Official LCM Sampler Support

[continue-revolution]

Description

This is official LCM sampler support. "Official" means that code has been reviewd by LCM author @luosiallen and LyCORIS author @KohakuBlueleaf

• The following issues should have been resolved
  • close [Feature Request]: Latent Consistency model #13730
  • close [Feature Request]: We need LCM merged model support  #13950
  • close [Feature Request]: new sampler "lcm" for lcm-loras #13952
• Code has been under beta test under my extension sd-webui-animatediff for nearly a month, and people are generally positive about it. If this PR is merged, LCM sampler will be removed in my extension when it updates to v2.0.0
• How to use:
  1. Download LCM LoRA to stable-diffusion-webui/models/Lora. Note that you should use different LoRA for SD15 and SDXL
  2. Use LCM sampler in this PR, or Euler a / Euler / DDIM / potentially others, need your experiment
  3. Use low CFG denoising strength (1-2 is recommended)
  4. Prompt engineering and generate your images / videos in 6-8 steps

Checklist:

• I have read contributing wiki page
• I have performed a self-review of my own code
• My code follows the style guidelines
• My code passes tests

[AUTOMATIC1111]

There should also be some way to convey to user that the lora is needed for this properly function. Maybe a warning message in the outputs (like the ones you get when you use a lora you don't have).

[KohakuBlueleaf]

There should also be some way to convey to user that the lora is needed for this properly function. Maybe a warning message in the outputs (like the ones you get when you use a lora you don't have).

I have tried LCM without LCM lora and it could actually generate some normal images
I think it is fine?

[light-and-ray]

There should also be some way to convey to user that the lora is needed for this properly function. Maybe a warning message in the outputs (like the ones you get when you use a lora you don't have).

It can be used with checkpoints with merged lcm lora, and it also works with sd_turbo

[continue-revolution]

It's challenging to verify whether or not user is using LCM LoRA. If we can verify in some way, then I tend to do things in this way

• if user is using this sampler, we check if user is using LCM LoRA
• if user is using LCM LoRA, we check if user is specifying a low CFG scale

One way to do so is to verify if the LoRA's name contains substring "lcm", but I'm not sure if this is a proper way.

[light-and-ray]

I think it was bad idea

[KohakuBlueleaf]

I think the best way is to add some kind of optional warning section
This will be good since we may add more features which only works if you enable/use some specific resources. Or some feature that is intended for specific hardware or whatever like that

[George0726]

Hi guys, it would be great to extend the LCM sampler to TCM sampler,
https://mhh0318.github.io/tcd/

[lalala-233]

Hi guys, it would be great to extend the LCM sampler to TCM sampler,
https://mhh0318.github.io/tcd/

That's good, but it's called TCD (

[bigmover]

There should also be some way to convey to user that the lora is needed for this properly function. Maybe a warning message in the outputs (like the ones you get when you use a lora you don't have).

It can be used with checkpoints with merged lcm lora, and it also works with sd_turbo

hi @

There should also be some way to convey to user that the lora is needed for this properly function. Maybe a warning message in the outputs (like the ones you get when you use a lora you don't have).

It can be used with checkpoints with merged lcm lora, and it also works with sd_turbo

Hi @continue-revolution This is great! LCM-lora is available for use with controlnet?

[continue-revolution]

@bigmover should be fine with any combination (no error on your terminal) but the performance will require your own experiments

[bigmover]

@bigmover should be fine with any combination (no error on your terminal) but the performance will require your own experiments

@continue-revolution Thanks very much for your prompt reply! I mean if controlnet use the same steps with SD base model, perhaps will have an influence on the quality on image. May I ask if there are any related suggestions or ideas?

[continue-revolution]

@bigmover Various experiments (ControlNet + AnimateDiff) show that you do not have to worry about this

[bigmover]

There should also be some way to convey to user that the lora is needed for this properly function. Maybe a warning message in the outputs (like the ones you get when you use a lora you don't have).

I have tried LCM without LCM lora and it could actually generate some normal images I think it is fine?

You mean that low CFG and low steps is ok on the quality of image? Besides whether newer TCD whill be supported?