Hacker News Reader

Hm that's true.

We're working on ways to evaluate if an image looks "off" - but its a hard problem because a lot of it is subjective

Yeah, I didn't mean that as a criticism, it's probably impossible to account for everything, and what is off in one context might be fine in another. Filtering out images that contain messed up objects and body parts still hugely reduces the amount of images a human has to look at, so that's valuable either way.

I have no clue about SD other than toying with it, so forgive if this is naive, but do you have any plans to maybe feed this back into the generation, e.g. jiggle the the seed or other parameters around a little until no fingers are malformed?

Yeah I think it's a great idea! Right now because the models are so finnicky changing the seed often does lead to an image with desirable qualities(ie: good hands) so this makes sense.

I do think over time the feedback loops will look much different, as the models get more solid.

Reply on news.ycombinator.com