ChatGPT has evolved from its earlier reliance on Dall-E for image generation. The newly enhanced GPT-4o model claims to produce images that are precise, accurate, and photorealistic. According to OpenAI, it excels in rendering text, following detailed instructions, and grasping the context of conversations. The model also allows users to transform uploaded images or use them as visual inspiration.
One of the standout features of GPT-4o is its ability to build upon existing images, maintaining consistency of subjects across various images. Unlike many other AI image generators, which typically manage around 5 to 8 objects, GPT-4o can handle between 10 to 20 different objects in a single image, showcasing notable advancements in complexity. However, OpenAI has pointed out some limitations of the GPT-4o model. Users might encounter issues with cropping images, experiencing hallucinations, and managing a high number of elements within a single image.
Additionally, the model may struggle with precise graphs, rendering text in non-Latin alphabets, accurate editing, and presenting dense text within small spaces. The upgraded 4o image generator is currently being rolled out for various ChatGPT users, including those on Free, Plus, Pro, and Team plans. Users with Enterprise and Edu accounts will gain access at a later date. For those who preferred the outputs from Dall-E, there remains an option to revert back to that version for image generation.
Overall, GPT-4o represents a significant step forward in AI-generated imagery, although it still has areas that require improvement.