Research map / Image generation / Image restoration
SLER-IR: Spherical Layer-wise Expert Routing for All-in-One Image Restoration
Shurui Peng, Xin Lin, Shi Luo, Jincen Ou, Dizhe Zhang, Lu Qi, Truong Nguyen, Chao Ren
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 |
|---|---|---|---|
| Add hyperspherical contrastive loss with spherical routing instead of classifier-based branch selection with L1 loss (ablation) | SLER-IR · five-task all-in-one average | PSNR: 28.30 -> 31.73 | Improved |
| Add auxiliary classifier loss to L1 (ablation) | SLER-IR with classifier branch selection · five-task all-in-one average | PSNR: 28.30 -> 28.76 | Improved |
| Add hyperspherical contrastive loss instead of classifier loss (ablation) | SLER-IR · five-task all-in-one average | PSNR: 28.76 -> 31.73 | Improved |
| Add Global-Local Granularity Fusion (GLGF) (ablation) | SLER-IR · five-task all-in-one average | PSNR: 30.64 -> 31.73 | Improved |
| Increase experts per layer from 2 to 3 (ablation) | SLER-IR · five-task all-in-one average | PSNR: 31.22 -> 31.73 | Improved |
| Increase experts per layer from 3 to 4 (ablation) | SLER-IR · five-task all-in-one average | PSNR: 31.73 -> 31.32 | Hurt |
| Increase contrastive weight alpha from 0.5 to 1 (ablation) | SLER-IR · five-task all-in-one average | PSNR: 31.03 -> 31.73 | Improved |
| Increase contrastive weight alpha from 1 to 2 (ablation) | SLER-IR · five-task all-in-one average | PSNR: 31.73 -> 31.15 | Hurt |
| Use SLER-IR instead of MoCE-IR on three-task setting | SLER-IR · three-task all-in-one average | PSNR: 32.73 -> 33.14 | Improved |
| Use SLER-IR instead of MoCE-IR on five-task setting | SLER-IR · five-task all-in-one average | PSNR: 30.58 -> 31.73 | Improved |
| Use SLER-IR instead of MoCE-IR on SOTS dehazing (five-task) | SLER-IR · five-task all-in-one SOTS dehazing | PSNR: 30.84 -> 33.43 | Improved |
| Use SLER-IR instead of MoCE-IR on GoPro deblurring (five-task) | SLER-IR · five-task all-in-one GoPro deblurring | PSNR: 30.05 -> 31.27 | Improved |
Closest papers by meaning
- Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration (NeurIPS 2026 · similarity 0.88)
- Complexity Experts are Task-Discriminative Learners for Any Image Restoration (CVPR 2025 · similarity 0.87)
- Expert Weaving: Marrying Masked AutoRegressive and Diffusion Models for Unified Image Restoration (ECCV 2026 · similarity 0.87)
- Retrieve-to-Restore: Efficient All-in-One Image Restoration with a Retrieval-Based Degradation Bank (CVPR 2026 · similarity 0.86)
- Learning Where and What to Restore for Composite Image Restoration (NeurIPS 2026 · similarity 0.84)
- Omni-Kernel Network for Image Restoration (AAAI 2024 · 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.
