers a range of custom settings users can choose from, including aspect ratio, generation speed (slower equals higher quality) and pre-sets such as design, 3D, anime and realistic.

Each generation costs a number of credits depending on processing demands, and delivers four sample images to choose from.

Ready to see how the model got on? Here we go.

Test One – People

Prompt: a beautiful family birthday party in an English country garden in the mid-afternoon, with a cake on which is written “Happy Birthday Maggie”.

Out of the twelve test images I created with this prompt, this one was the best overall. But only because of the prompt adherence.

There were other images with fewer glitches but they failed to deliver key features, for example a cake with writing or the garden scenario. Even with this winning image, while the overall composition is good, look closely and you’ll notice a few hand glitches which ruin the overall effect.

Test Two – Nature

Prompt: 4 majestic elephants on the savanna, medium shot, natural afternoon lighting, National Geographic, Nikon camera, award winning, with a small wooden building to the rear with a sign saying 'Safari' on it, photo, typography, cinematic

It took me twenty attempts to get this particular image. For some reason image AI has a problem with creating realistic looking elephants. And math, apparently. In this case four was a challenge, so we received three elephants which I guess is better than nothing.

My main gripe is with uniformity, so many of the animals looked like clones, which was a shame.

I selected this image in particular because of the excellent text adherence, which many of the other images fail to produce. And is that just a half elephant in the background?

Test Three – Place

Prompt: A street photograph, Canon EOS, a group of 5 friends enjoying a Sunday afternoon coffee at a pavement café in Paris called Gaucho, soft lighting, colorful

Once again out of the sixteen images I generated, this was the closest to the prompt, and featured a clear text element.

There are glitches unfortunately, but many of the other images, while better, had unnatural poses or no text at all. It seems surprisingly difficult to generate a clean and accurate image with this new model. Especially if there is text in it.

Test Four – Portrait

Prompt: A close-up portrait of an elderly fiercely charismatic Asian fisherman, looking past the camera with epic afternoon lighting beside a small village river.

It seems that portrait photography is a much easier task for the new 2a model. I generated twenty shots and a good half dozen of them were pretty decent.

There were one or two framed in city settings and one with a bizarre Scottish hat on, but overall the quality was good. My suspicion is when the subject is singular and there’s no requirement for text processing, the model has a much easier task of meeting the prompt requirements.

Test Five – Design

Prompt: A range of ultra luxury chocolates called Finesse.

Sixteen images of chocolates produced this rather nicely balanced shot, which you could imagine in a magazine advert.

Unfortunately again, as with so many of the generated images with this model, it’s so close and yet not quite right.

The text drifts off into nonsense, and while I think many image generators suffer the same problem, it’s still a shame. But overall, not bad at all.

Conclusion

I have to admit to being rather disappointed with the overall fidelity and quality of the images from the new Ideogram 2a model. It seems that the company has opted for cutting back the original Ideogram quality in favor of speed and a lower price per generation.

I learned to rely on the quality setting and Realistic pre-set in order to get the best out of the images I created. But even so there were far too many glitches and lackadaisical prompt adherence problems for a product at this level.

However there is one huge silver lining. I stumbled onto the edit Canvas function during the test, and it’s excellent.

It’s super easy to use, very powerful and allows for a huge amount of customization, editing, extending and polish which you won’t find on most other image generation tools.

It’s clear that many of the gallery images on the site have made considerable use of this editor to improve the original generations. But why not?

The example above is me inserting a chilled lion into a group of elephants on this heavily edited and extended image. Done in about 20 seconds.

Let me know what you think of the images and how well they adhere to the prompts in the comments below.

More from Tom's Guide

• Gemini just became the ultimate collaborator — everything you need to know about this huge new upgrade
• ChatGPT could soon work with Slack and Google Drive to make office life a bit easier
• AI Madness: The ultimate chatbot showdown

Leonardo and FreePik are two of the most established AI image generators on the market. Both companies have a solid reputation for producing quality images from text prompts, although they take significantly different approaches to that task.

While FreePik provides a fairly typical range of products and services for generating images, Leonardo has turned into a vast hub of image resources covering everything from templating and design to video creation.

One of the interesting things about FreePik is the fact that it offers third party image generation from Ideogram and Google Imogen 3, as well as its own models. Leonardo also offers images from Flux and Stable Diffusion, but focuses more on image styles than named models. It will be interesting to see how this contrast works in practice.

We thought it was about time to put Loenardo and FreePik in a head-to-head test to which service offered the best mix of quality and versatility to its users.

I decided to go for five prompts to test across a range of requirements. Fantasy Environment, Natural World, Abstract Design, Photorealistic Concept and Text & Typography Design.

Test One: Text & Typography Design

Prompt - "A stylish modern penthouse apartment in Paris, with a spectacular view of the city. Modern minimalist furniture, chrome and steel with an expensive look. One expensive modern painting on the wall, and a small stainless steel bauble which is the word 'Peace'.

