Use tomesd aka. Token Merging to speed up generation

Related: official PR of A1111 WebUI

• 2023/08/14: This project is no longer needed since token merging is already included in latest A1111 WebUI.
• 2023/05/14: Experimental support activating during hires fix! Requires my modified version to A1111 WebUI, which you can pull from [CLICK HERE]. Otherwise I can't detect when the logic enter & exit hires pass.
• 2023/05/13: Attach ToMe related settings into image generation infos, prompt paste parsing in the planning.

Open a terminal, activate your webui environment (typically, execute the venv/Scripts/activate from webui installation path)

Do anything necessary needed if you have a fancy environment settings like me.

And then follow the instruction of tomesd Installation

After successfully installed tomesd, installed this extension like other normal webui extensions (install via URL from webui or clone this repo to extensions folder manually)

Enable it by checking Enable ToMe optimization below generation UI, where many other extensions are (eg. ControlNet)

If you installed tomesd correctly, it should be enabled by default.

In Settings tab, you'll find a section called ToMe Settings, there are 3 major options and other advanced ones:

Major settings:

• ToMe Merging Ratio: higher the faster, at the cost of (sort of) generation quality, recommend <=0.6 according to tomesd document
• ToMe Min x/y: only apply ToMe when image size reach these values, since ToMe have few benefit when image size is small (when collab with xformers/SDP)

Advance settings:

• Use random perturbations: had been caused some artifacts in some sampling methods, fixed in tomsd v0.1.3
• and other stuff, leave them default if U don't know what you are doing

• Cannot apply ToMe only to hires fix pass since A1111 WebUI didn't expose the hires logic (it's enclosed in StableDiffusionProcessingTxt2Img's sample method). You can do a normal text2image first and then send to image2imamge for scaling up instead.
• It will change the image content. If your prompt is simple (like 1girl), it changes a lot. So I can't take hires size and batch size into account, or you will get a complete different image simply because you toggle hires fix or change batch size. The state of ToMe will be written into image generation info (how to load it when you paste is under examination)
• Feel free to turn on/off ToMe if you worry it affects your image quality. More over, you can pin tome_merging_ratio to your UI quick settings for fast tuning. Every change will apply the next time you click Generate button.

Tested on RTX 4090 24GB, Python 3.10.9, PyTorch 2.0, CUDA 11.8, CuDNN 8.8.1.3, xformers 0.0.17, with --skip-version-check --xformers --opt-sdp-attention --no-half-vae enabled, step 30, batch count 5, same seed, use best result

PS: ratio 0.9 is just for showcasing the performance, it's not the way it should be configured (according to tomesd document, ratio is limited by 1-(1/(s_x * s_y)), which is 0.75 by default (s_x and s_y default to 2)), and the genereation quality is not taken into account)

Works with big image size and big batch size, you will need total pixel of 4*512*512 = 1024*1024 or more to see a difference

The higher the total pixel there are, the more performance boost you'll get, on 2048*2048, it could be over +100% in extreme settings

In more common scenarios (512*512 with hires fix 2x), you can get around +30% speedup during the hires part, which is a definitely time saver

PS: after I did above test, I updated xformers from 0.0.17 to 0.0.18, it seems that there is overall ~10% speedboost, so the exact generation speed value may vary if I redo the test.