CLEAR: Color-Texture Latent Alignment for Single Image Reflection Removal

WACV 2027 · Code · Paper / OpenReview

Tu Vo, Sheir A. Zaheer, Chan Y. Park — KC Machine Learning Lab

CLEAR is a deterministic, single-timestep latent framework for single image reflection removal. A physics-guided dual-branch separator predicts the transmission, reflection and residue latents in one forward pass, and a cross-latent decoder with the Hybrid Feature Alignment Module (HFAM) re-aligns the input skip features to the clean transmission before decoding. There is no noise schedule and no iterative sampling.

Files

config.yaml                              Stage-1 training config of the released run
checkpoint/
├── diffusion/
│   ├── unet1/                           transmission branch  (config.json + diffusion_pytorch_model.bin)
│   └── unet2/                           reflection branch    (config.json + diffusion_pytorch_model.bin)
└── cld/
    └── cld_clear.pth                    Stage-2 HFAM cross-latent decoder

cld_clear.pth is the final Stage-2 checkpoint (212 training hours). It stores the state_dict plus the metadata refiner_type: hfam, total_hours: 212 and name: 0.

Quick start

git clone https://github.com/kc-ml2/clear.git
cd clear
pip install -r requirements.txt
python scripts/download_models.py --save_dir ./models            # SDv2 + Marigold backbones
CLEAR_HF_REPO_ID=ML2/CLEAR python scripts/download_models.py --weights --save_dir ./models

Run on a folder of images (the folder must contain a blended/ subfolder):

bash scripts/inference.sh \
    models/clear/checkpoint/diffusion \
    models/clear/checkpoint/cld/cld_clear.pth \
    <DATA_DIR> outputs/inference

This writes <name>_clear.png (the reflection-free output), <name>_r.png (predicted reflection), <name>_g1.png (residue) and the input / target copies.

Results

Average over six real-world benchmarks (Nature, SIR2, Real20, DRR-C, DRR-S, OpenRR-1k — 1,040 images), as reported in the paper:

Method	PSNR ↑	SSIM ↑	LPIPS ↓	MANIQA ↑	QUALICLIP ↑
DAI	25.98	0.869	0.159	0.460	0.696
WindowSeat	25.48	0.810	0.208	0.465	0.674
CLEAR (Ours)	26.59	0.906	0.106	0.504	0.731

See the paper for the per-dataset tables and ablations.

Notes

• config.yaml is the original Stage-1 training config and still refers to the authors' workspace (base_config: ../config/default.yaml). For inference use configs/stage1_clear.yaml from the code repository.
• Stage-1 is initialised from stabilityai/stable-diffusion-2 and prs-eth/marigold-depth-v1-0; both are downloaded by scripts/download_models.py.
• The VAE and text encoder are not duplicated here — they come from the SDv2 backbone.

License

The code is released under Apache-2.0. These weights are fine-tuned from Stable Diffusion 2 and are distributed under the CreativeML Open RAIL++-M License; the use restrictions of that license apply.

Citation

@inproceedings{vo2027clear,
  title     = {CLEAR: Color-Texture Latent Alignment for Single Image Reflection Removal},
  author    = {Vo, Tu and Zaheer, Sheir A. and Park, Chan Y.},
  booktitle = {The IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year      = {2027}
}