Leonardo: The penthouse created by Leonardo is definitely minimal and stylish, and judging by the Eiffel tower in the distance, it’s set in Paris. So the prompt adherence is spot on.

The one glitch is the word Peace is not really clear in the image, but this was the best of the multiple images I produced with Leonardo, so text is not a strong point here.

FreePik: I literally gasped when this image was produced by FreePik. I used the Mystic 2.5 Flexible model, and it’s an absolute triumph.

Not only did the model create this image in a single pass, but everything from the text adherence to the structure of the image is just about perfect.

This is one of the best outputs I’ve ever seen from a AI image generator; perhaps only Midjourney has done this well for me in the past.

Test Two: Photorealistic Concept

Prompt - “A forgotten lighthouse keeper's desk from the 1920s, with an abandoned journal, vintage navigation tools, and a steaming cup of tea. Light streams through a salt-crusted window, capturing dust particles in the air. The scene tells a story of solitude with detailed textures of weathered wood, tarnished brass, and dog-eared pages.”

Leonardo: This image is what I would call a workable generation. It provides a relatively close approximation to the prompt, although it would be an exaggeration to say this was a 1920s setting.

There’s nothing wrong with the image, but it’s a pretty bland. If I was a schoolteacher I’d mark this as "could try harder." And this was the best of at least eight generated images.

FreePik: This image is astounding. Not only is the prompt adherence 100%, but it oozes character and atmosphere.

You can almost smell the musty location, with dust in the air and the kind of brass implements and vintage tools you’d expect to find in a such a place. And just look at that beautifully textured wood. This is poetry in an image.

Test Three: Abstract Design

Prompt: "A geometric dreamscape where architecture defies gravity, featuring impossible staircases reminiscent of Escher's work but with a vibrant color palette. Flowing fabric-like structures intersect with crystal formations, while small human silhouettes provide scale. The composition balances between order and chaos with strong diagonal elements."

Leonardo: The colorful palette used by the Flux model, combined with some lovely design elements make this a great image. It’s vibrant, it follows the slightly abstract prompt, and gives the air of surrealism that’s demanded by the text.

The final result is let down perhaps by the fact that the image features not just human silhouettes but actual tiny scale people.

FreePik: A completely different take on the same subject, this is a much more organic interpretation of the prompt. That’s not to say it’s wrong, it’s just different to the Leonardo version.

They both flow, feature vibrant colors and echo Escher’s original works. Once again excellent prompt adherence even down to the black silhouettes on the stairs.

Test Four: Natural World

Prompt: "A macro photograph capturing the moment a dewdrop forms on a spider's web at dawn. Within the dewdrop, a perfect reflection of the surrounding forest appears, with details of moss-covered trees and morning mist. Soft pastel colors from the sunrise create rainbow prisms through the water droplet."

Leonardo: Both models struggled a little with this prompt, but Leonardo definitely comes out on top in this instance. The cobweb is realistic, evokes the kind of atmosphere you’d expect from an early morning forest, and captures the macro photo feel the prompt demands.

FreePik: For some reason I find this image, although beautifully constructed and true to the prompt, just a little disappointing.

The cobweb is almost an afterthought, while the dew drop is a huge improbable balloon of water dominating the screen. Not a great interpretation of the prompt in this case, although the reflection is bang on target.

Test Five: Fantasy Environment

Prompt:"A magical underwater library where bioluminescent sea creatures illuminate ancient books, detailed with vibrant coral shelves and floating text manuscripts. The scene is bathed in ethereal blue and green light filtering through the ocean surface above. The library extends into the depths with spiral staircases made of shells and pearls."

Leonardo: Once again this is the best generation out of at least eight that I made. It’s adequate, but lacks any real atmosphere.

Flux has made it a little sterile in my opinion. I chose it because it was one of the few that Leonardo generated which at least tried to follow the bioluminescent illumination aspect of the prompt. But it’s definitely a bit meh.

FreePik: My oh my, here’s Mystic 2.5 once again dominating the conversation. This is a stunning image in so many ways. It’s full of life, colorful, prompt friendly and genuinely interesting.

Our eye is drawn into the intriguing curves, and even though the staircase is not made of shells and pearls, we can forgive that because of the vibrant use of coral and lighting in the foreground.

Conclusion

What a shocking surprise. I went into this test thinking it was a slam dunk in favor of the powerful and well-established Leonardo product brand. But that opinion was instantly shattered with the first image generated by the Mystic 2.5 Flexible model.

I’m not sure what the Spanish have done to create such a powerful and world-class image generator, but it really has demonstrated that the AI image world is not as cut and dried as we might believe.

It’s not just the image generation that makes FreePik so impressive. I spent time playing around with a few of the other features, and was blown away by things like the Mockup generator with its simple interface and superb output.

