Research map / Video / Video generation
SwiftVideo: A Unified Framework for Few-Step Video Generation Through Trajectory-Distribution Alignment
Yanxiao Sun, Jiafu Wu, Yun Cao, Chengming Xu, Yabiao Wang, Weijian Cao, Donghao Luo, Chengjie Wang, Yanwei Fu
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 distribution alignment to CCD (ablation) | SwiftVideo · OpenVid-1M test, 4 steps | FVD: 172.54 -> 136.14 | Improved |
| use continuous-time instead of discrete-time consistency distillation before DA (ablation) | SwiftVideo · OpenVid-1M test, 4 steps | FVD: 151.23 -> 136.14 | Improved |
| add consistency distillation (CCD) before DA (ablation) | SwiftVideo · OpenVid-1M test, 4 steps | FVD: 163.54 -> 136.14 | Improved |
| add trajectory alignment (DPO with synthetic step preferences) (ablation) | SwiftVideo · OpenVid-1M test, 4 steps | FVD: 136.14 -> 118.89 | Improved |
| use lognorm(-0.6,1.4) instead of uniform(0,1) timestep sampler (ablation) | SwiftVideo · OpenVid-1M test, 4 steps | FVD: 140.36 -> 136.14 | Improved |
| use lognorm(-0.6,1.4) instead of lognorm(-0.8,1.0) timestep sampler (ablation) | SwiftVideo · OpenVid-1M test, 4 steps | FVD: 138.5 -> 136.14 | Improved |
| use synthetic preference pairs from different step counts instead of real data as preferred samples (ablation) | trajectory alignment · OpenVid-1M | FVD trend | Improved |
| add reflow loss regularization to DPO (ablation) | trajectory alignment · OpenVid-1M | Win Diff stability / peak performance | Improved |
| vary lambda_RF among 1, 2, 3 (ablation) | trajectory alignment · OpenVid-1M | training convergence | No clear effect |
| use tangent normalization instead of tangent clipping (ablation) | CCD · OpenVid-1M | quality | Improved |
| use SwiftVideo (4 steps) instead of DMD2 (4 steps) | DMD2 · OpenVid-1M test | FVD: 167.42 -> 118.89 | Improved |
| use SwiftVideo (4 steps) instead of teacher with 4 steps | Wan2.1 teacher · OpenVid-1M test | FVD: 321.95 -> 118.89 | Improved |
Closest papers by meaning
- VDOT: Efficient Unified Video Creation via Optimal Transport Distillation (CVPR 2026 · similarity 0.89)
- From Slow Bidirectional to Fast Autoregressive Video Diffusion Models (CVPR 2025 · similarity 0.89)
- CoDMD: Copula-aware Distribution Matching Distillation for Fast Video Generation (NeurIPS 2026 · similarity 0.88)
- OSV: One Step is Enough for High-Quality Image to Video Generation (CVPR 2025 · similarity 0.88)
- Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis (CVPR 2024 · similarity 0.88)
- FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video Synthesis (CVPR 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.
