OpenAI drastically improves ChatGPT image generation

Moving from Dall-E to native GPT 4o images makes for higher quality and more lifelike detail — and they also updated their terms of service.

Rolling out to users on Free, Pro and Plus tiers, the image generation should be live pretty much by the time of this writing, letting you generate just about anything you can imagine from a text prompt – within reason.

Accessing a world of knowledge
Since it uses GPT 4o, the generator has access to a world of knowledge and context that was missing from the previous generators, writes Mashable.

The model is not particularly fast, though, with Sam Altman warning that generating detailed images can sometimes take up to a minute.

Allowes users to be «offensive»
Perhaps the most interesting part is the update to the terms of service agreement, which, while still banning things like sexual deepfakes and child exploitation, does allow the user to generate «offensive» imagery.

This is in order to facilitate «Intellectual freedom,» OpenAIs Sam Altman said at launch. They want to put these creative choices in the hands of the user without overly censoring the output:

— What we’d like is for the model to not be offensive if you don’t want it to be, but if you want it to be within reason, really let people create what they want, said Altman in the launch stream.

Public figures? IT’s okay
There is also a provision on public figures, which you are now allowed to generate, according to simonwillison.net

This rule comes with a caveat, though, as the celebrities themselves are invited to opt out rather than opt in. OpenAi thinks this will allow for more political satire, and sees benefits in fields like «education, history and political speech.»

All images made with the model also comes with a watermark, to indicate that it’s an ai image.

Late, but good?
The «o» in gpt 4o stands for «omni» and the model was supposed to be multimodal at launch in May 2024, but was severely hampered by the low quality outputs of Dall-E. And Google has had high quality image outputs in its Gemini model from December 2024.

So OpenAI s playing catchup in the market, but they have a much larger user base, and these images could soon flood the net.

Read more: Detailed report on The Verge, Mashable on the live stream, Simon Willison gives it a spin, and OpenAIs official launch post.