FreePik will now be my go-to commercial image generator, no question.

More from Tom's Guide

• Forget ChatGPT Canvas — I just tried Gemini Canvas and I'm floored by the difference
• AI Madness: The ultimate chatbot showdown
• I test AI chatbots for a living and these are the best ChatGPT alternatives

Just in time for Valentine’s Day, TCL Studios has unveiled "Next Stop Paris," a romantic AI short film that premiered on February 13, 2025, on the company's streaming platform, TCLtv+. This 12-minute film represents a significant milestone as the first extended project to emerge from TCL's virtual studio and production hub.

We've spent hours testing some of the best AI video generators available, from Luma Labs and Pika to Sora and Showrunner, and none are as sophisticated enough to create a film of this caliber or length.

Director Stuart Acher expressed enthusiasm about the innovative process, stating, "When I was approached to re-envision 'Next Stop Paris,' I saw the opportunity to harness this innovative technology to serve storytelling and found that incredibly empowering as an artist. But as an actor's director, it was also imperative that we preserve the human element. We created two people who fall in love and challenge the world around them, and while you are looking at AI-generated characters, their hair, their bodies, their clothes, and the world — they are grounded in the real, raw performances of talented actors that we filmed. Without that, the tech is just, well, tech."

A look at the future of AI storytelling

Movie studios of all sizes have already begun integrating AI into their films, with two of the most highly nominated productions at the 2025 Oscars incorporating AI into their creation. For “Next Stop Paris,” TCL Studios employed a hybrid approach, blending live-action footage with advanced AI tools to craft the narrative and its Parisian backdrop. The creative team utilized a suite of technologies, including Midjourney, Runway ML, Nuke, Adobe After Effects, and proprietary tools from TCL's R&D division, to achieve this integration.

Most combined practical video was captured with professional cameras and various AI applications. The environments were imagined using AI, with many complex settings based on 3D models treated with exclusive TCL methodologies.

A culmination of an ambitious eight-month production cycle aimed at developing realistic human characters capable of expressing genuine emotion, over 80 professionals contributed to the project, encompassing roles such as directors, writers, actors, producers, designers, VFX experts, engineers, and graphic artists.

How to watch

I watched the short film as a casual viewer and then as an AI expert. I was impressed by the dialogue, the human-like element of the characters, and the natural flow of the story. For me, there wasn't ever a time where I caught a glimpse of "robotic" movements or dialgoue that didn't match the lips moving. The voiceover carried the story, and the AI humor sprinkled throughout the film made it fun to watch. The film delivered in a way that I've never seen before at this type of AI level.

"Next Stop Paris" is now available for streaming on TCLtv+ and YouTube. It will anchor the platform's new Romance O&O Channel just in time for Valentine's Day.

As TCL Studios continues to push the boundaries of technology and storytelling, audiences will surely anticipate more innovative content that blends traditional filmmaking with cutting-edge AI advancements.

More from Tom's Guide

• 7 reasons why Ray-Ban Meta smart glasses are now my go-to wearable for running
• I asked ChatGPT, Gemini, Perplexity, and DeepSeek to predict what Apple is announcing Feb. 19 — here's what to expect
• Make your own music with Riffusion AI — 5 prompts to get started

Early last year, I got a sneak peek at Fable Studio's latest innovation, Showrunner, the AI-driven platform that gives users the opportunity to create customized television episodes, heralding a new era in content creation.

AI is already integrated into Hollywood, so the opportunity to dive in on a smaller scale has always intrigued me. I've enjoyed playing around with what seems like the intersection of creativity and technology and can understand why Showrunner has been dubbed the ‘Netflix of AI.'

Fable Studio, co-founded by Edward Saatchi, has a history of pioneering immersive storytelling experiences. Earlier projects, such as the Emmy-winning VR adaptation of Neil Gaiman's "Wolves in the Walls," showcased the company’s commitment to blending narrative depth with technological innovation.

It makes sense that they would build upon that foundation, and venture into the realm of AI-generated content, leading to Showrunner.

How Showrunner works

Showrunner leverages advanced AI algorithms to transform user prompts into animated episodes. Users start by selecting a genre that resonates with their creative vision, such as sci-fi, fantasy or historical drama. They can then customize their creation with specific prompts.

For example, I might choose “cyberpunk” and add details like “mysterious detective” with “a touch of Blade Runner.” The outcome? A distinctive animated series that blends genres and reflects the user’s creative input. Whether it’s anime, classic noir, or a fusion reminiscent of Tarantino, Showrunner brings these visions to life.

Beyond genre selection, there are a range of features to set the tone and style of the series. Input specific prompts help guide the AI in generating unique content tailored to the user’s creative vision.

Episodes can range from short scenes to longer stories and AI-generated dialogue and voice help to bring the characters to life. Users have complete control over scene composition and easy editing ensures the final product aligns with the user’s creative direction.

