Inside image generation’s Renaissance moment — the OpenAI Podcast Ep. 19
- Source
- youtube.com
- Author
- OpenAI
- Date
Context
In an OpenAI Podcast episode, researcher Kenji Hata and product lead Adele Li of OpenAI say its Images 2.0 improved in three areas: text rendering, so words are real and legible; multilingual output; and photorealism, meaning images that keep a person's face and body true to them. They report usage up more than 50% in the two weeks after launch and over 1.5 billion images a week in ChatGPT. Those are company figures.
Hata describes an informal internal test: ask GPT for a list of 100 random objects, send it to the image generator as a grid, and count how many appear correctly. Hata says a grid held about 5 to 8 objects with DALL-E 3, about 16 with GPT Image 1 and 25 to 36 with version 1.5, and that the new model could probably exceed 100, usually almost all correct. It is a company check with no published method, offered as an illustration of steady progress.
- Model — A computational system that uses learned patterns to produce an output from information supplied to it.
Checking sign-in…
Loading comments…






