Research map / Video / Video generation
Mind the Generative Details: Direct Localized Detail Preference Optimization for Video Diffusion Models
Zitong Huang, Kaidong Zhang, Yukang Ding, Chao Gao, Rui Ding, Ying Chen, Wangmeng Zuo
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 region-aware DPO loss on top of DPO+SFT (ablation) | LocalDPO (CogVideoX-2B) · VBench prompts | Imaging quality: 0.6514 -> 0.7080 | Improved |
| add region-aware DPO loss on top of DPO+SFT (ablation) | LocalDPO (CogVideoX-2B) · VBench prompts | Aesthetic quality: 0.6308 -> 0.6499 | Improved |
| add region-aware DPO loss on top of DPO+SFT (ablation) | LocalDPO (CogVideoX-2B) · VBench prompts | VideoAlign Overall: 7.8003 -> 7.8568 | Improved |
| add SFT loss to DPO loss (ablation) | LocalDPO (CogVideoX-2B) · VBench prompts | Aesthetic quality: 0.6303 -> 0.6308 | No clear effect |
| region-aware corruption negatives instead of vanilla-DPO lose samples with real positives (ablation) | LocalDPO (CogVideoX-2B) · VBench prompts | Imaging quality: 0.6577 -> 0.7080 | Improved |
| region-aware corruption negatives instead of vanilla-DPO lose samples with real positives (ablation) | LocalDPO (CogVideoX-2B) · VBench prompts | VideoAlign Overall: 7.7778 -> 7.8568 | Improved |
| real-video positives with generated negatives instead of vanilla win/lose pairs (ablation) | DPO (CogVideoX-2B) · VBench prompts | VideoAlign Overall: 7.7862 -> 7.7778 | No clear effect |
| LocalDPO vs vanilla DPO | CogVideoX-5B · VBench prompts | VideoAlign Overall: 9.5910 -> 10.2930 | Improved |
| LocalDPO vs vanilla DPO | CogVideoX-5B · VBench prompts | Aesthetic quality: 0.5953 -> 0.6274 | Improved |
| LocalDPO vs SFT | Wan2.1-1.3B · VBench prompts | VideoAlign Overall: 7.7283 -> 7.9588 | Improved |
| vanilla DPO vs baseline | CogVideoX-5B · VBench prompts | Aesthetic quality: 0.6110 -> 0.5953 | Hurt |
| LocalDPO vs vanilla DPO | Wan2.1-1.3B · VideoJAM benchmark | VideoAlign Overall: 7.0548 -> 7.4849 | Improved |
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.
- DenseDPO — DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models (NeurIPS 2025) · better on 6 of 6 shared results
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
- DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models (NeurIPS 2025 · similarity 0.92)
- VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · similarity 0.91)
- $\mathcal{V}ista\mathcal{DPO}$: Video Hierarchical Spatial-Temporal Direct Preference Optimization for Large Video Models (ICML 2025 · similarity 0.90)
- Personalized Preference Fine-tuning of Diffusion Models (CVPR 2025 · similarity 0.88)
- Diffusion-NPO: Negative Preference Optimization for Better Preference Aligned Generation of Diffusion Models (ICLR 2025 · similarity 0.88)
- InPO: Inversion Preference Optimization with Reparametrized DDIM for Efficient Diffusion Model Alignment (CVPR 2025 · similarity 0.87)
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.
