Research map / Video / Video generation
Seeing Fast and Slow: Learning the Flow of Time in Videos
Yen-Siang Wu, Rundong Luo, Jingsen Zhu, Tao Tu, Ali Farhadi, Matthew Wallingford, Yu-Chiang Frank Wang, Steve Marschner, Wei-Chiu Ma
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 iterative prediction (3 unrolls) to the speed estimator (ablation) | speed estimator · Playback speed estimation (111 videos) | Pearson rho / RMSE: 0.680 -> 0.735 rho; 0.917 -> 0.649 RMSE | Improved |
| add iterative prediction to VideoLLM (ablation) | VideoLLM · Playback speed estimation (111 videos) | Pearson rho: 0.426 -> 0.552 | Improved |
| train speed estimator on SloMo-44K instead of standard videos only (no IP) (ablation) | speed estimator · Playback speed estimation (111 videos) | Pearson rho: 0.242 -> 0.680 | Improved |
| train speed estimator on SloMo-44K instead of standard videos (with IP) (ablation) | speed estimator · Playback speed estimation (111 videos) | Pearson rho / RMSE: 0.632 -> 0.735 rho; 0.737 -> 0.649 RMSE | Improved |
| train speed-conditioned video model on SloMo-44K instead of standard-FPS videos (ablation) | Wan2.1-I2V speed-conditioned model · Speed-conditioned generation (48 triplets) | FVD: 1392.9 -> 1114.1 | Improved |
| explicit speed conditioning with SloMo-44K fine-tuning versus prompt modifiers on Wan2.1 | Wan2.1-I2V · Speed-conditioned generation (48 triplets) | FID: 72.2 -> 68.4 | Improved |
| fine-tune Wan2.1-VACE LoRA on SloMo-44K for temporal super-resolution (blurred input) | Wan2.1-VACE · Temporal super-resolution, blurred input, SloMo-44K-Test | FVD: 436.1 -> 134.3 | Improved |
| fine-tune on SloMo-44K for temporal super-resolution (clear input) | Wan2.1-VACE · Temporal super-resolution, clear input, DAVIS | LPIPS / FID: LPIPS 0.203 vs FILM 0.200 | Mixed |
| use audio pitch shift as free supervision for visual speed-change detection | VideoMAEv2 detector · Speed-change detection (test split) | accuracy (%): 59.5 -> 92.4 vs Gemini 2.5 | Improved |
| speed-conditioned model on VBench quality metrics versus Wan2.1 | Wan2.1-I2V · VBench | imaging quality: 0.6367 -> 0.6337 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SpeedNet — SpeedNet: Learning the Speediness in Videos (CVPR 2020) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SlowFast-VGen: Slow-Fast Learning for Action-Driven Long Video Generation (ICLR 2025 · similarity 0.89)
- Before It Fades: Reinforcing Temporal Representations at Inference Time in VideoLLMs (NeurIPS 2026 · similarity 0.88)
- FlowMo: Variance-Based Flow Guidance for Coherent Motion in Video Generation (NeurIPS 2025 · similarity 0.87)
- Where Do We (Not) Need Temporal Context in Low-Resource Video Task Adaptation? (NeurIPS 2026 · similarity 0.87)
- DELTAVID: Enhancing Fine-Grained Spatiotemporal Perception with Cross-Video Differences (NeurIPS 2026 · similarity 0.86)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.86)
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.
