undefined | Better HN

0 pointsdrcode4y ago0 comments

I also read somewhere that this system has special logic added to it that judges how humans would aesthetically judge the final image, so in a way the impressive aesthetic qualities of these images isn't totally coincidental.

0 comments

1 comments · 1 top-level

amilios4y ago

This makes sense, OpenAI has been experimenting with human-in-the-loop type refinement systems for GPT-3 (see InstructGPT https://openai.com/blog/instruction-following/). It would make sense that they would use something similar for Dall-E 2 as well.

j / k navigate · click thread line to collapse