Research map / Video / Video generation
DynamicsBoost: Dynamic Plausible Video Generation via Annotation-Free Continuation Preference Optimization
Jiaxing Li, Jiepeng Wang, Junyao Gao, Yang Liu, Eric Li, Bo An, Hao-Xiang Guo
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 |
|---|---|---|---|
| compute DPO loss only on continuation regions (2)-(3), excluding shared prefix (ablation) | DynamicsBoost · VBench | Overall Consistency: 22.15 -> 25.64 | Improved |
| restrict loss to Region (3) only instead of Regions (2)-(3) (ablation) | DynamicsBoost · VBench | Dynamic Degree: 44.92 -> 38.21 | Hurt |
| use randomized negative and positive continuation lengths instead of fixed N2=N (ablation) | DynamicsBoost · VBench | Overall Consistency: 25.44 -> 25.64 | Improved |
| use pure T2V (N1=0) as negative samples (ablation) | DynamicsBoost · VBench | Overall Consistency: 25.64 -> 25.31 | Hurt |
| fix N1=0, N2=N for continuation lengths (ablation) | DynamicsBoost · VBench | Dynamic Degree: 44.92 -> 34.32 | Hurt |
| add more reference frames in continuation (ablation) | continuation model · VBench and VideoReward | Aesthetic Quality: 36.10 -> 46.64 | Improved |
| add pure T2V (0 reference frames) as the lowest-ordered sample (ablation) | continuation preference ordering · VBench | Aesthetic Quality: 36.10 -> 44.83 | Hurt |
| add learnable continuation task prompt (ablation) | continuation model (4/13) · VBench | Aesthetic Quality: 40.21 -> 43.81 | Improved |
| add timestep masking (ablation) | continuation model (4/13) · VBench | Imaging Quality: 49.29 -> 57.98 | Improved |
| apply annotation-free continuation preference pairs vs Flow-DPO with VideoReward labels | Flow-DPO · VBench | Dynamic Degree: 31.22 -> 44.92 | Improved |
| apply SFT cold start on small dataset | pretrained model · VBench | Motion Smoothness: 98.81 -> 96.84 | Hurt |
| compare proposed method with Flow-DenseDPO on imaging quality | DynamicsBoost · VBench | Imaging Quality: 67.91 -> 66.81 | Mixed |
Closest papers by meaning
- VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · similarity 0.91)
- Improving Video Generation with Human Feedback (NeurIPS 2025 · similarity 0.90)
- Hierarchical Fine-grained Preference Optimization for Physically Plausible Video Generation (NeurIPS 2025 · similarity 0.90)
- Boosting Text-to-Video Generative Model with MLLMs Feedback (NeurIPS 2024 · similarity 0.90)
- DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models (NeurIPS 2025 · similarity 0.90)
- SafeSora: Towards Safety Alignment of Text2Video Generation via a Human Preference Dataset (NeurIPS 2024 · similarity 0.89)
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.
