Creators publish 725+ ChatGPT Images 2 prompt-and-output corpus

A prompt library sounds small. But this one is turning into a real piece of infrastructure for AI image work. In the week after OpenAI shipped ChatGPT Images 2.0 on April 21, creat(github.com) and the biggest community repo has now grown past 700 examples. That matters because image models are finally being judged less like art toys and more like s(github.com)her people reproduce the result? (openai.com) ### What actually got published? The clearest thing to p(openai.com) EvoLinkAI. It collects prompt-and-image pairs for GPT-Image-2 across portraits, posters, UI mockups, character sheets, and other community experiments. The repo says most of the cases were curated from X, creator communities, public demos, and shared experiments — basically turning a fast-moving social feed into something you can actually browse and reuse. (github.com) ### Why are people treating this like a benchmark? Because the hard part with image models i(openai.com)ks people actually care about. A shared corpus lets designers and builders run the same kinds of prompts — posters, ads, product cards, app screens — and compare where models break. That is much closer to a real-world benchmark than a handful of cherry-picked launch images. The library’s categories make that practical by grouping prompts around tasks instead of vibes. (github.com) ### Why now? OpenAI changed the baseline. Cha(github.com)ort, and stronger visual reasoning. Those are exactly the features that make prompt libraries more useful, because the tests stop being “can it draw a cool scene?” and start being “can it render a menu board, a poster headline, or a UI mockup without falling apart?” (openai.com) ### What are creators stress-testing? A lot of typography and layout. Another repo built around GPT Image 2 says the model is currently the best public option for legible text inside images and ca(github.com)zine covers, UI screenshots, banners, and multilingual designs. That is a pretty specific shift. Earlier image-model prompt galleries were mostly about style. These are about production tasks. (github.com) ### How big is the collection already? Big enough that it is no longer just a side project. EvoLinkAI’s repo is sitting above 1(openai.com)il 29 alone, the maintainers added 22 new GPT-Image-2 cases covering e-commerce, ad creative, portraits, character work, posters, and comparisons. The repo also tracks prompt-only updates in a JSON file, which makes it easier to turn the collection into tooling later. (github.com) ### Is this the only library? No — and that is part of the story. A separate GPT Image 2 gallery from wuyoscar lists 162(github.com)g them. So the ecosystem is already splitting into two layers: public galleries for inspiration, and structured prompt libraries that agents and developer tools can actually call. That is how a meme format starts becoming workflow infrastructure. (github.com) ### What is the catch? These libraries are still community-curated, not neutral lab benchmarks. They reflect what creators on X and GitHub care about right now (github.com) also the point. Those are the tasks where teams spend real time and money, so a public corpus built around them is often more useful than a sterile benchmark. (github.com) ### Bottom line? The news is not just that people posted hundreds of prompts. It is that GPT-Image-2 seems good enough to trigger open, reusable test suites almost immediately. When creators start pu(github.com)sed into tool territory. (openai.com)