WindowSeat Reflection Removal in ComfyUI

A ComfyUI custom node plugin for removing reflections from images using the WindowSeat model.

<p align="center"> <a href="https://hf.co/spaces/huawei-bayerlab/windowseat-reflection-removal-web"><img src="https://img.shields.io/badge/%F0%9F%A4%8D%20Project%20-Website-blue"></a> <a href="https://arxiv.org/abs/2512.05000"><img src="https://img.shields.io/badge/arXiv-PDF-b31b1b"></a> <a href="https://huggingface.co/huawei-bayerlab/windowseat-reflection-removal-v1-0"><img src="https://img.shields.io/badge/🤗%20Hugging%20Face%20-Model-yellow"></a> <a href="https://huggingface.co/spaces/toshas/windowseat-reflection-removal"><img src="https://img.shields.io/badge/🤗%20Hugging%20Face%20-Space-yellow"></a> </p>

Installation

Option 1: Manual Installation

cd ComfyUI/custom_nodes
git clone https://github.com/toshas/ComfyUI-WindowSeat.git
cd ComfyUI-WindowSeat
pip install -r requirements.txt

Option 2: ComfyUI Manager

Search for "WindowSeat" in ComfyUI Manager and install.

Note: If you get a security error when installing, you can temporarily change the security policy as described here, restart ComfyUI, install this plugin, and then restore your security policy.

Usage

1. Right-click on canvas → Add Node → WindowSeat
2. Add WindowSeat Model Loader node
3. Add WindowSeat Reflection Removal node
4. Connect them: Loader → Reflection Removal
5. Add an image input (Load Image node) and connect to Reflection Removal
6. Add a Preview Image or Save Image node to see results
7. Click "Queue Prompt" to run

Nodes

WindowSeat Model Loader

Loads the model from HuggingFace Hub (auto-downloads on first use). Caches the model for reuse.

WindowSeat Reflection Removal

Required Inputs:

• model - Connect from WindowSeat Model Loader
• image - Input image (from Load Image node)

Optional Inputs:

| Parameter | Default | Description | |-----------|---------|-------------| | use_short_edge_tile | True | Use short edge for tile size | | tiling_size | 768 | Base tile size (512-1536) | | max_tiles_w | 4 | Max horizontal tiles (1-8) | | max_tiles_h | 4 | Max vertical tiles (1-8) | | min_overlap | 64 | Tile overlap pixels (16-256) | | tile_batch_size | 2 | Tiles per batch (1-4) |

Requirements

• Python 3.10+
• PyTorch 2.0+
• CUDA-capable GPU
• 24GB VRAM

Development

pip install -r requirements.txt
pre-commit install  # Required: install git hooks

Run tests:

./run_tests.sh           # Unit tests (no GPU)
./run_tests.sh --gpu     # All tests including GPU integration

Troubleshooting

Out of memory:

• Reduce tile_batch_size to 1
• Use smaller tiling_size (e.g., 512)
• Process smaller images

Model not loading:

• Check HuggingFace Hub access
• Ensure you're logged in: huggingface-cli login

Citation

Please cite our paper:

@misc{zakarin2025reflectionremovalefficientadaptation,
  title        = {Reflection Removal through Efficient Adaptation of Diffusion Transformers},
  author       = {Daniyar Zakarin and Thiemo Wandel and Anton Obukhov and Dengxin Dai},
  year         = {2025},
  eprint       = {2512.05000},
  archivePrefix= {arXiv},
  primaryClass = {cs.CV},
  url          = {https://arxiv.org/abs/2512.05000},
}

License

The code and models of this work are licensed under the Apache License, Version 2.0. By downloading and using the code and model you agree to the terms in LICENSE.