Accelerating Steady Diffusion

Do you find your Stable Diffusion too slow? Many options to speed up Stable Diffusion is now available. In this article, you will learn about the following ways to speed up Stable Diffusion.

  • cross-attention optimization
  • Token merging
  • Negative guidance minimum sigma

Software

We will use AUTOMATIC1111 Stable Diffusion GUI to create images. You can use this GUI on Google ColabWindows, or Mac.

Cross-attention optimization options

All optimization options focus on making the cross-attention calculation faster and using less memory. Below are all the options available to you in AUTOMATIC1111.

How to set cross-attention optimization

In AUTOMATIC1111 Web-UI, navigate to the Settings page. Select Optimization on the left panel. In the Cross attention optimization dropdown menu, select an optimization option. The initial selection is Automatic. Click Apply Settings.

Which one should you pick? See the explanation below.

Doggettx

In the early days of Stable Diffusion (which feels like a long time ago), the GitHub user Doggettx made a few performance improvements to the cross-attention operations over the original implementation.

It was a good speed-up then, but people mostly move on to other speed-up options listed below.

xFormers

The attention operation is at the heart of the Stable Diffusion algorithm but is slow.

xFormers is a transformer library developed by the Meta AI team. It speeds up and reduces memory usage of the attention operation by implementing memory-efficient attention and Flash Attention.

Memory-efficient attention computes attention operation using less memory with a clever rearrangement of computing steps. Flash Attention computes the attention operation one small patch at a time. The end result is less memory usage and faster operation.

xFormers is the go-to library before PyTorch implemented a native support. If you have a newer Nvidia video card, you should use SDP instead.

AUTOMATIC1111 command line argument: --xformers

Scaled-Dot-Product (sdp) attention

Scaled dot product attention is Pytorch’s native implementation of memory-efficient attention and Flash Attention. In other words, it is an alternative implementation of the xFormers option.

A drawback of this optimization is that the resulting images can be non-deterministic (a problem in older xFormer versions). You may be unable to reproduce the same image using the same generation parameters.

This is a new function that requires Pytorch 2 or above.

AUTOMATIC1111 command line argument: --opt-sdp-attention

sdp-no-mem

sdp-no-mem is the scaled-dot-product attention without memory-efficient attention.

Unlike SDP attention, the resulting images are deterministic. The same generation parameters produce exactly the same image.

This is a new function that requires Pytorch 2 or above.

AUTOMATIC1111 command line argument: --opt-sdp-no-mem-attention

sub-quadratic attention

Sub-quadratic (sub-quad) attention is yet another implementation of memory-efficient attention, which is part of xFormer and SDP. You can try this option if you cannot use xFormers or SDP.

AUTOMATIC1111 command line argument: --opt-sub-quad-attention

Split-attention v1

Split-attention v1 is an earlier implementation of memory-efficient attention.

You should use xFormers or SDP when turning this on. In AUTOMATIC1111, it is on by default.

AUTOMATIC1111 command line argument: --opt-split-attention-v1

To turn off: --disable-opt-split-attention

Invoke AI

Cross-attention optimization as used in the Invoke AI code base. This option is useful for MacOS users where Nvidia GPU is not available.

Memory management options

Med Vram

Use the command line argument --medvram to conserve GPU memory by only allowing some parts of the model in the VRAM. It slows down the generation a little bit.

Low VRAM

Use the command line argument --lowvram to conserve even more GPU memory by allowing only necessary parts in the VRAM. It drastically slows down the generation.

SDXL settings

Try adding the command line argument –medvram-sdxl or –lowvram experience significant slowdown or cannot run Stable Diffusion XL models. Here are the recommendations.

  • Nvidia 12GB+ VRAM: --opt-sdp-attention
  • Nvidia 8GB VRAM: --opt-sdp-attention --medvram-sdxl
  • Nvidia 4GB VRAM: --opt-sdp-attention--lowvram
  • AMD 4GB VRAM: --lowvram --opt-sub-quad-attention

Token merging

Token merging (ToMe) is a new technique to speed up Stable Diffusion by reducing the number of tokens (in the prompt and negative prompt) that need to be processed. It recognizes that many tokens are redundant and can be combined without much consequence.

