by foreverlurk » Mon Dec 05, 2022 10:59 pm
Yeah, that technology is still evolving, after having a fun weekend of messing around with it (not even SW-related), just taking pictures with different prompts, reading and understanding the denoising, samplers, etc. It's really fun.
But I've also come across some of the current limitations, like the 512x512 render and sampling (sampling is actually 64x64, compressed), so anything higher res gets really weird because it can only rendy the first 512 block correctly. There's this "high resolution fix", but it also brings it own problems, like repeated faces (very hard to make a scene with two characters from scratch).
Don't get me wrong, this is pure magic and super fun to mess with, but so far I haven't been able to train it to understand the notion of "size difference". I'm sure it'll be able to, eventually. The inpainting mask is a nice workaround in the meantime. Oh and I've read that they are working on voice sythesis and video tech next! Imagine, being able to produce our own animations...
Edit: forgot about the hands... you know, like in hand-helds, or grabbing a SW, etc... Stable Diffusion gets fingers, limbs, arms, sooo messed up and crank out... **shudders** man-made horrors beyond my comprehension / I've seen things you people wouldn't believe, all that jazz. The horror. Apparently the 2.x releases are better with fingers and limbs, we'll see.
Yeah, that technology is still evolving, after having a fun weekend of messing around with it (not even SW-related), just taking pictures with different prompts, reading and understanding the denoising, samplers, etc. It's really fun.
But I've also come across some of the current limitations, like the 512x512 render and sampling (sampling is actually 64x64, compressed), so anything higher res gets really weird because it can only rendy the first 512 block correctly. There's this "high resolution fix", but it also brings it own problems, like repeated faces (very hard to make a scene with two characters from scratch).
Don't get me wrong, this is pure magic and super fun to mess with, but so far I haven't been able to train it to understand the notion of "size difference". I'm sure it'll be able to, eventually. The inpainting mask is a nice workaround in the meantime. Oh and I've read that they are working on voice sythesis and video tech next! Imagine, being able to produce our own animations...
[b]Edit[/b]: forgot about the hands... you know, like in hand-helds, or grabbing a SW, etc... Stable Diffusion gets fingers, limbs, arms, sooo messed up and crank out... **shudders** man-made horrors beyond my comprehension / I've seen things you people wouldn't believe, all that jazz. The horror. Apparently the 2.x releases are better with fingers and limbs, we'll see.