ChatGPT 4o gets new update for image generation

GPT-4o solves previous weaknesses in complex image tasksCurrently, many AI models can already create images. However, numerous models fail when faced with complex tasks, such as creating text or retaining details. OpenAI wants to solve these problems with the big update for ChatGPT. Instead of Dall-E, GPT-4o is now used as the model for image generation.

Detailed texts in images are becoming a reality

In a detailed post, OpenAI first explains that it is now possible to create detailed texts with the AI. While other models often only produce rough hieroglyphs instead of legible text, ChatGPT can better implement precise instructions for text thanks to GPT-4o. For example, OpenAI showed an image of a man placing magnets with words on them on a refrigerator. Thanks to the new model, ChatGPT can generate each individual word separately and position it in the correct place. According to OpenAI, the final result only needed five attempts to achieve this.

Complex requests are no longer a problem. Another improvement can be seen in particularly detailed requests to ChatGPT image generation. Many other AI models failed miserably when given more than five to eight instructions.

Note the limitations of GPT-4o

GPT-4o image generation is initially being made available to Plus, Pro, and Team account users, as well as the free version. However, free version users must continue to accept restrictions that already apply to Dall-E. Therefore, they can only generate a few images per day. Enterprise and education customers will have access to the new model at a later date. Dall-E will remain available, but only via a custom GPT model.

Challenges from AI hallucinations continue

OpenAI recognizes that problems may still arise when using the new image generation. For example, so-called hallucinations still occur in isolated cases. When creating a world map with all the capital cities of a country, ChatGPT mentions incorrect country names, invents new terms or lists countries twice. There is also a risk of parts being cut off in particularly large images. OpenAI promises to continue working on these problems after publication and to make appropriate improvements.