LocalDPO: A CVPR 2026 Method for Fine‑Grained Preference Optimization in Video Diffusion Models

Paper Background

CVPR is the premier computer‑vision conference; the 2026 edition received over 16,000 submissions with an acceptance rate of ~25%. This paper was selected and focuses on post‑training of large video diffusion models, where generating clearer, more stable, and human‑preferred videos remains challenging.

Problem Statement

Existing video DPO approaches rely on multiple sampled videos, external reward models, or costly human annotations, and they supervise the whole video globally. This leads to high annotation cost, ambiguous global scores, and insufficient attention to local artifacts such as flickering, texture collapse, and temporal inconsistency.

Method Overview (LocalDPO)

LocalDPO eliminates the need for external scoring or manual labeling by treating high‑quality real videos as positive samples. Negative samples are created by applying controlled spatio‑temporal degradation to local regions of the same video, ensuring semantic and global structure consistency while degrading only the targeted details.

Automatic Preference Pair Construction

Positive sample: directly use a real high‑resolution video.

Negative sample: apply a random 3D Bezier‑curve mask to a local spatio‑temporal region, then use a frozen pre‑trained video diffusion model (VDM) to locally repaint the masked area, producing a version with degraded texture, blur, or temporal artifacts.

The positive and negative videos differ only in the degraded region, providing a high‑confidence preference pair (real > degraded).

Dataset Construction

The authors collected and filtered 63 K high‑quality video clips covering diverse scenes and stable motion. Each clip is paired with a structured textual description generated by a vision‑language model (VLM) to support text‑conditioned training.

Region‑Aware DPO Loss

During training, the loss is computed only on the degraded region (region‑aware DPO loss) to focus the model on correcting local defects. To preserve global structure and motion, the region‑aware loss is combined with standard DPO loss and supervised fine‑tuning (SFT) loss in a mixed objective.

Experiments

LocalDPO was evaluated on three mainstream video diffusion models: CogVideoX‑2B, CogVideoX‑5B, and Wan2.1‑1.3B. Baselines included SFT, Vanilla DPO, DenseDPO, and other prior methods. Evaluation metrics covered automatic benchmarks (VBench, VideoJAM) and human‑centric scores (aesthetic, clarity, HPSv2, PickScore, Image Reward, Video Align).

Quantitative Results

Across all models and benchmarks, LocalDPO achieved the highest scores, especially on visual‑quality metrics, demonstrating superior modeling of high‑frequency texture, edges, and fine details. Table 1 (shown below) summarizes the VBench and VideoJAM improvements.

Human Evaluation

A 20‑person study compared LocalDPO with the strongest baseline (DenseDPO) on four dimensions: visual quality, motion quality, text alignment, and overall quality. Results showed a clear preference for LocalDPO, with higher average win rates.

Qualitative Evaluation

Visual comparisons (Figures 6‑8) illustrate that LocalDPO produces richer local textures, sharper frames, fewer artifacts, more stable temporal dynamics, and better semantic alignment with prompts compared to SFT, Vanilla DPO, and DenseDPO.

Conclusion and Outlook

LocalDPO provides an efficient, stable, and fine‑grained approach for aligning video diffusion models with human preferences without extra annotation or external reward models. Future work will incorporate semantic masks from grounding models such as Grounding‑DINO and SAM to further improve control over key object regions and extend preference alignment to multimodal generation.