正文 · AI 翻译
Images in as little as 6 seconds
Hengyu reporting from Aofeisi
QbitAI | WeChat public account QbitAI
Hengyu reporting from Aofeisi
QbitAI | WeChat public account QbitAI
GPT-Image-2.6 has suddenly launched!
Overall generation time is down 50%, with images produced in as little as 6 seconds; it natively supports generating animated GIFs; and it supports direct export in SVG vector format.
For the first time, OpenAI has separated a model’s “scene-planning brainpower” and “pixel-rendering quality” into 2 independent parameters in its API, substantially reducing the cost and time required for multiple rounds of edits.
GPT-Image-2.6 is now fully available in ChatGPT, Codex, and the API, and all free and paid users can use it directly.
Separating “brainpower” from “image quality,” with images in as little as 6 seconds
This time, the GPT-Image-2.6 API has introduced 2 distinct models for the first time: Flare and Sunburst.
Flare focuses on speed and layout, with layout optimizations specifically for commercial visuals such as posters and e-commerce product images. At the low-quality setting, the time to generate a single image has been squeezed down to 6 seconds.
Sunburst tackles the harder problems, providing greater detail at medium and high quality settings. Its main focus is consistency across multiple rounds of conversational edits, so people and products do not change appearance as edits accumulate.
The changes to the underlying control logic are also worth a look.
Previously, when calling an image-generation API, developers could often adjust only the resolution or a broad image-quality setting. Even changing just the background color required the model to run through the entire composition-planning process again.
Images 2.6 has added a parameter called complexity, which controls how much computing power the model spends planning a scene before it starts drawing. The existing quality parameter now deals solely with visual detail and fidelity.
Separating the effort spent on composition from rendering precision follows much the same idea as adjusting the depth of reasoning in large language models.
In practice, for routine local tweaks or quick image edits, developers can set complexity to low while keeping output quality high.
That saves the time the model would spend planning the layout again, leaving it to do only a high-precision redraw.
Complexity needs to be raised only for difficult scenes with many elements in the same frame or a great many constraints.
A 3×2 sequence of frames as a shortcut: native support for animated GIFs
Native support for animated GIFs is the most immediately apparent new feature in Images 2.6.
Whether you want to give a brand mascot a little movement, make a few reaction stickers for social media, or simply hand the model an owl casually doodled with a ballpoint pen on lined paper, the animated stickers it produces have clean backgrounds and fluid motion.
How can an image-generation model quickly make a smooth GIF without help from a heavyweight video-generation model?
The answer lies in the prompt from an official OpenArt collaboration example.
For this set of animations showing an office worker shedding macho tears while watching stocks plunge, then throwing money in celebration as stocks surge, the prompt given to the model actually looked like this:
“Generate a single animation sprite sheet for a seamlessly looping GIF ... Include exactly 6 consecutive frames, arranged in a precise grid of 3 columns × 2 rows, read from left to right and top to bottom. Every panel must be exactly the same size. Use a fixed front three-quarter camera angle, with the same character, desk, monitor, clothing, and background ... The 6th frame should transition smoothly back to the 1st. No captions, borders, or empty space.”
“Generate a single animation sprite sheet for a seamlessly looping GIF ... Include exactly 6 consecutive frames, arranged in a precise grid of 3 columns × 2 rows, read from left to right and top to bottom. Every panel must be exactly the same size. Use a fixed front three-quarter camera angle, with the same character, desk, monitor, clothing, and background ... The 6th frame should transition smoothly back to the 1st. No captions, borders, or empty space.”
See how it works?
Images 2.6 does not use the complex spatiotemporal diffusion calculations of video generation at all. Instead, it borrows the idea of sequential frames from the gaming industry, drawing all 6 closely matched stop-motion frames in a single image—a sprite sheet.
Slice up the image and put the frames together, and there is your GIF.
This approach sidesteps the high inference latency of video models while keeping each frame’s character features, clothing folds, and background objects relatively consistent.
Companies are already putting this capability to practical use.
For example, Town, a startup that makes an AI office assistant, has integrated it through the API specifically to generate real-time animations of subtle actions, such as typing and responding, for virtual assistants in work software.
Direct SVG output, and local image markup opened up to the API
This update also brings 2 quite practical engineering capabilities.
One is SVG vector export; the other is image markup through the API.
Anyone who has worked on front-end development or UI knows that the edges of AI-generated bitmap images look awful when enlarged ... To put them into a website or app, the usual approach is to use a third-party tool to convert them to vector format.
It is quite a hassle.
Images 2.6 natively supports direct SVG export.
When writing front-end code in Codex, programmers can have the model generate icons, components, and illustrations that can be scaled indefinitely, then add them straight to the codebase.
All those intermediate conversion steps are gone!
The local selection and redrawing feature, image markup, which was previously available only on the ChatGPT website, is now officially available through the API as well.
Enterprise teams can use coordinates or masks to mark areas for editing directly in their own product dashboards. Combined with the Sunburst model’s high consistency, they can process image edits in batches like an assembly line.
One More Thing
A small surprise spotted outside the image-generation model:
GPT-6.2 Astra has now appeared on Google Cloud.
References:
[1] https://openai.com/index/introducing-images-2-6/
[2] https://x.com/lyraxana/status/2118123296738677133