The amount of token merging is controlled by the percentage of token merged.

Below are a few samples with 0% to 50% token merging.

Token Merging produces similar images (Image: Original paper)

A drawback of token merging is it changes the images. You should not turn it on if you want others to reproduce the exact images.

Using token merging in AUTOMATIC1111 WebUI

AUTOMATIC1111 has native support for token merging. You don’t need to install an extension to use it.

To use token merging, navigate to the Settings page. Go to the Optimizations section. Set the token merging ratio. 0.2 means merging 20% of the tokens, for example. Click Apply Settings.

Negative guidance minimum sigma

The Negative guidance minimum sigma option turns off the negative prompt under certain conditions that are believed to be inconsequential. Emprical testing suggests that negative conditioning can be turned off in some steps without affecting the image.

Testing minimum sigma using the prompt in the next section.

Although changes are small, there are still noticeable minor changes. So don’t use this setting if you want to document the parameters and reproduce exactly the same image later.

Using negative guidance minimum sigma

To use negative guidance minimum sigma, navigate to the Settings page. Go to the Optimizations section. Set the value for the negative guidance minimum sigma. Click Apply Settings.

Speed and memory benchmark

Test setup

Below are the prompt and the negative prompt used in the benchmark test. The exact prompts are not critical to the speed, but note that they are within the token limit (75) so that additional token batches are not invoked.

(close-up editorial photo of 20 yo woman, ginger hair, slim American sweetheart), (freckles:0.8), (lips parted), realistic green eyes, POV, film grain, 25mm, f/1.2, dof, bokeh, beautiful symmetrical face, perfect sparkling eyes, well defined pupils, high contrast eyes, ultra detailed skin

(semi-realistic, cgi, 3d, render, sketch, cartoon, drawing, anime:1.4), text, close up, cropped, out of frame, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation

Model: Chillout Mix

Sampling method: Euler

Size: 512×512

Sampling steps: 20

Batch count: 2

Batch size: 8

CFG Scale: 7

Seed: 100

The following results are from running on an RTX 4090 GPU card and CUDA 11.8.

Benchmark for cross-attention optimization

Option Time (sec) Peak VRAM
None 35.1 23.1 GB
doggettx 14.4 14.9 GB
xFormers 12.8 3.5 GB
sdp-no-mem 10.6 3.5 GB
sdp 10.5 3.5 GB
sub-quadratic 18.8 5.9 GB
v1 21.5 3.6 GB
Invoke AI 14.4 14.9 GB
Cross-attention optimization

Scaled dot product optimization performs the best, followed by xFormers. Both are good options for speed up and reduced memory usage.

Benchmark for token merging

Token merge % Time (sec) Peak VRAM
0 10.6 3.5 GB
0.3 10.3 3.5 GB
0.5 10.0 3.5 GB
0.7 9.9 3.5 GB
Token merging.

Token merging improves generation speed, though I don’t find it to be significant.

It does have the potential to alter images substantially. So only use it if you don’t care about reproducibility.

Benchmark for negative guidance minimum sigma

Negative guidance Time (sec) Memory (with reserved)
0 10.6 3.5 GB
0.3 10.5 3.5 GB
0.5 10.2 3.5 GB
0.7 10.1 3.5 GB
Negative guidance minimum sigma.

Negative guidance minimum sigma resulted in speed up similar to token merging, but changes the images much less.

Which speed-up option should you use?

If you care about reproducing your images, use xFormers or SDP-no-mem. Don’t use token merging or negative guidance minimum sigma.

If you don’t care about reproducibility, feel free to choose between xFormers, SDP-no-mem and SDP. Use token merging and negative guidance minimum sigma for additional speed up.

If you are on MacOS, Invoke IA is your best bet.

Useful resources

Optimizations · AUTOMATIC1111/stable-diffusion-webui Wiki – Documentation of optimization options from the official AUTOMATIC1111 wiki.

[2303.17604] Token Merging for Fast Stable Diffusion – Research article on token merging.

aizmin: