InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment (CVPR 2025 Highlight)

arXiv Paper | Project Page | Poster | SmPO-Diffusion Project Page

The repository provides the official implementation, experiment code, and model checkpoints used in our research paper.

---

📖 News & Updates

• [2025-03-24] 🎉 Preprint paper released on arXiv!
• [2025-03-24] ✅ Initial model checkpoints published
• [2025-06-04] 📊 Poster and project page
• [2025-06-29] 🚀 Training code release

---

🌟 Key Features

• Preservation of Pre-Trained Knowledge
• Targeted Performance Improvement
• Optimized Computational Efficiency
• Interpretability and Parametric Control
• Cross-Domain Broader Applicability

---

🔧 Quick Start

Installation

conda create -n ddiminpo python=3.10
conda activate ddiminpo
git clone https://github.com/JaydenLyh/InPO.git
cd InPO
pip install -r requirements.txt

---

Preparation of dataset and base models

InPO/
├── assets/                   
│   └── inpo.png            
├── checkpoints/            
│   ├── stable-diffusion-v1-5/          
│   ├── sdxl-vae-fp16-fix/            
│   └── stable-diffusion-xl-base-1.0/         
├── datasets/                 
│   └── pickapic_v2/   
├── train.py            
├── README.md              
├── LICENSE.txt            
└── requirements.txt       

---

Training for SDXL

export MODEL_NAME="checkpoints/stable-diffusion-xl-base-1.0"
export VAE="checkpoints/sdxl-vae-fp16-fix"
export DATASET_NAME="pickapic_v2"
PORT=$((20000 + RANDOM % 10000))

CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" accelerate launch --main_process_port $PORT --mixed_precision="fp16" --num_processes=8 train.py \
  --pretrained_model_name_or_path=$MODEL_NAME \
  --pretrained_vae_model_name_or_path=$VAE \
  --dataset_name=$DATASET_NAME \
  --train_batch_size=1 \
  --dataloader_num_workers=16 \
  --gradient_accumulation_steps=128 \
  --max_train_steps=400 \
  --lr_scheduler="constant_with_warmup" --lr_warmup_steps=200 \
  --learning_rate=1e-8 --scale_lr \
  --checkpointing_steps 50 \
  --beta_dpo 5000 \
  --sdxl  \
  --output_dir="ddiminpo-sdxl" 

---

Our Models

Model	Download Links
InPO-SD1.5	Hugging Face
InPO-SDXL	Hugging Face

Citation

@inproceedings{lu2025inpo,
  title={InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment},
  author={Lu, Yunhong and Wang, Qichao and Cao, Hengyuan and Wang, Xierui and Xu, Xiaoyin and Zhang, Min},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={28629--28639},
  year={2025}
}

Acknowledgments

The implementation of this project references the DiffusionDPO repository by Salesforce AI Research. We acknowledge and appreciate their open-source contribution.