Research map / Image generation / Fast diffusion sampling
Generative Manifold Distillation: Aligning Restoration Trajectories with the Natural Image Prior
Yuyang Hu, Mojtaba Sahraee-Ardakan, Kangfu Mei, Arpit Bansal, Chenyang Qi, Peyman Milanfar, Mauricio Delbracio
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 GMD offline adaptation | LDM-Deblur · REDS | FID: 37.67 -> 31.64 | Improved |
| add GMD offline adaptation | LDM-Deblur · RealBlur-J | FID: 25.74 -> 20.33 | Improved |
| add GMD offline adaptation | LDM-SR · DIV2K 4x SR (weak->strong) | MUSIQ: 54.41 -> 58.29 | Improved |
| add GMD offline adaptation | LDM-SR · DPED-iPhone | NIQE: 5.467 -> 4.300 | Improved |
| add quality-gated NIMA filtering (ablation) | GMD · REDS | FID: 41.26 -> 31.64 | Improved |
| add source ID data at ratio 0.9 (vs pseudo-targets only) (ablation) | GMD · REDS | FID: 53.56 -> 31.64 | Improved |
| raise source ratio from 0.6 to 0.9 (ablation) | GMD · REDS | FID: 39.09 -> 31.64 | Improved |
| lower source ratio from 1.0 (no adaptation) to 0.95 (ablation) | GMD · REDS | FID: 37.67 -> 36.75 | Improved |
| use RF-Solver with attention injection instead of SDEdit for targets (ablation) | GMD · REDS | FID: 34.09 -> 31.64 | Improved |
| use SDEdit instead of RF-Solver targets (ablation) | GMD · REDS | MUSIQ: 63.67 -> 64.05 | Mixed |
| offline GMD instead of online FLUX projection (SDEdit) | LDM-Deblur · REDS | PSNR: 20.18 -> 24.35 | Improved |
| apply GMD to SwinIR student | SwinIR · REDS | FID: 38.40 -> 35.93 | Improved |
Closest papers by meaning
- Efficient Degradation-agnostic Image Restoration via Channel-Wise Functional Decomposition and Manifold Regularization (ICLR 2026 · similarity 0.88)
- Flow Straight to Reality: Perceptually Consistent Flow Matching for Efficient Image Restoration (ECCV 2026 · similarity 0.88)
- Generative Model Inversion Through the Lens of the Manifold Hypothesis (NeurIPS 2025 · similarity 0.86)
- Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration (NeurIPS 2026 · similarity 0.86)
- Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild (CVPR 2024 · similarity 0.85)
- Reversing Flow for Image Restoration (CVPR 2025 · similarity 0.85)
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.
