Research map / Video / Video generation
SGMD: Score Gradient Matching Distillation for Few-Step Video Diffusion Distillation
Zhuguanyu Wu, Ruihao Gong, Yang Yong, Yushi Huang, Xiangyu Fan, Lei Yang, Dahua Lin, Xianglong Liu
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 reverse-KL DMD2 objective with teacher stop-gradient Fisher (ablation) | DMD2 · VBench-T2V 4-step | OptFlow: 4.51 -> 8.18 | Improved |
| replace reverse-KL DMD2 objective with teacher stop-gradient Fisher (ablation) | DMD2 · VBench-T2V 4-step | VBench Quality: 85.05 -> 82.98 | Hurt |
| add dual potentials NR/RC with Fake-R reduced 5 to 1 (ablation) | TSG-Fisher · VBench-T2V 4-step | FVD: 126.7 -> 100.3 | Improved |
| add dual potentials NR/RC with Fake-R reduced 5 to 1 (ablation) | TSG-Fisher · VBench-T2V 4-step | VBench Quality: 82.98 -> 84.77 | Improved |
| add dual potentials NR/RC (ablation) | TSG-Fisher · VBench-T2V 4-step | VBench Semantic: 71.50 -> 75.64 | Improved |
| SGMD instead of DMD2 | DMD2 · VBench-T2V 4-step | OptFlow: 4.51 -> 9.29 | Improved |
| SGMD instead of DMD2 | DMD2 · VBench-T2V 4-step | VBench Dynamic Degree: 80.56 -> 93.06 | Improved |
| SGMD instead of DMD2 | DMD2 · VBench-T2V 4-step | VBench Semantic: 77.46 -> 75.64 | Hurt |
| SGMD instead of SIM-style coupling (TSG-SIM) | TSG-SIM · VBench-T2V 4-step | FVD: 193.0 -> 100.3 | Improved |
| increase lambda from 0.1 to 0.5 (ablation) | SGMD · VBench-T2V | VBench Total: 82.95 -> 79.54 | Hurt |
| decrease lambda from 0.1 to 0.05 (ablation) | SGMD · VBench-T2V | VBench Total: 82.95 -> 81.92 | Hurt |
| increase lambda from 0.1 to 0.2 (ablation) | SGMD · VBench-T2V | VBench Dynamic Degree: 93.06 -> 94.23; Total 82.95 -> 82.01 | Mixed |
Closest papers by meaning
- Distribution Backtracking Builds A Faster Convergence Trajectory for Diffusion Distillation (ICLR 2025 · similarity 0.89)
- Distribution Matching Distillation without Fake Score Network (NeurIPS 2026 · similarity 0.89)
- Continuous-Time Distribution Matching for Few-Step Diffusion Distillation (NeurIPS 2026 · similarity 0.88)
- CoDMD: Copula-aware Distribution Matching Distillation for Fast Video Generation (NeurIPS 2026 · similarity 0.88)
- One-step Diffusion with Distribution Matching Distillation (CVPR 2024 · similarity 0.88)
- Improved Distribution Matching Distillation for Fast Image Synthesis (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.
