Research map / Image generation / Editing and style transfer
Rectifying Latent Space for Generative Single-Image Reflection Removal
Mingjia Li, Jin Hu, Hainuo Wang, Qiming Hu, Jiarui Wang, Xiaojie Guo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train the VAE with the reflection-equivariance loss in addition to reconstruction loss (ablation) | GenSIRR (FLUX-based) · Real20 | PSNR: 25.79 -> 27.27 | Improved |
| add equivariance loss to VAE training (ablation) | GenSIRR re-VAE · Real20 | SSIM (SIRR): 0.842 -> 0.871 | Improved |
| finetune VAE with reconstruction loss only versus FLUX VAE (ablation) | VAE · Real20 | SSIM (SIRR): 0.841 -> 0.842 | No clear effect |
| add equivariance loss to VAE reconstruction (ablation) | VAE · synthetic reconstruction (non-mixed) | SSIM: 0.918 -> 0.919 | No clear effect |
| use learnable task-specific text embedding instead of fixed text prompt (ablation) | GenSIRR · Real20 | PSNR: 26.52 -> 27.27 | Improved |
| use learnable embedding instead of fixed prompt (ablation) | GenSIRR · Real20 | SSIM: 0.830 -> 0.871 | Improved |
| initialize prompt randomly or with RDNet prompt generator instead of fixed-text initialization (ablation) | GenSIRR · Real20 | convergence | Hurt |
| add DEBS with k=4 (ablation) | GenSIRR · Real20 | PSNR: 27.27 -> 27.58 | Improved |
| increase DEBS candidates from 4 to 16 (ablation) | GenSIRR · Real20 | PSNR: 27.58 -> 27.74 | Improved |
| increase DEBS candidates from 4 to 16 (ablation) | GenSIRR · Real20 | SSIM: 0.881 -> 0.882 | No clear effect |
| increase DEBS candidates from 1 to 16 (ablation) | GenSIRR · Inference cost (256x256) | inference time (ms): 2061.9 -> 3952.4 | Hurt |
| use GenSIRR instead of DAI and RDNet | reflection removal · OpenRR-val (human evaluation) | average success rate (%): 41.2% -> 96.6% | Improved |
Closest papers by meaning
- PRISM: Latent Composition Consistency for Single-Image Reflection Removal (ECCV 2026 · similarity 0.90)
- Fine-Grained Erasure in Text-to-Image Diffusion-based Foundation Models (CVPR 2025 · similarity 0.88)
- Masked Autoencoders Are Effective Tokenizers for Diffusion Models (ICML 2025 · similarity 0.87)
- NoiseCLR: A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Directions in Diffusion Models (CVPR 2024 · similarity 0.87)
- Progressive Text-to-Image Diffusion with Soft Latent Direction (AAAI 2024 · similarity 0.87)
- Unsupervised Region-Based Image Editing of Denoising Diffusion Models (AAAI 2025 · similarity 0.87)
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.
