Research map / Image generation / Image restoration
Flash-Split: 2D Reflection Removal with Flash Cues and Latent Diffusion Separation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Flash-Split separates transmission and reflection from a possibly misaligned flash/no-flash pair by running a dual-branch latent diffusion separation in VAE latent space followed by a cross-latent decoder that restores high-frequency detail.
Problem
Flash/no-flash reflection removal requires perfectly aligned image pairs; hand shake or shutter motion misaligns them and breaks the difference-based separation, and prior alignment via optical flow or homography is not robust.
Approach
Encode the flash/no-flash pair into a VAE latent pair and perform reflection separation there with a dual-branch diffusion UNet (cross-attention between transmission and reflection branches), then decode with a cross-latent decoder whose skip connections come from the unseparated input image to avoid hallucinated detail.
Key results
- On the Lei et al. dataset, the full method reaches 32.06 PSNR / 0.964 SSIM / 0.042 LPIPS for transmission, versus 28.60 / 0.956 / 0.071 for Lei et al. 2023.
- For reflection, the full method reaches 30.37 PSNR / 0.937 SSIM / 0.161 LPIPS, versus 26.29 / 0.889 / 0.383 for Lei et al. 2023.
- A model trained from scratch on synthetic data still outperforms all baselines on real scenes, with pretraining from SD v2 giving only a marginal further gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add inter-branch cross-attention (ablation) | Flash-Split with cross-latent decoder · Lei et al. flash/no-flash dataset | Reflection PSNR: 28.43 -> 30.37 | Helped |
| add inter-branch cross-attention (ablation) | Flash-Split with cross-latent decoder · Lei et al. flash/no-flash dataset | Transmission PSNR: 31.61 -> 32.06 | Helped |
| use cross-latent decoding instead of vanilla decoding (ablation) | Flash-Split with cross-attention · Lei et al. flash/no-flash dataset | Transmission PSNR: 30.23 -> 32.06 | Helped |
| use cross-latent decoding instead of vanilla decoding (ablation) | Flash-Split with cross-attention · Lei et al. flash/no-flash dataset | Reflection LPIPS: 0.187 -> 0.161 | Helped |
| use VAE latents rather than full-resolution RGB separation (ablation) | Flash-Split latent separation · Real images from [61] | reflection separation quality | Helped |
| downsample inputs by a factor of 8 rather than process full-resolution images (ablation) | Flash-Split latent separation · Real images from [61] | reflection separation quality | Helped |
| fine-tune from pretrained Stable Diffusion instead of training from scratch (ablation) | Flash-Split · Real-world testing | visual separation quality | Helped |
| add flash/no-flash cues instead of using a single composite image (ablation) | Flash-Split latent separation · Real-world scenes | reflection removal quality | Helped |
| use cross-latent decoding instead of vanilla VAE decoding (ablation) | Flash-Split · Qualitative reconstruction | detail faithfulness | Helped |
Surprising findings
- A model trained from scratch performs only marginally worse than the Stable Diffusion-pretrained model, so pretraining is not the dominant factor.
- A single-image model in latent space fails to remove reflections, showing flash cues remain essential even after latent compression.
Limitations
- Relies on flash/no-flash capture, so it still requires a flash-equipped camera and two shots.
- Evaluation on the quantitative benchmark uses a single dataset; most real-world results are qualitative.
- Tonemapped-image variant was shown qualitatively only.
Experimental setup
Datasets: Lei et al. 2023 flash/no-flash dataset (synthetic and real) data from Zhu/Yu et al. [61]
Benchmarks: Lei et al. 2023 flash/no-flash dataset
Metrics: PSNR SSIM LPIPS
Baselines: Lei et al. 2023 Dong et al. 2021 Hu et al. 2023 Zhu et al. 2024 Chugunov et al. 2024 SDN
Builds on: Stable Diffusion Marigold Lei et al. 2023 flash/no-flash reflection removal
Closest papers by meaning
- Robust Reflection Removal With Reflection-Free Flash-Only Cues (CVPR 2021 · similarity 0.91)
- Polarized Reflection Removal With Perfect Alignment in the Wild (CVPR 2020 · similarity 0.88)
- Reflection Separation from a Single Image via Joint Latent Diffusion (CVPR 2026 · similarity 0.88)
- PolarFree: Polarization-based Reflection-Free Imaging (CVPR 2025 · similarity 0.86)
- FIRM: Flexible Interactive Reflection ReMoval (AAAI 2025 · similarity 0.85)
- Removing Reflections from RAW Photos (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.