
PCM acceleration LoRAs as individual .safetensors files.
These are already available in https://civitai.com/models/487073?modelVersionId=541627 , but as a single zip file, so can't be used on the Civitai generator.
Some quick guidelines:
-
all versions are very sensitive to CFG values: if you get too much contrast or burned-out images, try lowering the CFG. A higher CFG may also need a higher generation step count to fix inconsistencies, artifacts, etc;
-
smallCFG versions work pretty much like LCM: they essentially ignore the negative prompt, and need CFG to be very low (<= 2). Try CFG 1 first, since generation will be faster;
-
any n-step LoRA should work as long as you keep generation steps above n, but for much higher step counts a higher-step LoRA tends to generate better images;
-
for the scheduler, I had good results with DDIM, DEIS, DPM2, TCD, Euler and Euler A; in some cases, UniPC and even LCM work too. The DPM family doesn't seem to work very well;
For the on-site gen, I suggest starting with the 8-step normalCFG Lora, and CFG around 3 or 4, leaving the step count at 10. DDIM, Euler, DPM2 and Euler A schedulers seem to work best.
All credits go to https://civitai.com/user/G_U_N for this great resource.
Project page: https://g-u-n.github.io/projects/pcm/
Original repository: https://huggingface.co/wangfuyun/PCM_Weights
描述:
8-step normal CFG (use CFG 2-9 for generation, depending on the number of steps), and DDIM or Euler for sampling
训练词语:
名称: pcm_sd15_normalcfg_8step_converted.safetensors
大小 (KB): 131466
类型: Model
Pickle 扫描结果: Success
Pickle 扫描信息: No Pickle imports
病毒扫描结果: Success