Research map / Image generation / Fast diffusion sampling
SDDM: Score-Decomposed Diffusion Models on Manifolds for Unpaired Image-to-Image Translation
Shikun Sun, Longhui Wei, Junliang Xing, Jia Jia, Qi Tian
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 low-pass filter manifold with BAdaIN manifold (ablation) | SDDM · AFHQ Cat→Dog | FID: 67.56 -> 62.29 | Improved |
| Add multi-objective optimization on manifold (vs. without MOO) (ablation) | SDDM · AFHQ Cat→Dog | PNI and FID: FID 64.93 -> 62.29; PNI 0.024 -> 0 | Improved |
| Use 16x16 chunking blocks instead of 8x8 blocks (ablation) | SDDM · AFHQ Cat→Dog | FID and SSIM: 8x8 FID 54.56 SSIM 0.359; 16x16 FID 62.29 SSIM 0.422 | Mixed |
| Use 32x32 chunking blocks instead of 16x16 (ablation) | SDDM · AFHQ Cat→Dog | FID and SSIM: FID 62.29 -> 68.03; SSIM 0.422 -> 0.426 | Mixed |
| Epsilon policy 1 (single manifold iteration) instead of policy 3 (ablation) | SDDM · AFHQ Cat→Dog | FID and SSIM: FID 61.33 vs 62.29 (better); SSIM 0.413 vs 0.422 (worse) | Mixed |
| Epsilon policy 2 (iterate twice at every step) instead of policy 3 (ablation) | SDDM · AFHQ Cat→Dog | FID: 64.05 vs 62.29 | Hurt |
| Increase T0 manifold-stage iterations (policy 3) versus uniform single iteration (ablation) | SDDM · AFHQ Cat→Dog | SSIM: SSIM 0.413 -> 0.422 | Improved |
| SDDM with 100 steps vs EGSDE with 1000 steps | SDDM · AFHQ Cat→Dog | FID: 62.29 vs 65.82 (EGSDE*, 1000 steps) | Improved |
| SDDM vs EGSDE at 200 steps (EGSDE**) | SDDM · AFHQ Cat→Dog | FID: 62.29 vs 70.16 | Improved |
| SDDM vs EGSDE at 1000 steps on Wild→Dog | SDDM · AFHQ Wild→Dog | FID: 57.38 vs 59.75 (FID better) but SSIM 0.328 vs 0.343 (worse) | Mixed |
| SDDM vs EGSDE on Male→Female (1000 steps) | SDDM · CelebA-HQ Male→Female | FID: 44.37 vs 41.93 (EGSDE* better) | Hurt |
| Use SDDM with T0=0.6T (SDDM†) instead of default T0=0.5T | SDDM · AFHQ Cat→Dog | FID and SSIM: FID 62.29 -> 49.43; SSIM 0.422 -> 0.361 | Mixed |
Later papers that compared with this method
4 later accepted papers report SDDM as a baseline in their results tables; 3 of them report a better number on most of the results they share with it.
- Unified Discrete Diffusion for Categorical Data (JMLR 2025 · better on 3 of 3 shared results)
- Learning to Match Unpaired Data with Minimum Entropy Coupling (ICML 2025 · better on 0 of 1 shared result)
- Formulating Discrete Probability Flow Through Optimal Transport (NeurIPS 2023 · better on 11 of 20 shared results)
- Constructing Non-isotropic Gaussian Diffusion Model Using Isotropic Gaussian Diffusion Model for Image Editing (NeurIPS 2023 · better on 2 of 2 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- EGSDE* — EGSDE: Unpaired Image-to-Image Translation via Energy-Guided Stochastic Differential Equations (NeurIPS 2022) · better on 5 of 6 shared results
- SDEdit* — SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations (ICLR 2022) · better on 2 of 2 shared results
- StarGAN v2* — StarGAN v2: Diverse Image Synthesis for Multiple Domains (CVPR 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Adapting to Unknown Low-Dimensional Structures in Score-Based Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.88)
- DCTdiff: Intriguing Properties of Image Generative Modeling in the DCT Space (ICML 2025 · similarity 0.87)
- TCFG: Tangential Damping Classifier-free Guidance (CVPR 2025 · similarity 0.87)
- High-Fidelity Diffusion-Based Image Editing (AAAI 2024 · similarity 0.87)
- Score Distillation via Reparametrized DDIM (NeurIPS 2024 · 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.