What users can expect

I joined the waiting list almost immediately after Showrunner was announced and long before the website was final. I was so eager to get on the platform and start creating. But when I was finally allowed in to the ultra-secreative site, I was underwhelmed.

While the concept is pretty cool, there are a few drawbacks. For one, getting off the waiting list. I was among the first users, but now the site says there are currently 50,000 people on the waiting list. If you’re like I was and eager to get started generating your animated series, don't sweat, you can skip the line if you write a good enough reason why you want to join.

By scrolling to the bottom of the screen and applying, you can submit a short but thorough questionnaire about what you will do with your Showrunner alpha access.

Once you are “in” you’ll be sworn to secrecy. You will have to promise that you won’t share any of the animation you generate outside of the confines of Showrunner until the official site launches. It’s bittersweet. You have access to all the cool tools once you’re off the waiting list, yet, no matter what you create, you can’t show the world for fear of getting kicked off the site.

However, upon engaging with Showrunner, users can anticipate a community of like-minded animators on Discord and an opportunity to share within that community of AI-driven storytellers. You can share your animations with fellow users and get feedback. I’ve found this to be helpful and inspiring, but maybe that's just because that's the only place I can share.

The user interface of the platform is intuitive and simplifies the content creation process. This is especially helpful for users who may have creative goals but not technical expertise. While users can expect a surprisingly quick turnaround from concept to finished episode, each episode can only be between two and sixteen minutes.

A 30-minute episode of a regular animated series is 22-minutes to leave room for commerials. It's a bummer that Showrunner didn't give us those extra minutes to really give us the Matt Groening vibe.

Future prospects

Showrunner opens the door to a future where technology and creativity coexist, offering novel possibilities for storytellers and audiences alike. Looking ahead, Showrunner aims to expand its AI-generated animation capabilities significantly. Plans are underway to introduce multi-season story arcs, enhancing the complexity and depth of episodes.

This evolution promises to elevate the platform beyond basic animations to more intricate narratives that unfold across multiple episodes. Hopefully, the site will allow me to share my animations with you soon. Until then, maybe I'll see you in the Showrunner Discord community.

More from Tom's Guide

• OpenAI CEO Sam Altman just shared a massive update on what's next for ChatGPT
• Adobe Firefly Video is here to take on Sora with new AI video generator
• I tested Gemini 2.0 vs Perplexity with 7 prompts created by DeepSeek — here's the winner

Adobe today unveiled its latest innovation in generative AI within it's Firefly application, now featuring the industry's first commercially safe AI video generation model. This development aims to provide users with an all-inclusive creative tool to generate images, vectors, and videos while ensuring intellectual property (IP) safety.

The new Firefly app gives users the opportunity to ideate and create production-quality work with creative control and multi-modal workflows all within Adobe’s Creative Cloud applications.

Users can generate images, edit them, transform them into videos, and apply cinematic movements before transitioning to applications like Photoshop, Premiere Pro, and Adobe Express for further refinement.

Meet Firefly Video

Firefly Video Model is currently available in public beta. This AI tool lets users generate video clips from text prompts or images, control camera angles, create professional-quality images from 3D sketches, craft atmospheric elements and develop custom motion design elements.

Touted as the industry's first commercially safe AI video generation model, it powers the "Generate Video" (beta) feature within the Firefly application and the "Generative Extend" (beta) in Adobe Premiere Pro. This model ensures that the generated video content is IP-friendly and suitable for production use.

A key differentiator of Adobe's Firefly Video Model is its focus on producing high-quality, usable short clips, as opposed to longer, less usable ones. Alexandru Costin, Adobe's vice president of generative AI, emphasized the company's commitment to quality, stating, "We actually think that great motion, great structure, great definition scheme, making the actual clip look like it was film, is more important than making a longer clip that's unusable."

The Firefly Video Model initially supports 1080p resolution, and Adobe plans to introduce a 4K model for professional-level production work in the near future.

New subscription plans

As part of the release, Adobe has introduced two new subscription plans: Firefly Standard and Firefly Pro. Available today with early access pricing, both options give creators unlimited access to Firefly imaging and vector features as well as tiered capacity for new video and audio features.

Starting at $9.99 USD, Firefly Standard offers access to 2,000 video/audio credits per month and up to 20 five-second 1080p video generations.

Firefly Pro offers users access to 7,000 video/audio credits per month, up to 70 five-second 1080p video generations and starts at 29.99 USD.

Additionally, a new Firefly Premium plan designed for professionals looking to generate video content on a regular basis will be coming soon to provide even more audio and video capacity for high-volume creators and teams.

Safety and what's next

By focusing on IP safety and providing robust tools for content creation, Adobe aims to ensure creatives can produce high-quality, production-ready content confidently.

