OpenAI is introducing a new image generation capability directly into ChatGPT. This feature is based on the GPT-4o model and will be available to everyone in the near future.
The free usage limit remains the same as DALL-E, but the company didn’t specify a specific number, as it could vary depending on demand. According to ChatGPT’s FAQ, free users were previously able to generate “three images per day using DALL-E 3.” Generation via DALL-E will also remain available via the corresponding GPT.
Because GPT-4o is an “omnimodal” model that works with text, images, audio, and video, it is able to better understand user prompts when creating multiple objects in an image.
Users will also notice improvements in text generation in images. The system uses an autoregressive approach, creating images gradually, from left to right and top to bottom, like writing text. In contrast, most models, including DALL-E, use a diffusion technique, generating the entire image at once. The new system takes longer to generate images, but OpenAI considers this a justifiable trade-off.
The OpenAI team emphasized that the system has robust safeguards to prevent abuse: the tool prevents watermark removal, blocks the creation of sexual fakes, and rejects requests to generate CSAM.
OpenAI’s new image generation system does not include any visual watermarks or indicators that indicate that the images were created by artificial intelligence. However, all generated images will have standard C2PA metadata that identifies them as being created by OpenAI.