LucianoCirino/efficiency-nodes-comfyui

Document support for LCM-lora

halr9000 opened this issue · 0 comments

Oh cool, LCM-lora works!

Here is an sd1.5 workflow with a 5 step generation @ 1.5 cfg, and it's as useless as you'd expect:
eff-no-lcm

But here is the same workflow where I've added in lcm-lora, and the output is decent! (Not saying that I'm perfectly happy with the technique yet, but you can't argue with its SPEED.)
eff-with-lcm-lora

Below are some why bother articles, as well as a workflow example from theally @ civitai, which I've tested and works as expected. Given the way it only patches the model, I was not sure if this would work with a lora stack, but it seems fine. I also have no idea what the model sampling discrete node is for.

Please check my work and maybe tidy things up in the docs and etc so people know that LCM works with your nodes?

https://huggingface.co/blog/lcm_lora

https://medium.com/@ayush-thakur02/what-is-lcm-lora-a-stable-diffusion-acceleration-module-b8959601fc4f

https://civitai.com/models/195519?modelVersionId=225222

LCM Workflow.json
t2i-sdxl-lcm