ChatGPT takes on Gemini's advanced image generator with a massive upgrade powered by GPT-4o

Google greatly improved Gemini’s image generation capabilities earlier this year with the Flash 2.0 model. The company expanded access to the upgraded image generation experience earlier this month, causing quite a stir when users discovered how good it was at removing watermarks from images. Not wanting to be outdone by Google’s latest image generation tool, OpenAI has now upgraded ChatGPT’s image generation feature with its advanced GPT-4o model.

The improved native image generation experience powered by the GPT-4o model replaces ChatGPT’s legacy DALL-E 3 model, and is rolling out to users on ChatGPT’s Plus, Pro, Team, and Free subscription tiers. OpenAI says it will soon be available on the Enterprise and Edu tiers, and developers will also get access to it via the company’s API service in the coming weeks.

As for the performance improvements, OpenAI claims that the new model delivers images that are significantly more accurate and detailed than those generated by the legacy model. However, it takes a bit longer to achieve these results.

Image generation powered by GPT-4o offers several noteworthy improvements, including better text rendering, the ability to edit and transform existing images or draw inspiration from them, enhanced photorealism and artistic styles, and improved binding of objects to their traits when generating images with multiple objects.

Although better than the legacy model, GPT-4o with image output still has multiple limitations. OpenAI says it can occasionally struggle with cropping long images, rendering non-Latin languages, precise editing, and rendering detailed information in a very small size. It also suffers from hallucinations and can make up information, especially in low-context prompts.

OpenAI has also implemented some safety measures to highlight the AI origin of the resulting images and prevent misuse. All images generated using the tool will come with C2PA metadata, linking them to GPT-4o. ChatGPT will also block prompts to make images with CSAM, sexual deepfakes, and objectionable photos of real people.