SDXL Acceleration Model Original Model Download Link: https://huggingface.co/wangfuyun/PCM_Weights

4-Step Link: https://bizyair.cn/community/models/publicity/32037

Please use DDIM or Euler instead of LCM for sampling! When using DDIM, please set timestep_spacing="trailing", clip_sample = False, and set_alpha_to_one = False.

The name of each LoRA weight indicates how many inference steps it should be applied to.

The name of each LoRA weight indicates whether it supports normal CFG or small CFG.

NormalCFG means the model equipped with LoRA can generate images using a CFG value between 2–9. However, you should adjust the CFG value according to the number of steps used. When using fewer steps, a smaller CFG value should be used. For example, use CFG 2.5–3.5 for 4 steps, and CFG 3–6 for 8 steps. This is because the fewer the steps, the less chance the model has to fix issues caused by CFG. SmallCFG means the model equipped with LoRA can generate images using a CFG value between 1–2.

Note:

normalCFG LoRAs are more sensitive to prompts. Setting appropriate positive and negative prompts can improve quality. I found that running 4 steps with normalCFG doesn't work very well. I am currently trying to fix this issue.