Super Fast Image Generation in stable diffusion using LCM LoRA

#lcm #stablediffusion #aiimages #a1111 #aiupdates #LoRA #art

LCM Latent Consistency Models

This new LoRA Model requires only 4 to 8 sampling steps instead of the usual 25 to 50 steps, to produce complete detailed images.
LCM LoRA doesn’t require a special plugin or installation, it can be used like a normal LoRA model, The Latent Consistency Models or LCM are trained in a different way than normal LoRA models, which allows to infer images faster.
for model downloads and more reading please check the links bellow.

Hugging face page
https://huggingface.co/latent-consistency

SD 1.5 LoRA model
https://huggingface.co/latent-consistency/lcm-lora-sdv1-5/tree/main

SDXL LoRA model
https://huggingface.co/latent-consistency/lcm-lora-sdxl/tree/main

Dreamshaper v7 LCM
https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7/tree/main

More reading
https://huggingface.co/blog/lcm_lora

Note: for Hires fix, I found it to work best at higher denoising strength such as 0.7, producing very reliable results with hires steps of 5 for example when using LCM LoRA.
Note: for img2img, you should double check the actual number of steps being used, for instance, if you didnt enable an option in the User Interface settings that allows using same number of steps for img2img then 15 steps would only render 5 actual steps at 0.35 denoising strength (0.35 * 15) so you should take this into account to make sure that img2img will process images at at least 4 steps in the console window.