Until now, Dall-E was responsible for generating images in ChatGPT. The model can still be used in a custom GPT, but it will be replaced by the "Images in ChatGPT" feature (so OpenAI will never know what to name its inventions). As with Dall-E, the visuals are generated directly in the conversation thread, but they should be of much higher quality.
What changes with the switch to GPT-4o
Images in ChatGPT is based on the so-called autoregressive GPT-4o AI model. It takes a little longer to think, but in return the text rendering is much better and without the usual dross of diffusion models like Dall-E. Object attributes are much better managed, it is up to 20 objects that can be rendered correctly.
While the resolution doesn't change between Images in ChatGPT and Dall-E, OpenAI announces that images can be edited, including those containing people, by retouching the foreground or background elements ("inpainting"). The model was trained using public data — in other words, by digging through the internet — and through partnerships, notably with Shutterstock.
OpenAI, regularly accused of plundering creators, is very careful to respect artists' rights and prohibits the direct reproduction of styles from living artists. Opt-out mechanisms are available for creators who wish to exclude their works from the training data.
The images generated by ChatGPT do not contain any watermarks or indicators that would allow us to say that they are AI visuals. OpenAI nevertheless explains that in the long term, they will integrate C2PA metadata, the standard that is gradually becoming established. small.
The feature, already available to Pro subscribers (the $200 per month plan), will be offered to all paying Plus users, as well as to developers via the API and to users who do not pay for access to ChatGPT. However, they will be limited to a handful of visuals generated per day (OpenAI doesn't specify the number, but with Dall-E it's 3 per day).
Source: OpenAI
0 Comments