XReflection is a neat toolbox tailored for single-image reflection removal(SIRR). We offer state-of-the-art SIRR solutions for training and inference, with a high-performance data pipeline, multi-GPU/TPU support, and more!

Equipped with XReflection, our team won the 1st place in the NTIRE 2025 Challenge on Single Image Reflection Removal in the Wild.

• [2026年06月02日] DIRS, built upon XReflection, is now open-sourced: https://github.com/mingcv/DIRS
• [2026年05月18日] Xiaomi achieved the runner-up award in the NTIRE 2026 Challenge using our XReflection. Read more: https://mp.weixin.qq.com/s/OwkKXKNkT_8pbc2LRtFQuA
• [2025年10月26日] XReflection is now production-ready and have been applied to multiple research projects inside our team. DSIT is available in the model zoo. More models are on the way!
• [2025年07月16日] DSRNet is available in the model zoo. More models are on the way!
• [2025年05月26日] Release a training/testing pipeline.

• All-in-one intergration for the state-of-the-art SIRR solutions. We aim to create an out-of-the-box experience for SIRR research.
• Multi-GPU/TPU support via PyTorchLightning.
• Pretrained model zoo.
• Fast data synthesis pipeline.

Please visit the documentation for more features and usage.

# Build from source
git clone https://github.com/hainuo-wang/XReflection.git
cd XReflection
# Install dependencies
pip install -r requirements.txt
python setup.py develop

python tools/train.py --config configs/train_config.yaml --test_only pretrained.ckpt

python tools/train.py --config configs/train_config.yaml

python tools/train.py --config configs/train_config.yaml --resume

python tools/train.py --config configs/train_config.yaml --resume your_checkpoint_path.ckpt

• 7,643 images from the Pascal VOC dataset, center-cropped as 224 x 224 slices to synthesize training pairs;
• 90 real-world training pairs provided by Zhang et al.;
• 200 real-world training pairs provided by IBCLN.

• 45 real-world testing images from CEILNet dataset;
• 20 real testing pairs provided by Zhang et al.;
• 20 real testing pairs provided by IBCLN;
• 454 real testing pairs from SIR^2 dataset, containing three subsets (i.e., Objects (200), Postcard (199), Wild (55)).

Download all in one from https://checkpoints.mingjia.li/sirs.zip

The performance of previous methods are improved with our new training pipeline. Access pretrained models for various SIRR algorithms. More are on the way.

This project is licensed under the Apache License 2.0. See the LICENSE file for details. The authors would express gratitude to the computational resource support from Google's TPU Research Cloud.