As the Firefly Video Model progresses through its beta phase, the creative community eagerly anticipates the innovative possibilities this technology will unlock. We will put Firefly Video to the test ourselves to see if just how good this model is.

More from Tom's Guide

• I tested Gemini 2.0 vs Perplexity with 7 prompts created by DeepSeek — here's the winner
• Google Gemini now lets you summarize YouTube videos — here's how to do it
• I write about AI for a living — here's my 7 favorite free AI tools to try now

What if your selfie could do more than just capture a single moment? What if it could transform into a lifelike AI avatar that talks, moves, and expresses itself just like you? Well, mostly like you.

That’s the promise behind Selfie Avatars, the latest experimental feature from Synthesia, a leader in AI-powered avatars. This cutting-edge AI tool allows users to generate personalized avatars from just a few images of themselves, unlocking new possibilities for self-expression, digital storytelling, and communication.

According to Sundar Solai, Senior Product Manager at Synthesia, Selfie Avatars provide a new level of personalization, control, and creative freedom for users looking to enhance their digital presence. By uploading a small set of selfies, users can create an avatar capable of wearing different outfits, appearing in various settings, and delivering voiceover-driven short-form video content.

“This experimental release gives users the opportunity to shape the future of AI avatars,” said Solai. “With our upcoming EXPRESS-2 model, we aim to take it even further—allowing for enhanced body language and greater control over avatar performance.”

How to try it

To try Selfie Avatars, visit the official landing page at Synthesia.io. From there you’ll be asked to set up an account and upload several selfies. I recommend uploading at least ten images of yourself using a variety of angles and include full body shots. The better the lighting and clarity of your selfies, the better the outcome of your Selfie Avatars.

While the innovation is exciting, the company has been upfront about technical limitations. Some users may notice that avatars don’t always perfectly capture their likeness, and minor discrepancies in facial expressions or proportions may occur.

I have been playing with the tool and testing it for about a week. I noticed that at times the Selfie Avatar seemed to struggle with hand motions and the facial expressions did not always match the tone. Also, sometimes the background noise that has been added to the videos is more distracting than realistic. The team at Synthesia welcomes feedback and is continually working to improve their AI tools.

"We believe in transparency," Solai explained. "User feedback during this phase is critical in helping us refine the technology and deliver an experience that exceeds expectations."

Synthesia is actively encouraging users to try the feature and share their experiences on social media platforms like Instagram, X (formerly Twitter), YouTube, and TikTok by tagging @SynthesiaIO.

Pushing the boundaries of AI-generated avatars

In an era where deepfake concerns and AI ethics are at the forefront of discussions, Synthesia has taken safety measures to ensure that Selfie Avatars are developed responsibly. The company has partnered with Risky Business Solutions, a team of trust and safety experts, to identify and mitigate potential risks.

To prevent unauthorized deepfakes, a biometric consent flow is required during avatar creation. For instance, while you could create a Selfie Avatar of you in what looks like the Oval Office, it is against Synthesia’s policies to include anyone who might resemble the president. Similarly, no celebrities or prominent figures should be included within the prompts.

Strict policies are in place to block hate symbols, nudity, and explicit material, ensuring an inclusive platform. Recognizing that AI image generation can reinforce racial and gender biases, Synthesia has fine-tuned its system to reduce these risks.

A glimpse into next-gen AI

Synthesia has positioned Selfie Avatars as a glimpse into the next generation of AI-driven self-expression. While this initial rollout offers users the ability to experiment with their avatars, it also sets the foundation for future advancements in AI video production, digital communication, and content creation.

“Selfie Avatars are just the beginning,” said Solai. “We’re building towards a future where digital identity is not only customizable but also safe, ethical, and empowering for everyone.”

As AI avatars become more integrated into digital communication, content creation, and virtual interactions, Synthesia’s Selfie Avatars could be a game-changer. The feature offers creators, professionals, and everyday users a new way to engage with audiences in an interactive, dynamic format.

More from Tom's Guide

• I used ChatGPT to pick out the perfect Valentine's Day gift — here's what happened
• Here are 7 of my favorite AI tools — and they're all free
• Elon Musk reportedly bidding $97.4 billion on OpenAI

Today, Pika, the AI-driven social video platform, released its latest feature, "Pikadditions," further expanding its suite of creative tools for its 12 million users. This launch follows the recent release of the 2.1 model upgrade, which introduced high-quality 1080p resolution, improved character control, and cinematic camera movements.

Meet Pikadditions

Pikadditions offers a video-to-video model that allows users to incorporate any person or object into their existing videos. Whether it's their daily Instagram Reel or a favorite movie clip, creators can now elevate their content by adding unique, surreal, or comedic elements designed to captivate their audiences.

Pika's Founder and CEO, Demi Guo, says, "Our goal with Pikadditions is to empower users to push the boundaries of their creativity," said Guo. "We believe that by making advanced AI tools user-friendly and accessible, we can inspire a new wave of content creation that is both innovative and fun."

