yichaodu/DiffusionDPO-bias-hps-2.1

Aligned Diffusion Model via DPO

Diffusion Model Aligned with thef following reward model and DPO algorithm

close-sourced vlm: claude3-opus  gemini-1.5  gpt-4o  gpt-4v
open-sourced vlm: internvl-1.5
score model: hps-2.1

How to Use

You can load the model and perform inference as follows:

from diffusers import StableDiffusionPipeline, UNet2DConditionModel

pretrained_model_name = "runwayml/stable-diffusion-v1-5"

dpo_unet = UNet2DConditionModel.from_pretrained(
        "path/to/checkpoint",
        subfolder='unet',
        torch_dtype=torch.float16
    ).to('cuda')

pipeline = StableDiffusionPipeline.from_pretrained(pretrained_model_name, torch_dtype=torch.float16)
pipeline = pipeline.to('cuda')
pipeline.safety_checker = None
pipeline.unet = dpo_unet

generator = torch.Generator(device='cuda')
generator = generator.manual_seed(1)

prompt = "a pink flower"

image = pipeline(prompt=prompt, generator=generator, guidance_scale=gs).images[0]

Citation

@misc{mjbench2024mjbench,
  title={MJ-BENCH: Is Your Multimodal Reward Model Really a Good Judge?},
  author={Chen*, Zhaorun and Du*, Yichao and Wen, Zichen and Zhou, Yiyang and Cui, Chenhang and Weng, Zhenzhen and Tu, Haoqin and Wang, Chaoqi and Tong, Zhengwei and HUANG, Leria and Chen, Canyu and Ye Qinghao and Zhu, Zhihong and Zhang, Yuqing and Zhou, Jiawei and Zhao, Zhuokai and Rafailov, Rafael and Finn, Chelsea and Yao, Huaxiu},
  year={2024}
}