ChatGPT's new Images 2.0 model is surprisingly good at generating text
TechCrunch • • general
Key Points:
- AI image generation has significantly improved, with ChatGPT Images 2.0 now creating realistic and usable images, such as a Mexican restaurant menu, without obvious errors that were common two years ago.
- Earlier diffusion-based models struggled with small details like text, but newer approaches, including autoregressive models, enhance image accuracy by predicting image content more effectively.
- OpenAI has not disclosed the specific model behind ChatGPT Images 2.0 but highlights its advanced "thinking capabilities," enabling web searches, multiple image outputs from one prompt, and self-verification of generated content.
- The new model supports better rendering of non-Latin scripts and can produce complex visuals like multi-paneled comics with high fidelity and up to 2K resolution, though generation times are longer than simple text queries.
- Despite its advanced features, the model’s knowledge is limited to data up to December 2025, which may affect the accuracy of images related to very recent events or information.
