Birchlabsさんのイラストまとめ

ah, that was the parent commit. here's the real one (with ceiling).
unscale each latent channel (÷0.18215), center on mean. if any channel has a max exceeding 42: threshold by 99.95%.
so only sigmas 8.56, 4.80, 2.55 get thresholded. some clipping remains.
https://t.co/oiSBUldN7j

CFG20, told to get its 99.95%ile down to the same as CFG7.5's.
code here (see CFGDynTheshDenoiser):
https://t.co/1Zn1aefIFL

high CFG scale doesn't always give you stylish pop art. dynamic thresholding can hugely repair an image.
#stablediffusion

think I got dynamic thresholding working.
CFG scale 20. left is what you'd usually get.
salvaged the dynamic range by thresholding the latents on every samping step.
#stablediffusion

fumo embedding works in waifu-diffusion too (no retraining required). WD forgot everything about photos or plushies, but it still understands "make it chibi, give it little T-rex arms and make the sleeves too big".
prompt coherence is way better too (I trained it on booru labels)

I tried to fix the hips but she turns into megaman.
sometimes you have to work within the constraints of what noise you get in the high sigmas of your seed! the Unet predicts that sigma 3.8167 denoises to a wide cloud of legs, so that's what you're up against.

this is with waifu-diffusion 1.3, final epoch.
if you get too many legs: try specifying what legwear was worn

trying out the new DPM-Solver samplers from k-diffusion 0.0.2+ (~1 week ago). outperforms Heun.
dpm_adaptive is great for producing final masters of images; you don't need to guess how many steps to use.
dpm_fast just gave me noise, so sadly can't try it out at low step counts.

IIRC: left vaporwave girl was 8 sampler steps, right was 30. left has a plasticky, specular look which is interesting in its own way. I definitely recommend exploring low sample counts. they're not as "true" as a converged result, but they're frequently interesting!

I also made one for epoch 5 but @haruu1367 released epoch 6 faster than I could finish the animation