Research map / Video / Video generation
Temporal-aware Flow Matching for Video Generation with Temporally Coherent Motion
Zirui Pan, Xin Wang, Yipeng Zhang, Yuwei Zhou, Wenwu Zhu
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 |
|---|---|---|---|
| fine-tune with TFM flow instead of standard FM flow (ablation) | Wan2.1-T2V-14B · VideoJam-Bench | VBench motion score: 0.815 -> 0.961 | Improved |
| fine-tune with TFM flow instead of standard FM flow (ablation) | Wan2.1-T2V-14B · Movie-Gen | VBench motion score: 0.810 -> 0.857 | Improved |
| use residual-type loss instead of solution-type loss (ablation) | TFM · VideoJam-Bench | VBench motion score: 0.894 -> 0.961 | Improved |
| use residual-type loss instead of solution-type loss (ablation) | TFM · Movie-Gen | VBench motion score: 0.811 -> 0.857 | Improved |
| apply TFM fine-tuning to Wan2.1 base model | Wan2.1-T2V-14B · VideoJam-Bench user study motion | % votes favoring TFM: 74.3% preference for TFM | Improved |
| apply TFM fine-tuning to Wan2.1 base model | Wan2.1-T2V-14B · Movie-Gen user study motion | % votes favoring TFM: 64.6% preference for TFM | Improved |
| apply TFM fine-tuning to Wan2.1 base model | Wan2.1-T2V-14B · VideoJam-Bench appearance | VBench appearance score: 0.780 -> 0.783 | No clear effect |
| apply TFM fine-tuning to Wan2.1 base model | Wan2.1-T2V-14B · Movie-Gen appearance | VBench appearance score: 0.791 -> 0.795 | No clear effect |
| fine-tune with TFM instead of FM | Wan2.1-T2V-1.3B · Temporal perception, frame permutation | relative loss increase after permutation | Improved |
| use TFM instead of FM during training | Wan2.1-T2V-14B · Training step cost | overall time per step: 60.51s -> 61.92s (+2.3%) | No clear effect |
| use TFM instead of FM during training | Wan2.1-T2V-1.3B · Training step cost | overall time per step: 18.71s -> 19.66s (+5.1%) | Hurt |
| ablated variants without TFM flow or residual loss (ablation) | TFM · Movie-Gen / VideoJam motion smoothness | Motion Smooth | No clear effect |
Closest papers by meaning
- Contextual Flow Matching: Adaptive Step Selection in Flow Models for Efficient Visual Generation (NeurIPS 2026 · similarity 0.92)
- FlowMo: Variance-Based Flow Guidance for Coherent Motion in Video Generation (NeurIPS 2025 · similarity 0.91)
- Optical-Flow Guided Prompt Optimization for Coherent Video Generation (CVPR 2025 · similarity 0.90)
- FlowVid: Taming Imperfect Optical Flows for Consistent Video-to-Video Synthesis (CVPR 2024 · similarity 0.89)
- Pyramidal Flow Matching for Efficient Video Generative Modeling (ICLR 2025 · similarity 0.89)
- Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · 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.
