README.md

---
base_model:
- stabilityai/stable-diffusion-2
library_name: diffusers
pipeline_tag: image-to-image
tags:
- reflection-removal
- reflection-separation
- image-restoration
- diffusion
- stable-diffusion
- cvpr2026
---

# Reflection Separation from a Single Image via Joint Latent Diffusion

This repository contains pre-trained checkpoints for the CVPR 2026 paper [Reflection Separation from a Single Image via Joint Latent Diffusion](https://huggingface.co/papers/2606.04107).

Given a single photo taken through glass, the model jointly generates the **transmission** (reflection-free) and **reflection** layers using a fine-tuned Stable Diffusion 2 architecture.

- **Authors:** Zheng-Hui Huang, Zhixiang Wang, Yu-Lun Liu, and Yung-Yu Chuang
- 🌐 **Project page:** [https://brian90709.github.io/diff-reflection-separation/](https://brian90709.github.io/diff-reflection-separation/)
- 💻 **Code:** [https://github.com/Brian90709/diff-reflection-separation-code](https://github.com/Brian90709/diff-reflection-separation-code)
- 📄 **Paper:** [arXiv:2606.04107](https://arxiv.org/abs/2606.04107)

## Method Overview
Single-image reflection separation is challenging under conditions like glare or weak reflections. This method leverages generative diffusion priors to simultaneously generate transmission and reflection layers through a unified diffusion model, incorporating a novel cross-layer self-attention mechanism for better feature disentanglement and a disjoint sampling strategy to reduce interference between layers.

## Files

| File | Size | Description |
| --- | --- | --- |
| `iter_016000/unet/diffusion_pytorch_model.bin` | ~3.5 GB | Trained layer-separation UNet. |
| `fuse_blocks.bin` | ~264 MB | CFW refiner for the VAE decoder. |
| `lrm/iter_008000/aux_net.bin` | ~1.3 MB | Latent composition module (LRM), used by `--optimization`. |

## Usage

### Installation
Clone the [official repository](https://github.com/Brian90709/diff-reflection-separation-code) and set up the environment:

```bash
conda create -y -n diffrs python=3.10 && conda activate diffrs
pip install torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt
```

### Inference
Download the weights into `./checkpoints`:

```bash
huggingface-cli download Brian9999/diff-reflection-separation --repo-type model --local-dir ./checkpoints
```

Run the inference script on a directory of images:

```bash
python infer_layersep.py --input_dir ./samples --save_to_dir ./outputs
```
Each input yields three files: `*_transmission.png` (reflection-free result), `*_reflection.png`, and `*_ori_transmission.png` (transmission before CFW refinement).

## Citation

```bibtex
@inproceedings{huang2026reflection,
  title     = {Reflection Separation from a Single Image via Joint Latent Diffusion},
  author    = {Huang, Zheng-Hui and Wang, Zhixiang and Liu, Yu-Lun and Chuang, Yung-Yu},
  booktitle = {CVPR},
  year      = {2026}
}
```