hnmr293 / ssasd

SSaSD - Shrinked Self-attention for Stable Diffusion

SSaSD - Shrinked Self-attention for Stable Diffusion

Installation

for diffusers

$ git clone https://github.com/hnmr293/ssasd.git

from diffusers import StableDiffusionPipeline
import ssasd

pipe = StableDiffusionPipeline.from_single_file(...)
ssasd.apply_patch(pipe)

for A111 WebUI

$ cd "your extension directory"
$ git clone https://github.com/hnmr293/ssasd.git

What is this?

This is the script to reduce the size of the latent input to the self-attention layer. Then computation time of the self-attention layer will be reduced and generation speed will be improved.

Examples

• horizontal - tomesd off/on
• vertical - ssasd(this) off/on

• horizontal - ssasd off/on

[setting]
  WebUI
  DPM++ 3M SDE / 30 steps / 640x1024 / cfg scale 3.0
  ssasd: max_downsample=2, scale=2, end_step=0.5(15)

[performance]
  off: 7.4s
  on: 6.3s

• horizontal - ssasd off/on

[setting]
  WebUI
  7th_anime_v3_A
  DPM++ 2M Karras / 20 steps / 640x1024 / cfg scale 7.0
  ssasd: max_downsample=1, scale=2, end_step=0.5(10)

[performance]
  off: 7.1s
  on: 6.1s

Performance

[setting]
  diffusers
  DPM++ 3M SDE / 30 steps / 640x1024 / cfg scale 3.0
  ssasd: max_downsample=2, scale=2, end_step=0.5(15)

setting	n	avg (s)	max (s)	min (s)	std. dev. (s)
normal	10	27.10	27.64	26.54	0.41
tomesd	10	20.01	20.41	19.36	0.37
ssasd (this)	10	22.07	22.66	21.49	0.45
tomesd + ssasd	10	19.44	20.13	18.93	0.48

Yes, tomesd is nice ^^;