Source: New Intelligence Source
Get ready, OpenAI is about to make a big move!
Just now, there was a warm-up, ChatGPT Images 2.6 has received a major upgrade, with the speed of image generation soaring by 50%.
This is the first time, and also a first for OpenAI, AI can output GIFs with one click.
Not only that, GPT Images 2.6 also supports frame-by-frame editing and playback speed adjustment, with reference image functionality launched simultaneously.
A silly rooster stands firm, and after speeding up, it directly performs a chaotic dance.
Today, these two new models, Flare and Sunburst, officially go live.
The former can generate images in as fast as 6 seconds, while the latter emphasizes high quality and editing consistency, supporting SVG export and localized retouching.
Coincidentally, the next generation GPT-6.2 Astra and Sol also appeared on the Google platform, and many in the circle are betting: an official announcement could come as soon as tomorrow!
ChatGPT can finally make GIFs, making a picture come to life
UIS1, the upgrade of GPT Images 2.6 by OpenAI focuses on practicality.
In the past, if you asked the AI to draw a kitten, you would usually receive a static image.
If you wanted the kitten to dance, blink, or wag its tail, you would need to use additional animation software or other generation tools.
This time, ChatGPT has taken over the entire process.
Generated images are no longer monotonous; if you want a photo to come to life, just throw in a line to ChatGPT: "Can you make it into a GIF?"
In no time, the star of the photo "comes alive."
A cat wearing a daisy knitted hat immediately twists its body, while frogs under a disco ball, a proud rooster, and a rabbit on a pink car roof all start dancing to the rhythm.
The most crucial part is that after it's generated, you can still modify it, adding lasers to the blue bird and speeding it up to 2x.
With just a few words, you can totally change the image. The most interesting part of GPT Images 2.6 is that it can also do "frame-by-frame editing."
Pet photos taken casually on a phone can also instantly turn into dancing animations.
Upload photos of your turtle, hamster, and parrot to the chat window, and they quickly transform into a dancing teddy bear in a pleated skirt, a Pomeranian winking amidst rose petals, and a raccoon holding a microphone singing.
Even a hand-drawn sketch of an owl can instantly gain brown feathers, big eyes, and wave to greet you when given to ChatGPT.
Using it to create everyday chat emojis is also very convenient.
A chibi worker on the left crying while staring at a plummeting red line, and on the right celebrating and throwing money while looking at a rising green line; these scenes are cyclical animations automatically stitched together after ChatGPT generated 6 frames of actions at once.
Transforming memes, we tried playing around with it
Since this is the first time ChatGPT can create dynamic images, we couldn't resist giving it a try.
We first tested it with our newly debuted super IP "Super DouDou", with one line, instantly generating a cute image of a Pomeranian frantically typing.
Moreover, we can also create emojis, and the character's traits and hair color details are perfectly captured.
The other IP, Super Pony, gives off a solid sci-fi protagonist vibe just by appearing.
We also asked ChatGPT to create a meme with the "backhanded possum" theme.
With its two little paws reaching back, gazing quietly out the window, the melancholic vibe is perfectly captured.
The blue fat fish has come to life under ChatGPT's brush, eating happily while waving at the programmer.
Continuing to lock in the core persona of the fat fish, we asked ChatGPT to generate an image that flips from "deep thought" to "give me more bowls" in one second.
Next, we let it create an animation seamlessly connecting the set of micro-expressions and small actions like holding the bowl, blinking, and wagging its tail.
Classic scenes from movies featuring Stephen Chow, Zhen Huan, and others can also be played around with.
As soon as the new feature was released, X users already jumped in for some fun.
Tibo has got the reset button.
Professor Derya Unutmaz from Jackson Laboratory generated a GIF of the classic character Cartman from "South Park," who is laughing while looking up and slapping his thigh, with the character's expression well replicated.
Tech blogger TestingCatalog had even more fun, creating an infinite loop animation of a row of small people wearing wide-brimmed hats reading "Daily AI News," which looks extremely amusing.
Besides GIFs, the "save reference image" feature that everyone was anticipating has also been arranged.
Save photos of yourself, family, friends, or pets once, and in future chats, you can directly @ to pull them out anytime, making it both convenient and fun.
4.84 seconds rapid image generation, with a 50% speed increase
I must say, the ultra-smooth user experience on the front end relies on the dual new models logged in at the same time —
GPT-Image-2.6 Flare
GPT-Image-2.6 Sunburst
OpenAI also showcased some might, as global users have generated over 4 billion images weekly on ChatGPT, Codex, and APIs, averaging over 500 million per day.
With such enormous scale, the 2.6 version has increased the overall image generation speed by 50%.
The Flare model is specifically designed for ultra-fast image output and layout design, with the lowest quality setting generating images in under 6 seconds.
Rendering a luxurious travel advertisement poster, the previous generation 2.5 Flare took 7.08 seconds, while 2.6 Flare completed rendering in just 4.84 seconds, increasing speed by 46%.
The image editing interaction has also received a significant upgrade, now supporting redline doodles.
Circle it, draw an X, or drag an arrow; Image-2.6 can accurately grasp these cues and automatically erase the marks afterward.
The following living room remodel image is immediately clear. The right side achieves the final image in one step, even the sunlight casting shadows on the wooden floor transitions quite naturally.
The other core model, Sunburst, emphasizes high detail and consistency in multi-round continuous editing.
In a three-step continuous image editing test featuring a red-haired knight, the final image maintains consistency in her facial features, expressions, and sword-holding posture.
It's worth mentioning that as soon as the two new models were announced, they immediately topped several rankings in blind tests by third-party DesignArena.
Sunburst topped the overall leaderboard in the image arena and won first place in the product and scenery categories; Flare grabbed first place in portraits, architecture, and typography, breaking the quality and speed state-of-the-art benchmarks.
Direct GIFs from a nine-slice layout, native SVG vector image export, more quantity at no extra cost
This time, OpenAI has integrated GIF animations, SVG vector graphics, and control over compositional complexity directly into the native API.
For developers, it’s extremely simple to use.
Just set the output format to gif in the API, and you can freely specify the number of frames between 2 and 64 and the millisecond duration of each frame.
The model first runs out a coherent "nine-slice" sprite sequence, then automatically slices, aligns, and combines it into a seamless looping animated koi fish swimming or carrying a lantern in the snow.
The early testing team, Town, has already utilized this capability to give their AI office assistant, Monty, a lively typing waiting animation.
In response to the pain points of front-end engineers and UI designers, the new models also support native SVG export for the first time.
The generated geometric animal icons and interface buttons come with complete XML vector paths, which can be infinitely enlarged in design software without losing quality and can be easily disassembled for color changes.
OpenAI researcher Gabriel Chua stated that Images 2.6 can also better control image complexity.
Moreover, this major version upgrade has achieved a complete "more quantity without extra cost."
The API pricing for Flare and Sunburst remains the same as the previous generation 2.5, with text inputs costing only $5 per million tokens, image inputs at $8, and image outputs at $30, without any additional charges for doodle annotations and complexity control.
Currently, ChatGPT Images 2.6 is fully open to all subscription tiers of ChatGPT and Codex web, desktop, and mobile users worldwide.
What are you waiting for, a wave of GIFs is already on the way.
References:
https://openai.com/index/introducing-images-2-6/
https://x.com/OpenAI/status/2118124944928432327
https://developers.openai.com/api/docs/guides/image-generation
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到[email protected]，本平台相关工作人员将会进行核查。
