Research map / Image generation / Fast diffusion sampling
Adaptive Noise Covariance Scheduling under Riemannian Metrics for Diffusion Models
Bolin Deng, Die Hu, Bin Tan, Jun Wu
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 |
|---|---|---|---|
| replace Cholesky interpolation with BW geodesic path (fixed schedule) (ablation) | BNDM · AFHQ-Cat (64x64) | FID: 7.95 -> 6.40 | Improved |
| replace LE/AIRM geodesic with BW geodesic path (fixed schedule) (ablation) | BNDM · AFHQ-Cat (64x64) | FID: 7.95 -> 6.40 | Improved |
| use LE/AIRM geodesic instead of Cholesky interpolation (fixed schedule) (ablation) | BNDM · AFHQ-Cat (64x64) | FID: 7.95 -> 7.95 | No clear effect |
| add adaptive transition schedule to BW geodesic (ablation) | BW+F · AFHQ-Cat (64x64) | FID: 6.40 -> 6.17 | Improved |
| add adaptive transition schedule to Cholesky interpolation (ablation) | CH+F · AFHQ-Cat (64x64) | FID: 7.95 -> 6.43 | Improved |
| add adaptive transition schedule to LE/AIRM geodesic (ablation) | LA+F · AFHQ-Cat (64x64) | FID: 7.95 -> 8.36 | Hurt |
| add adaptive schedule to BW geodesic (ablation) | BW+F · AFHQ-Cat (64x64) | KID: 0.0034 -> 0.0034 | Improved |
| replace BNDM path and schedule with BW geodesic and adaptive schedule | BNDM · AFHQ-Cat (64x64) | KID: 0.0052 -> 0.0034 | Improved |
| replace BNDM with ours at 5 sampling steps | BNDM · AFHQ-Cat (64x64), T=5 | FID: 82.93 -> 57.35 | Improved |
| replace BNDM with ours at 250 sampling steps on AFHQ | BNDM · AFHQ (64x64), T=250 | FID: 12.79 -> 9.51 | Improved |
| extend the method to DDIM | DDIM · AFHQ-Cat (64x64) | FID: 9.82 -> 6.33 | Improved |
| replace BNDM with ours for super-resolution | BNDM · AFHQ-Cat super-resolution (64 to 128) | PSNR: 29.30 -> 30.01 | Improved |
Closest papers by meaning
- Diffusion Priors for Variational Likelihood Estimation and Image Denoising (NeurIPS 2024 · similarity 0.89)
- The Geometry of Noise: Why Diffusion Models Don't Need Noise Conditioning (NeurIPS 2026 · similarity 0.89)
- Free Hunch: Denoiser Covariance Estimation for Diffusion Models Without Extra Costs (ICLR 2025 · similarity 0.89)
- Listening to the noise: Blind Denoising with Gibbs Diffusion (ICML 2024 · similarity 0.88)
- A Variational Perspective on Solving Inverse Problems with Diffusion Models (ICLR 2024 · similarity 0.88)
- Diffusion Models With Learned Adaptive Noise (NeurIPS 2024 · similarity 0.88)
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.
