Now ChatGPT's Images 2.0 can generate a restaurant menu that looks restaurant-ready. No misspellings. No made-up dishes.
Just clean, usable output.
That's a real shift.
The old problem was the model type.
Diffusion models treat text in images as just more pixels, so letters came out garbled.
Researchers moved toward autoregressive models, which think more like a language model and actually plan what the image should look like.
OpenAI won't say exactly what's powering Images 2.0. But the results speak for themselves.
It can now handle small text, icons, UI elements, dense layouts, and multilingual text in Japanese, Korean, Hindi, and Bengali.
It can also generate multiple sizes of a marketing asset from one prompt or build a multi-panel comic strip in a few minutes.
Paid ChatGPT users get more advanced outputs. The API is live too, with pricing based on quality and resolution.
For content creators and small business owners, this is the first time AI image output actually looks close to production-ready.
Not perfect. But close enough that the gap between "AI made this" and "a designer made this" just got a lot smaller.
Full breakdown with images is on TechCrunch if you want to see the before/after comparison yourself.