---
license: apache-2.0
language:
- en
library_name: pytorch
pipeline_tag: image-to-image
tags:
- reflection-removal
- reflection-separation
- computational-photography
- image-restoration
- pytorch
---

# DIRS Models and Data

This repository hosts the released checkpoints and TJReflection real-world data
for **DIRS: Principled Reflection Separation via Nonlinear Superposition and
Feature Interaction**.

- [Code](https://github.com/mingcv/DIRS)
- [Project Page](https://mingcv.github.io/DIRS-Page)
- [Paper](https://openreview.net/pdf?id=Shwtw8uV8l)

DIRS studies reflection separation under a nonlinear sRGB image formation model
and provides three released variants:

- **DIRS-YTMT**: CNN interaction through feature recycling.
- **DIRS-MuGI**: CNN interaction through mutual gating.
- **DIRS-PAIR**: Transformer-based dual-stream joint attention.

## Repository Contents

```text
checkpoints/
  dirs_ytmt_lors.ckpt
  dirs_mugi_lors.ckpt
  dirs_pair_lors.ckpt
  dirs_pair_lors_nature.ckpt
  rsr_supplement.ckpt
  polarized_dirs_pair_lors.ckpt
  pretrained/
    swin_large_o365.pth
datasets/
  TJReflection/
```

`datasets/TJReflection/` contains 175 real-world reflection images used by the
DIRS release.

## Checkpoints

Metrics are reported on the LORS test setting at 256 x 256 resolution. Runtime
is measured on a single NVIDIA RTX 3090.

| Model | Type | Params | FLOPs | Time | PSNR | SSIM | File |
|---|---:|---:|---:|---:|---:|---:|---|
| DIRS-YTMT | CNN, activation interaction | 32.42M | 102.91G | 31.35 ms | 24.94 | 0.902 | `checkpoints/dirs_ytmt_lors.ckpt` |
| DIRS-MuGI | CNN, mutual gating | 84.47M | 153.98G | 49.95 ms | 25.63 | 0.913 | `checkpoints/dirs_mugi_lors.ckpt` |
| DIRS-PAIR | Transformer, joint attention | 48.80M | 200.22G | 75.36 ms | 26.37 | 0.918 | `checkpoints/dirs_pair_lors.ckpt` |
| DIRS-PAIR + Nature | Transformer, joint attention | 48.80M | 200.22G | 75.36 ms | 26.95 | 0.926 | `checkpoints/dirs_pair_lors_nature.ckpt` |

Supplementary checkpoints:

| Task | File | Notes |
|---|---|---|
| Reflection scene reconstruction | `checkpoints/rsr_supplement.ckpt` | Supplementary model for reconstructing reflection scenes. |
| Polarized reflection separation | `checkpoints/polarized_dirs_pair_lors.ckpt` | DIRS-PAIR adapted to polarized inputs. |
| Swin prior | `checkpoints/pretrained/swin_large_o365.pth` | Pretrained prior used by DIRS-PAIR and polarized DIRS. |

## Usage

Clone this repository or download it from the Hugging Face UI, then place
`checkpoints/` and `datasets/` at the root of the
[DIRS code repository](https://github.com/mingcv/DIRS):

```bash
git clone https://github.com/mingcv/DIRS.git
cd DIRS
git clone https://huggingface.co/huqiming513/DIRS-Models DIRS-Models
cp -r DIRS-Models/checkpoints .
cp -r DIRS-Models/datasets .
```

Example evaluation command:

```bash
python -m xreflection.test \
  --config options/test_dirs_pair_lors.yml \
  --resume checkpoints/dirs_pair_lors.ckpt
```

Available evaluation configs in the code repository:

| Model | Config | Checkpoint |
|---|---|---|
| DIRS-YTMT | `options/test_dirs_ytmt_lors.yml` | `checkpoints/dirs_ytmt_lors.ckpt` |
| DIRS-MuGI | `options/test_dirs_mugi_lors.yml` | `checkpoints/dirs_mugi_lors.ckpt` |
| DIRS-PAIR | `options/test_dirs_pair_lors.yml` | `checkpoints/dirs_pair_lors.ckpt` |
| DIRS-PAIR + Nature | `options/test_dirs_pair_lors.yml` | `checkpoints/dirs_pair_lors_nature.ckpt` |

## Intended Use

These files are intended for academic research and reproducibility in image
reflection separation, reflection scene reconstruction, and polarized reflection
separation. The models are not designed as a safety-critical restoration system
and may fail on images outside the training and evaluation distribution,
including unusual glass materials, severe saturation, extreme low light, or
strong misalignment.

## Citation

If you use these checkpoints, data, or code, please cite:

```bibtex
@article{hu2026dirs,
  title={Principled Reflection Separation via Nonlinear Superposition and Feature Interaction},
  author={Hu, Qiming and Li, Mingjia and Li, Yuntong and Guo, Xiaojie},
  journal={arXiv preprint},
  year={2026}
}
```