Here's how to try it

Users can enhance their videos in two distinct ways. Firstly, users can add a little extra magic to their videos and enrich their original footage by inserting surprising objects, people, or fantastical elements yet still preserving the original audio. Want to add a unicorn to your daily commute or a pirate to your pool party? You can do that with Pikadditions. This feature encourages imaginative enhancements to selfies, point-of-view videos and anything else users can create.

Another way users can use the feature is by inserting themselves into iconic clips. Perhaps you’ve wanted to be in your favorite film – now’s your chance. Pikadditions enables users to place themselves into any video, maintaining the original audio. Imagine becoming a character in a blockbuster film, walking on the moon, or sitting beside a celebrity at an awards show. Here’s your chance to explore the world in any way you can imagine.

Accessible to all users

Pikadditions is available across all subscription tiers, from Basic (free) to Fancy. Basic users receive 15 free Pikaddition generations per month, ensuring that everyone can experiment with and enjoy this new feature.

With the launch of Pikadditions, Pika continues to solidify its position as a leader in AI video innovation, making creative video editing accessible and enjoyable for all users, not just professionals. As AI technology continues to evolve, Pika remains at the forefront, dedicated to bringing the joy of video creation to a global audience.

More from Tom's Guice

• Amazon announces Alexa event on February 26 — what we expect to see
• Google’s Super Bowl ad had to be changed last minute — and it’s all Gemini’s fault
• ChatGPT Search is now open to everyone — no account required

Artificial intelligence has already begun to change the way we conduct research, and now OpenAI just released a fresh update to ChatGPT that, it says, will open up "deep research" to just about anyone.

Following just days after its 03-mini reasoning model, the new deep research tool is launching today for ChatGPT Pro subscribers.

"Today we’re launching deep research in ChatGPT, a new agentic capability that conducts multi-step research on the internet for complex tasks," a blog post reads, promising that the model "accomplishes in tens of minutes what would take a human many hours."

This new agent "can do work for you independently", acting on a series of prompts to "find, analyze, and synthesize hundreds of online sources to create a comprehensive report at the level of a research analyst".

It's powered by the upcoming o3 model of ChatGPT and is geared towards finance, science, policy and engineering research, while OpenAI also says that it can helpful for making purchasing decisions, too, helping run detailed product research (another shot across Google's bow, perhaps).

Using the new tool is simple, too. Users just need to select 'deep research' in the ChatGPT message composer and enter the query, with the option to attach additional context via files and spreadsheets.

Doing so will see articles cited in the sidebar, and the progress can be tracked there too. OpenAI says it can take five minutes to half an hour to complete a deep research task.

Limitations

OpenAI has stressed that there are limitations, however.

"It can sometimes hallucinate facts in responses or make incorrect inferences, though at a notably lower rate than existing ChatGPT models, according to internal evaluations."

"It may struggle with distinguishing authoritative information from rumors, and currently shows weakness in confidence calibration, often failing to convey uncertainty accurately. At launch, there may be minor formatting errors in reports and citations, and tasks may take longer to kick off. We expect all these issues to quickly improve with more usage and time."

Still, if you want to give it a whirl, you can do so today — providing you're paying for ChatGPT Pro.

More from Tom's Guide

• I tested DeepSeek vs Qwen 2.5 with 7 prompts — here’s the winner
• I had 7 AI chatbots predict the Super Bowl — here’s the overall winner
• iOS 18.3 proves Apple Intelligence is far from finished

Luma Labs has given its popular Dream Machine AI creativity platform a major upgrade, bringing the new Ray2 video model into the system. This is a huge upgrade over the previous Ray 1.6, offering better realism and more natural motion.

Ray2 was announced last year as part of a new partnership with Amazon AWS. It has finally been integrated into Dream Machine, available as the default option when you create a video.

The AI startup describes Ray2 as “a new frontier in video generative models.” To achieve the level of visual and motion realism they increased compute power 10x compared to previous models, which has “unlocked new freedoms of creative expression and visual storytelling.”

I’ve been testing Luma's Ray 2 since launch and the video generations are very impressive, but it's slow due to demand with some issues around clips refusing to generate or taking too long to be useful. These are all the same teething problems any platform has when launching a new model. So I'd say Ray2 is certainly in the running to be one of the best AI video generators available.

Putting Ray2 to the test

Being built into Dream Machine already gives Ray2 a leg up compared to other video models because of how impressive Dream Machine is to work with. It makes creating content with AI more collaborative and less about throwing a prompt into the wind and hoping for the best.

Accessing Ray2 is as simple as starting a new board in Dream Machine, selecting Video from the prompt bar and typing your prompt. The AI will handle the rest, showing you two videos and giving the usual adaptable interface where you can change elements of the prompt.

