OSDiff

This repository provides a PyTorch implementation of the paper "One-Step Diffusion for Perceptual Image Compression".

---

🧩 Introduction

We propose OSDiff, a diffusion-based perceptual image compression method that performs one-step diffusion, drastically reducing inference latency and computational cost. To further boost perceptual quality, we introduce a feature-space discriminator operating on intermediate UNet representations, allowing the model to better align reconstructed features with those of the original images.

---

📦 Requirements

conda create -n osdiff python=3.8
conda activate osdiff
pip install -r requirements.txt

---

⚙️ Inference

Download Pre-trained Model

Download the pre-trained checkpoint from pretrain_weight.

---

Run the following command

python inference.py \
    --ckpt_lc ./weight/lambda_1.ckpt \
    --input path to input images \
    --output path to output files

---