Hacker News Reader

Online, stable diffusion looks incredible. But it's been so, so bad for me, and when you look closely at lots of other "good" pictures, they have lots of flaws too, we just aren't good at picking them out. I wonder if there will be a way to make SD align with what humans want more, maybe with some GAN approach, without needed actual "understanding" of the real world.

View on HN

Using img2img and making subtle alterations/substitutions to the prompts and/or images as you go can allow one to serve the adversarial role oneself as the user. Ultimately it's sort of like sitting across the table from an artist who can't speak but you can pass drawings back and forth iterating on them until they are complete. Of course it can be drawings, 3D animation style rendering, painting, photorealistic renders, etc.

Try MidJourney instead. Very little is needed to get amazing results. You just set the theme and let it run wild. Or you can also be very specific and it will handle it.

Reply on news.ycombinator.com