Research map / Video / Video generation
Transition Matching Distillation for Fast Video Generation
Weili Nie, Julius Berner, Nanye Ma, Chao Liu, Saining Xie, Arash Vahdat
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 flow head rollout (TMD) to one-step DMD2-v distillation | DMD2-v · VBench, Wan2.1 14B one-step | overall score: 83.69 -> 84.24 | Improved |
| add flow head (N2H5) to two-step DMD2-v | DMD2-v · VBench, Wan2.1 1.3B | overall score: 84.39 -> 84.68 | Improved |
| add flow head to two-step DMD2-v on 14B | DMD2-v · VBench, Wan2.1 14B two-step | overall score: 84.79 -> 84.92 vs rCM 85.05 | Hurt |
| use Conv3D discriminator head instead of Conv1D-2D or attention (ablation) | DMD2-v · VBench, one-step 1.3B | overall score: 82.32 -> 83.24 | Improved |
| add GAN loss (ablation) | DMD2-v · VBench, one-step 1.3B | overall score: 81.63 -> 83.24 | Improved |
| add KD warm-up (ablation) | DMD2-v one-step · VBench, 1.3B | overall score: 83.06 -> 83.24 | Improved |
| add KD warm-up (ablation) | DMD2-v two-step · VBench, 1.3B | overall score: 84.39 -> 83.79 | Hurt |
| apply timestep shift to student timesteps (ablation) | DMD2-v two-step · VBench, 1.3B | overall score: 83.44 -> 84.39 | Improved |
| apply timestep shift to DMD loss noise level (ablation) | DMD2-v one-step · VBench, 1.3B | overall score: 83.22 -> 83.24 | No clear effect |
| use MeanFlow instead of flow matching for stage-1 pretraining (ablation) | TMD N4H5 · VBench, 1.3B two-step | overall score: 84.29 -> 84.67 | Improved |
| use MeanFlow pretraining (ablation) | TMD N2H5 · VBench, 1.3B two-step | overall score: 84.61 -> 84.68 | No clear effect |
| add flow head rollout during distillation (ablation) | TMD · VBench convergence | overall score | Improved |
Closest papers by meaning
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.92)
- Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation (NeurIPS 2024 · similarity 0.91)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.90)
- CoDMD: Copula-aware Distribution Matching Distillation for Fast Video Generation (NeurIPS 2026 · similarity 0.90)
- Transition Matching: Scalable and Flexible Generative Modeling (NeurIPS 2025 · similarity 0.90)
- Emergent Temporal Correspondences from Video Diffusion Transformers (NeurIPS 2025 · similarity 0.90)
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.
