Birchlabsさんのイラストまとめ

I was actually trying to render a different video, but neglected to fix the seed. so I came back to 60 random images, a transition between Reimu and Marisa. the random Marisa came out so well, I used her random seed as the basis for the video you see above.
#stablediffusion

made #stablediffusion a bit more stable.
now we know the seed for *every* sample generated, rather than just the first one.
saved into the filename.
works with --n_iter and --n_samples.
https://t.co/pr3DYGBvHF

I like the clock tower take on Starry Night
#stablediffusion

@loretoparisi @soumithchintala MPS works fine
can generate an image in 9.3 secs on M1 Max

@iScienceLuvr @Arman_point0 is this a fair fight? the samplers from the Karras paper (k_heun, k_euler) were published in the context of a novel noise schedule (get_sigmas_karras).
Heun performs well at as low as 7 steps with the right noise schedule.

#stablediffusion #7stepchallenge
show me your best generation with 7 sampler steps

fixed k-diffusion's k_lms sampler on M1 GPU!
before/after pics.
#stablediffusion
https://t.co/uorxaRAIZs

fixed k-diffusion Heun sampler on M1 GPU.
with Karras noise schedule, we can get this quality in 8 steps of sampling. that's very few steps.
before/after pics.
#stablediffusion
https://t.co/4QVWOyeNE5

k-diffusion sampling now running fully* on-GPU on M1
#stablediffusion
*okay fine I fell back to CPU to sort a wee 2x2 matrix lemme have this
https://t.co/GouHsCLStK

yes you can run GFPGAN on it to generate a more realistic face