AI images
(media.kotakuinaction2.win)
You're viewing a single comment thread. View all comments, or full comment thread.
Comments (43)
sorted by:
It's not that hard, conceptually, to define "a finger" alongside "a wrist", "a thumb", and that a hand has exactly 4, 1, and 1 in such and such approximate locations.
It would be "seeding" info, rather than a "pure" learning algorithm, though I don't see why having hard-coded prompts in a tool such as an art-asset-maker is a bad thing. You can seed into a chatterbot that it is to roleplay as a game show host, you can certainly seed into a painterbot the concept of a "hand" in its Platonic pure form.
But at that point you're writing an art-making program, not fleecing the public with fancy AI words, so...
I think it just needs a step back and new approach to that, with something like the seeding info yeah. I think working off of a 3d model though would solve it.
3d generation of 3d models, with an adjustable default base mesh, then rendered pictures from that, that's what will fix the fingers. The issue with the hard coded seed is that you don't always see 5
3d generation has just started, it's as primitive as dall-e mini was when everything went wild. So give it another 2 years.
My dude nobody here is a normie.