huawei-bayerlab/windowseat-reflection-removal-v1-0

News

2026-09-13: Mirrored on ModelScope:
2026-06-04: Presented at the NTIRE workshop at CVPR 2026.
2025-12-05: Initial release: inference code, the released checkpoint, and the demo.

WindowSeat Model Card

This is a model card for the windowseat-reflection-removal-v1-0 model for single-image reflection removal. The model is derived from Qwen/Qwen-Image-Edit-2509 using LoRA adaptation as described in our paper titled "Reflection Removal through Efficient Adaptation of Diffusion Transformers" by Daniyar Zakarin, Thiemo Wandel, Anton Obukhov, Dengxin Dai.

See the Quick Start section of the paper's code repository for instructions on how to set up the environment and process photos with this model.

• Model Name: windowseat-reflection-removal-v1-0
• Task: Single Image Reflection Removal
• Base Model: Qwen/Qwen-Image-Edit-2509
• Model Type: End-to-end single-step latent diffusion reflection removal from a single image.
• Resources for more information: Project Website, Paper, Code.
• Framework: PyTorch, Transformers, PEFT
• Language: English.
• License: Apache-2.0
• Developed by: HUAWEI Bayer Lab
• Cite as:

@InProceedings{Zakarin_2026_CVPR,
  author    = {Zakarin, Daniyar and Wandel, Thiemo and Obukhov, Anton and Dai, Dengxin},
  title     = {Reflection Removal through Efficient Adaptation of Diffusion Transformers},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
  year      = {2026},
  pages     = {2776--2785}
}