Due to the issues I mentioned earlier I was only able to get about half of the prompts I tried to actually generate, and because of how slow it was I couldn’t make use of the re-prompting and collaboration features that make Dream Machine so good. Despite that, I was still impressed.

In one example I asked Luma's Ray2 to create a video of a knife slicing into an onion. Knife skills is something no video model — with the exception of Google Veo 2 — has been able to do consistently well. While it wasn’t perfect, the motion was spot on and it did slice.

Ray2 is also particularly good at animal motion. I asked it to generate several videos of dogs — including one stretching and another catching butterflies — and it did both very well. There were some elements of the butterfly video that were not perfect, but with Dream Machine that can be relatively easily corrected by replying to the original video and specifying what to change.

When I was able to get videos to generate it happened extremely fast. It doesn’t look like Luma had to sacrifice the speed of generation (something Dream Machine was famous for) in favor of improving quality. We seem to get performance and speed all in a single model.

Luma Labs Ray2 Final thoughts

Overall it does appear that Ray2 is a significant step forward in generative video. Its leap feels very similar to the jump we saw when Luma first launched Dream Machine. It is also slightly better in terms of motion than OpenAI’s flagship Sora model.

It is not perfect. There are still artifact issues; sometimes the motion doesn’t make sense; and its currently only text-to-video (although image-to-video is coming soon). However, these are all problems that plague every other model I’ve tried, including Sora, Runway, Kling and Pika.

The biggest takeaway is just how fast AI video is evolving. Being able to generate ten seconds of high resolution video nearly indistinguishable from something filmed with a camera would have been unthinkable two years ago. Today, its commonplace and available from multiple companies.

More from Tom's Guide

• Best food delivery services: Grubhub vs Uber Eats vs Doordash
• The best cast iron skillets 2024: Tested and rated
• How to make images with AI using Leonardo

Luma Labs, the startup behind the popular Dream Machine AI creativity platform, is releasing a new video model. Ray2 promises to be faster, more realistic, and offer a better understanding of real-world physics than other, similar models such as Runway’s Gen-3 or even OpenAI Sora.

According to Luma, Ray2 can distinguish interactions between different objects and object types. This includes between humans, creatures and vehicles, adding to the realism.

Ray2 will be able to create up to 10 seconds of high-resolution video from a text or an image prompt and will be available through Amazon's AWS Bedrock as well as in Dream Machine.

According to Luma, the model will create clips showcasing advanced cinematography, smooth motion and "eye-catching drama." It is a native multimodal architecture that was trained directly on video data to improve the accuracy of characters.

While it isn’t available yet, there are dozens of clips made using Ray2 appearing on social media and I’ve gathered some of the best I’ve seen.

1. A ball of matter

First up is a video shared by Luma CEO Amit Jain showing a ball of opaque matter floating and rotating counter to the motion of the camera. This would be great for planet scenes. He said: “Scaling video model pretraining is leading to far more accurate physics!”

2. Woman on a motorbike

The next clip to draw my attention was shared by self-titled “AI rockstar” Ryan J Phillips (aka Uncanny Harry) and shows stunning Mad Max-esque footage of a woman on a motorbike in a desert. He said he was “enjoying the improvements they have made to motion.”

3. Sweeping mountain vista

I loved this sweeping mountain vista from Luma’s Theo Panagiotopoulos. It shows how well the model handles camera motion over complex terrain. Other models develop a slight “game-like” style in these scenes.

4. Cat on a couch

Slow but complex motion can be difficult for AI models, especially when accurately depiction a living creature but this video shows how Ray2 gets us closer to the ideal. Shared by Luma’s William Shen it depicts a cat tentatively moving across a couch before leaping.

5. Man on a tightrope

Finally, this might be one of my favorite videos I’ve ever seen and was created by filmmaker Allen T using Ray2. In it a man tentatively crosses a ravine on a tight rope and shows some very accurate footwork. Allen wrote: “Look at the physics in the Ray2 generation… Holodeck is coming.”

Final thoughts

AI video is advancing faster than almost any other form of artificial intelligence. The next big leap is in an understanding of reality, and with Google’s Veo2 and the new Ray2 from Luma, we might be getting close.

I haven’t tried Luma Ray2 yet, it hasn’t been released. But from what I’ve seen from both Luma staff and others that were given early access — its one to watch.

More from Tom's Guide

• Best food delivery services: Grubhub vs Uber Eats vs Doordash
• The best cast iron skillets 2024: Tested and rated
• How to make images with AI using Leonardo

I’m always looking at ways to streamline my workflow. Take editing photos, which a single image could take at least 10 minutes to tweak and edit — and even more when I have to painstakingly remove unwanted stuff. I thought that I could shave down my editing time with the Pixel 8 Pro’s AI-assisted features, like Magic Eraser and Magic Editor, but I would soon learn that there’s still room for improvement.

