Research map / Video / Video generation
Inference-time Motion Calibration for Video Generation
Zile Huang, Ser-Nam Lim
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 Dynamic Latent Control (VFM gradient steering) (ablation) | HunyuanVideo · HunyuanVideo component ablation | FVD: 430.2 -> 358.6 | Improved |
| Add Dynamic RoPE Control alone (ablation) | HunyuanVideo · HunyuanVideo component ablation | FVD: 430.2 -> 401.3 | Improved |
| Combine DLC and DRC (ablation) | HunyuanVideo · HunyuanVideo component ablation | FVD: 358.6 -> 318.9 | Improved |
| Amortize with ATR (N=8) instead of full DLC (ablation) | HunyuanVideo · HunyuanVideo component ablation | FVD: 318.9 -> 328.5 | No clear effect |
| Use RAFT instead of frame difference for motion loss (ablation) | HunyuanVideo · HunyuanVideo VFM ablation (full DLC, no ATR) | FVD: 405.7 -> 354.8 | Improved |
| Start DLC at t=0.5 instead of t=1.0 or t=0.2 (ablation) | HunyuanVideo · HunyuanVideo DLC start step | FVD: 412.3 / 358.4 -> lower | Improved |
| Gradient steering vs Universal Guidance | HunyuanVideo · HunyuanVideo gradient-baseline comparison | FVD: 385.4 -> 328.5 | Improved |
| Gradient calibration vs search (EvoSearch, Noise Trajectory Search) | HunyuanVideo · VBench 2.0 T2V (HunyuanVideo) | Motion Rationality: 34.60 -> 37.20 (%) | Improved |
| Add calibration to Wan 2.1 I2V | Wan 2.1 · VBench I2V (Wan 2.1) | Camera Motion: 33.50% -> 36.80% | Improved |
| Add calibration to HunyuanVideo T2V | HunyuanVideo · VBench 2.0 T2V (HunyuanVideo) | Motion Smoothness: 98.70% -> 96.90% | Hurt |
| Add calibration to CogVideoX-1.5 T2V | CogVideoX-1.5 · VBench 2.0 T2V (CogVideoX-1.5) | Motion Smoothness: 96.92% -> 97.40% | Improved |
| Raise ATR interval to N=16 on HunyuanVideo (ablation) | HunyuanVideo · VBench 2.0 T2V (HunyuanVideo) | Motion Smoothness: 96.90% -> 96.70% | Mixed |
Closest papers by meaning
- Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout (CVPR 2026 · similarity 0.89)
- Time-to-Move: Training-Free Motion-Controlled Video Generation via Dual-Clock Denoising (ICLR 2026 · similarity 0.89)
- Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance (NeurIPS 2025 · similarity 0.88)
- MotionPro: A Precise Motion Controller for Image-to-Video Generation (CVPR 2025 · similarity 0.87)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.86)
- MotionClone: Training-Free Motion Cloning for Controllable Video Generation (ICLR 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.
