[LCM模型]LCM-LoRA-SDXL

LCM-LoRA - Acceleration Module!

Tested with ComfyUI, although I hear it's working with Auto1111 now!
Step 1) Download LoRA
Step 2) Add LoRA alongside any SDXL Model (or 1.5 model, if using the SD 1.5 version)
Step 3) Set CFG to ~1.5 and Steps to 3
Step 4) Generate images in ~<1 second (instantaneously on a 4090)
Basic LCM Comfy workflow attached as "Training Images" in zip format.

Read more about the concept here：
https://github.com/luosiallen/latent-consistency-model/blob/main/LCM-LoRA%20Technical%20Report/LCM-LoRA-Technical-Report.pdf
Original HF link - SDXL：
https://huggingface.co/latent-consistency/lcm-lora-sdxl/tree/main
Original HF link - SD 1.5：
https://huggingface.co/latent-consistency/lcm-lora-sdv1-5

测试提示词和参数：
prompt:
a woman in a hat holding a cell phone, fantasy art, using a magical tablet, sci-fi fantasy wallpaper, hearthstone style art, psytrance artwork, witchy, golden turquoise steampunk, a witch, amazing wallpaper, [[fantasy]], colorful spells, mages. steampunk engine style

Negative prompt:
text, watermark, low quality, worst quality, poor composition, poor quality, effortless, childish, infantile, young (worst quality, grayscale), watermark, username, signature, text, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, jpeg artifacts, bad legs, extra fingers, mutated hands, poorly drawn hands, poor proportions, extra limbs, disfigured, poor anatomy, rough proportions, deformed limbs, lack of arms, lack of legs, extra arms, extra legs, mutated arms, fused fingers, too many fingers, long neck

参数
Steps: 30, Size: 832x1216, Seed: 612380400, Sampler: DPM++ 2M, CFG scale: 4, Clip skip: 2