"the moon overrun by flowers, tempera" A grid of 256 samples, made in 4 minutes 34 seconds (20 steps PLMS).

Feb 9, 2022 · 5:33 AM UTC

6
9
1
97
The 4 minutes 34 seconds was on *one GPU* btw (an A100). The sampling takes four model outputs per timestep for the first three timesteps and one model output per timestep for all subsequent timesteps, so it did computation equivalent to 29 DDIM steps.
1
12
It's using the fast sampling method from "Pseudo Numerical Methods for Diffusion Models on Manifolds" (openreview.net/forum?id=PlKW…), which consist of diffusion-specialized versions of higher order ODE solvers.
1
2
20
I recommend this paper, their method seems to work quite well for classifier-free guidance! It probably works well for other conditional diffusion model types like upscalers too.
13
Replying to @RiversHaveWings
What model?
1
1
cc12m_1_cfg (github.com/crowsonkb/v-diffu…, colab.research.google.com/dr…), just with a work in progress fast sampling method that produces outputs of similar quality to the current sampling method.

Text to Image (CC12M Diffusion).ipynb

Colaboratory notebook

colab.research.google.com
1
4
Replying to @RiversHaveWings
@viccpoes should I bookmark this?
Replying to @RiversHaveWings
You are an incredible person, bringing so much knowledge to the community and always sharing what you learn. I'm having an amazing time just experimenting the things you share. I left the default prompt on the colab, and I just love the results. 32 images is not possible, but 12
2
Replying to @RiversHaveWings
2022 is gonna be even crazier for AI art than 2021
7
Replying to @RiversHaveWings
🤯
1