AI will be one of the pillar features in phones going forward, so expect to see it in next year’s batch of flagship phones such as the Samsung Galaxy S24 Ultra, iPhone 16 Pro Max, and many more. So far, the Pixel 8 Pro is just the beginning for how AI is helping us with our everyday tasks. I’m still impressed by its upgraded Call Screen feature, where Google Assistant takes calls for me when I’m unavailable.

After trying out some of its AI photo editing features, I realized that I was still spending time fine-tuning its editing — so it really wasn’t saving me a tremendous amount of time. Sometimes the Pixel 8 Pro was unable to properly isolate subjects that I wanted to erase/move, which required me to manually add my selections. In another situation, I found its AI image generator failing to fill in gaps with something practical.

Read on to see how the Pixel 8 Pro’s AI photo editing tools failed me.

Blurry under the CES arch

CES is a zoo filled with people. You don’t have to walk far before you’re surrounded by a gang of people. I tried using Magic Eraser to remove all the people underneath the CES arch near Central Hall in the Las Vegas Convention Center, but I’m not convinced by the results.

While Google Photos will recommend what subjects to remove, I had to manually make additional selections to get most of the people out of the show. Unfortunately, there are clear instances of where it struggled to camouflage the subjects I removed. You might not see it as much when you look at the overall picture, but you can when you zoom in.

To be fair, though, Magic Eraser works best when you're trying to delete one or two subjects or objects. 

Cat-on-cat bot

I take more pictures of my cats than I do of people. In this shot, I wanted to resize my cat, Vader, to make him look smaller.

I used the Pixel 8 Pro’s Magic Editor feature to do this, but it hilariously failed with what it generated behind Vader. When you resize a subject using Magic Editor, or move them to a different area in the photo, it’s using an AI image generator to fill in the area — and in this case, it looks like a gray cat standing up with his back to us in the photo.

I’m not sure why it couldn’t just apply a softening effect to that area, but this shows that there’s more improvement needed for machine learning to properly analyze a photo and deliver a proper image generation.

In here it could be Friday

I went out to grab drinks at the local TGI Friday’s near me, so I pulled up to the bar and saw that classic line sprawled out in the dining room of “in here it’s always Friday.” There were a few people on the other side of the bar that were captured in the shot, so I tried using Magic Eraser to remove them.

From a cursory look, it seems to do a decent job at erasing them. Upon closer inspection, I can visibly see oddities in those areas. The vicinity where the couple’s sitting at the table almost looks more like a mosaic than an actual photo.

Greetings from the mosiac brewery

That mosaic looking effect becomes even more prominent in this photo of people lining up at one of my favorite craft breweries at the Jersey Shore. This photo is very busy, and as I suspected, it would give any photo editing removal software trouble.

Fortunately, Magic Eraser was able to discern the outline of the building, so it remained relatively boxy — while the rest, well, looks pretty bad. There are clear areas of visible masking applied to the photo and it couldn’t retain the straight lines running on the side of the building. It almost looks like someone painted over some areas of the photo.

Can't fool the yacht down by the bay

And as a finale, here’s the result of using Magic Eraser to once again remove unwanted subjects from the scene. I was actually surprised by this one because it could almost pass. Although, I did spend over 10 minutes manually selecting the subjects because the Pixel 8 Pro had difficulty doing it on its own.

I do like how it manages to erase the chairs and people from the shot, but any photo editor can tell it’s a poor masking job because of the inconsistencies with the ripples in the water — which appears much softer in tone than the rest

The Pixel 8 Pro's AI is great when it’s not a busy scene

There’s certainly room for improvement here, but there are circumstances when I found the Pixel 8 Pro’s AI photo editing features delivering good results. Starting with Vader underneath the Christmas tree, I was really impressed by Magic Editor’s ability to realistically resize him and generate a convincing addition to where he was originally — making the tree look incredibly humongous.

Secondly, I like how Magic Eraser was able to remove this woman from the foreground in this shot of the Bethesda fountain in Central Park. If you’re a pixel peeper such as myself, then you’ll notice the odd brick pattern where the woman was removed.

The takeaway from all of this is that these AI photo editing features best work in photos that aren’t too busy and have enough contrast to allow them to distinguish foreground subjects from the background. I’m excited to see what the next generation of AI-powered phones have to offer.

I would love for a phone to have the same generative AI features similar to Adobe Photoshop’s Generative Fill feature, which lets you make selections and let AI fill them in with whatever you want. I’m just hoping generative AI features such as this don’t end up being locked away behind a paywall, which is what’s rumored to happen with the Galaxy S24 Ultra. I hope not.

More from Tom's Guide

• WALT is a new AI video tool that creates photorealistic clips from a single image — you have to see it to believe it
• Watch out ChatGPT - this OpenAI rival just unveiled a new chatbot that's totally free
• I just tried Meta’s new AI image generator — and it gives DALL-E a run for its money