Research map / Video / Video generation
TempoMaster: Efficient Long Video Generation via Next-Frame-Rate Prediction
Yukuo Ma, Cong Liu, Junke Wang, Junqi Liu, Haibin Huang, Zuxuan Wu, Chi Zhang, Xuelong Li
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 |
|---|---|---|---|
| generate video as next-frame-rate prediction (coarse-to-fine in frame rate) instead of autoregressive or bidirectional generation | TempoMaster (Wan2.2 14B) · VBench long video (500 frames) | Total score: 80.30 vs best baseline 79.52 | Improved |
| use next-frame-rate prediction on short videos | TempoMaster (Wan2.2 14B) · VBench short video (121 frames) | Total score: 80.76 vs best baseline 80.55 | Improved |
| compare against autoregressive-like methods on long vs short videos | MAGI-1 and SkyReels-V2 · VBench | Total score: MAGI-1 79.05 -> 78.50; SkyReels-V2 80.54 -> 79.17 | Hurt |
| VBench consistency and smoothness metrics reward low-motion videos | FramePack · VBench vs human study | Motion quality: Dynamic degree 16.82 on VBench; human motion quality 2.88 | Mixed |
| TempoMaster vs baselines in human evaluation | TempoMaster · Human study, long videos | Total score: 3.69 vs LongCat 3.58 | Improved |
| use two stages (6, 24 fps) with more parallel segments (1, 8) instead of (1, 4) (ablation) | TempoMaster inference · VBench long video (500 frames) | PFLOPs / total score: 526.48 -> 400.19 PFLOPs; 80.30 -> 80.29 | Improved |
| use parallel segments (1, 4, 8) instead of (1, 2, 4) in three stages (ablation) | TempoMaster inference · VBench long video (500 frames) | PFLOPs / total score: 774.23 -> 460.02 PFLOPs; 80.30 -> 80.30 | Improved |
| use parallel segments (1, 8, 8) instead of (1, 2, 4) in three stages (ablation) | TempoMaster inference · VBench long video (500 frames) | Total score: 80.30 -> 80.04 | Hurt |
| drop the intermediate 12 fps stage (two stages 6, 24 fps with segments (1, 4)) (ablation) | TempoMaster inference · VBench short video (121 frames) | Total score: 80.76 -> 80.55 | Hurt |
| use more aggressive parallelism (1, 4, 8) in three stages (ablation) | TempoMaster inference · VBench short video (121 frames) | Total score: 80.76 -> 80.26 | Hurt |
| add randomized temporal position indices during training (ablation) | TempoMaster · VBench long video (500 frames) | Total score: 80.00 -> 80.19 | Improved |
Closest papers by meaning
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.90)
- LongDiff: Training-Free Long Video Generation in One Go (CVPR 2025 · similarity 0.89)
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.89)
- WorldWeaver: Generating Long-Horizon Video Worlds via Rich Perception (NeurIPS 2025 · similarity 0.88)
- AdaState: Self-Evolving Anchors for Streaming Video Generation (NeurIPS 2026 · similarity 0.88)
- LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity (CVPR 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.
