SHIFT: Motion Alignment in Video Diffusion Models with Adversarial Hybrid Fine-Tuning
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866917351359774720 |
|---|---|
| author | Ye, Xi Yang, Wenjia Xu, Yangyang Liu, Xiaoyang Su, Duo Xia, Mengfei Zhu, Jun |
| author_facet | Ye, Xi Yang, Wenjia Xu, Yangyang Liu, Xiaoyang Su, Duo Xia, Mengfei Zhu, Jun |
| contents | Image-conditioned Video diffusion models achieve impressive visual realism but often suffer from weakened motion fidelity, e.g., reduced motion dynamics or degraded long-term temporal coherence, especially after fine-tuning. We study the problem of motion alignment in video diffusion models post-training. To address this, we introduce pixel-motion rewards based on pixel flux dynamics, capturing both instantaneous and long-term motion consistency. We further propose Smooth Hybrid Fine-tuning (SHIFT), a scalable reward-driven fine-tuning framework for video diffusion models. SHIFT fuses the normal supervised fine-tuning and advantage weighted fine-tuning into a unified framework. Benefiting from novel adversarial advantages, SHIFT improves convergence speed and mitigates reward hacking. Experiments show that our approach efficiently resolves dynamic-degree collapse in modern video diffusion models supervised fine-tuning. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_17426 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | SHIFT: Motion Alignment in Video Diffusion Models with Adversarial Hybrid Fine-Tuning Ye, Xi Yang, Wenjia Xu, Yangyang Liu, Xiaoyang Su, Duo Xia, Mengfei Zhu, Jun Computer Vision and Pattern Recognition Image-conditioned Video diffusion models achieve impressive visual realism but often suffer from weakened motion fidelity, e.g., reduced motion dynamics or degraded long-term temporal coherence, especially after fine-tuning. We study the problem of motion alignment in video diffusion models post-training. To address this, we introduce pixel-motion rewards based on pixel flux dynamics, capturing both instantaneous and long-term motion consistency. We further propose Smooth Hybrid Fine-tuning (SHIFT), a scalable reward-driven fine-tuning framework for video diffusion models. SHIFT fuses the normal supervised fine-tuning and advantage weighted fine-tuning into a unified framework. Benefiting from novel adversarial advantages, SHIFT improves convergence speed and mitigates reward hacking. Experiments show that our approach efficiently resolves dynamic-degree collapse in modern video diffusion models supervised fine-tuning. |
| title | SHIFT: Motion Alignment in Video Diffusion Models with Adversarial Hybrid Fine-Tuning |
| topic | Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2603.17426 |