OpenAI just released Images 2.6, and ChatGPT can finally generate animated GIFs.
Naturally, I immediately tried to break it. But first, the official news.
ChatGPT Images 2.6 introduces:
Animated GIF generation: Create looping animations from text prompts or still images.
Saved reference images: Reuse your face, characters, pets, and other visual references across conversations.
50% faster image generation: According to OpenAI's announcement.
Better editing consistency: Preserve character appearance, clothing, products, and other details across revisions.
Two image models: GPT-Image-2.6 Flare for speed and Sunburst for higher-quality generation.
Developer features: Animated GIF output, image markup, SVG generation, and adjustable complexity through the API.
Both Flare and Sunburst are already available in the API. And yes, you can now ask ChatGPT to animate a character, create a looping reaction, or bring a reference image to life.
Sounds wonderful.
So I decided to see what happens when you push this thing beyond the beautiful examples in the announcement.
Here's what I discovered.
First, resolution.
I tried generating GIFs at different sizes.
1024 × 1024: Couldn't get it to work.
883 × 883: Only managed two frames.
308 × 308: A much more practical result with 16 frames.
These are results from my experiments, not universal resolution limits published by OpenAI.
Then I started increasing the frame count.
And things got interesting.
16 frames: The sweet spot.
The animation was relatively coherent, movement looked reasonable, and the result was actually usable.
64 frames: Things started getting weird.
The animation became unstable.
Objects began jumping around, the composition drifted, and even when I explicitly requested a static camera, the entire image started shaking.
128 frames: Structural collapse.
Characters and background elements increasingly lost their consistency.
256 frames: Congratulations. You've invented a new form of digital suffering.
The animation basically fell apart.
Motion, geometry, continuity, everything.
At that point, I wasn't testing an animation tool anymore.
I was documenting the gradual destruction of visual reality.
And there's an important technical distinction here.
The current public OpenAI image API documentation specifies 2 to 16 frames per GIF, with nine frames as the default.
Yet in my ChatGPT experiments, I was able to request and receive much longer animations, including 128 and 256 frames.
Being able to request that many frames, however, doesn't mean the model can maintain useful visual consistency across them.
And based on my tests, it definitely can't.
Now for the interesting part: Reference Images.
I uploaded a photo of my own face.
And suddenly, the results became noticeably more stable.
Less random visual noise.
Better character consistency.
The reference helped the model maintain recognizable features instead of constantly reinventing them.
In some of the higher-resolution results, I even got a few visual flashbacks to Sora.
Not always in a good way.
But the difference was significant.
Another limitation worth knowing:
ChatGPT can edit GIFs it generated itself, but it currently doesn't support editing an arbitrary animated GIF that you upload.
You can upload a still image and animate it, though.
And now for the genuinely interesting news.
While experimenting with my own face, I started looking into OpenAI's rules for generating real people.
Because remember when AI image generators seemed terrified of anything resembling an actual human being?
Well, the situation is more nuanced than that.
There's no longer a blanket technical prohibition on generating images of adult public figures.
OpenAI explained this approach in its earlier native image generation system card.
Instead of blocking all depictions of public figures, the company introduced more targeted safeguards.
That means certain historical, educational, artistic, or satirical depictions can be generated.
OpenAI's Service Terms still prohibit reproducing someone's likeness without express consent and the necessary rights.
The rules also prohibit using visual capabilities to identify people or extract private or sensitive information about them. And there are additional restrictions around harmful, deceptive, and exploitative content.
Official Images 2.6 announcement: https://openai.com/index/introducing-images-2-6/
ChatGPT image generation and GIF documentation: https://help.openai.com/en/articles/11084440-images-in-chatgpt
Image generation API documentation: https://developers.openai.com/api/docs/guides/image-generation
OpenAI Service Terms: https://openai.com/policies/service-terms/
OpenAI's earlier policy explanation on public figures: https://deploymentsafety.openai.com/4o-native-image-generation/public-figures
#ChatGPT #Images26 #AIGIF #OpenAI #AIAnimation #ImageGeneration #GenerativeAI #AIVideo
YouTube / ABVX