Research map / Image generation / Fast diffusion sampling
Be Tangential to Manifold: Discovering Riemannian Metric for Diffusion Models
Shinnosuke Saito, Takashi Matsubara
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 |
|---|---|---|---|
| use score-Jacobian Riemannian metric geodesics instead of SLERP | Stable Diffusion v2.1-base · image interpolation (MB, CelebA-HQ, AFHQ) | PPL: e.g. 0.871 -> 0.767 on AFHQ | Improved |
| use score-Jacobian metric instead of density-based GeoDiff | Stable Diffusion v2.1-base · image interpolation | FID: 25.80 -> 21.01 on AFHQ | Improved |
| use score-Jacobian metric instead of FIM-based metric | Stable Diffusion v2.1-base · image interpolation | PPL: 5.249 -> 0.767 on AFHQ | Improved |
| use metric instead of SLERP for PDV | Stable Diffusion v2.1-base · image interpolation | PDV: best on MB(A) 0.021; SLERP better on AFHQ 0.022 vs 0.023 | Mixed |
| use metric geodesics for video frame interpolation | Stable Diffusion v2.1-base · video frame interpolation | MSE: 13.253 -> 8.777 on DAVIS (vs GeoDiff) | Improved |
| use LERP vs metric on DAVIS MSE | LERP · video frame interpolation DAVIS | MSE: 12.135 -> 8.777 | Improved |
| add metric-based guidance correction to CFG | Stable Diffusion v2.1-base · MS-COCO 30k | FID: 17.28 -> 16.04 at w=12.5 | Improved |
| add metric-based guidance correction to CFG | Stable Diffusion v2.1-base · MS-COCO 30k | CLIP Score: 0.315 -> 0.315 at w=12.5 | No clear effect |
| replace CFG with CFG++ | Stable Diffusion v2.1-base · MS-COCO 30k | FID: 14.29 -> 13.98 at w=7.5; 11.69 -> 11.87 at w=5.0 | Mixed |
| apply guidance correction during LCM distillation | LCM student · MS-COCO 4-step | FID: 17.91 -> 16.98 | Improved |
Closest papers by meaning
- Generative Modeling on Manifolds Through Mixture of Riemannian Diffusion Processes (ICML 2024 · similarity 0.91)
- Improving the Euclidean Diffusion Generation of Manifold Data by Mitigating Score Function Singularity (NeurIPS 2025 · similarity 0.90)
- Diffusion Models Encode the Intrinsic Dimension of Data Manifolds (ICML 2024 · similarity 0.90)
- What Matters for Diffusion-Friendly Latent Manifold? Prior-Aligned Autoencoders for Latent Diffusion (NeurIPS 2026 · similarity 0.90)
- Efficient Diffusion Models for Symmetric Manifolds (ICML 2025 · similarity 0.89)
- Manifold Diffusion Fields (ICLR 2024 · similarity 0.89)
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.
