siliconflow / onediff

OneDiff: An out-of-the-box acceleration library for diffusion models.
https://github.com/siliconflow/onediff/wiki
Apache License 2.0
1.41k stars 85 forks source link

The lora_scale parameter in the load_and_fuse_lora() function doesn't work #924

Open liorplaytika opened 1 month ago

liorplaytika commented 1 month ago

I tried to use load_and_fuse_lora function with lora scale parameter but don't see an effect on the results - using lora scale equal to 0/0.5/1, creates the same results.

Example code:

from onediffx.lora import load_and_fuse_lora, unfuse_lora
import torch
from PIL import Image
from diffusers import AutoPipelineForImage2Image, StableDiffusionPipeline, AutoencoderKL, DiffusionPipeline

def run_inference(model_id, pipe, model_params):
      load_and_fuse_lora(pipe, model_id, lora_scale=model_params['lora_scale'])
      res = pipe(**model_params).images
      unfuse_lora(pipe)
      pipe.unload_lora_weights()
      torch.cuda.empty_cache()
      return res

def main():
      device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
      dtype = torch.float16 if device != "cpu" else torch.float32
      vae = AutoencoderKL.from_pretrained('sdxl-vae-fp16-fix', torch_dtype=dtype)
      pipe = AutoPipelineForImage2Image.from_pretrained('stable-diffusion-xl-base-1.0', variant='fp16',     vae=vae,                  torch_dtype=dtype).to(device)

    if torch.cuda.is_available():
        from onediffx import compile_pipe
        pipe = compile_pipe(pipe, ignores=("vae.encoder", "vae.decoder"))

        # warmup
        pipe(prompt="prompt",
                  negative_prompt="negative_prompt",
                  num_inference_steps=5,
                  width=128,
                  height=128,
                  num_images_per_prompt=1,
                  image=Image.open('ref_img.png'))

    model_params = {"prompt": "woman wearing pink dress'",
                    "negative_prompt": "Ugly, Dots",
                    "num_images_per_prompt": 1,
                    "num_inference_steps": 50,
                    "width": 1024,
                    "height": 1024,
                    "lora_scale": 0.5
                    }

    res = run_inference("nerijs/pixel-art-xl", pipe, model_params)

if __name__ == "__main__":
    main()

I added pipe.unload_lora_weights() after unfuse because without it, from run to run, the picture became more and more like gibberish

strint commented 1 month ago

@marigoold is seeing this

Amitg1 commented 3 weeks ago

Hi, any news here? got the same problem. Thanks!

strint commented 1 week ago

Fixed here: https://github.com/siliconflow/onediff/pull/926

@Amitg1 @liorplaytika