ChatGPT’s new Images 2.0 model is amazingly good at generating text

It used to be easy to tell the difference between human-created images and AI-generated images. Just two years ago, you couldn’t use image models to create a Mexican restaurant menu without inventing new dishes like “entuitas,” “churillos,” “burtos,” and “margartas.”

Now, if you ask the new ChatGPT Images 2.0 model to create a Mexican menu, it will create something your restaurant can use right away without your customers even knowing anything is wrong. (However, for a $13.50 ceviche, you might question the quality of the fish.)

For comparison, here are the results I got from DALL-E 3 two years ago (ChatGPT didn’t generate images back then).

AI image generators have historically struggled to explain themselves, as they commonly used diffusion models that worked by reconstructing images from noise.

“Diffusion model” […] Asmelash Teka Hadgu, founder and CEO of Lesan AI, told TechCrunch in 2024. “The writing on the image can be assumed to be a very small portion, so the image generator learns patterns that cover more of these pixels.”

Researchers have since investigated other image generation mechanisms, such as autoregressive models that predict what an image should look like and work like LLMs.

Unfortunately, OpenAI declined to answer questions about the type of model powering ChatGPT Images 2.0 at a press conference this week.

But the company said the new model has “thinking capabilities” that let you search the web, create multiple images from a single prompt, and double-check what you’ve created. This allows Images 2.0 to create marketing assets of various sizes, as well as multi-panel comics.

OpenAI also says that Images has a better understanding of text rendering in non-Latin languages ​​such as Japanese, Korean, Hindi, and Bengali. This model’s knowledge will end in December 2025, which could impact how accurately it can generate specific prompts about recent news.

“Images 2.0 brings an unprecedented level of specificity and fidelity to image creation. You can not only conceptualize more sophisticated images, but actually realize that vision effectively, following instructions, preserving requested detail, and rendering fine-grained elements that often break image models (such as small text, iconography, UI elements, dense composition, subtle stylistic constraints, etc.) all at up to 2K resolution,” OpenAI said in a press release.

These features mean that while image generation isn’t as fast as typing a question into ChatGPT, it takes just a few minutes to generate something as complex as a multi-panel comic.

All ChatGPT and Codex users will have access to Images 2.0 starting Tuesday. Paid users will be able to generate more advanced output. The company will also offer a gpt-image-2 API, with pricing depending on output quality and resolution.