Research map / Video / Video generation
SHIFT: Motion Alignment in Video Diffusion Models with Adversarial Hybrid Fine-Tuning
Xi Ye, Wenjia Yang, Yangyang Xu, Xiaoyang Liu, Duo Su, Mengfei Xia, Jun Zhu
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 noise alignment (shared noise level for SFT and AWR terms) (ablation) | IMR-only SHIFT (A0) · SVD ablation | VBench Appearance: 81.62 -> 82.53 | Improved |
| add noise alignment (ablation) | IMR-only SHIFT (A0) · SVD ablation | FVD: 502.60 -> 472.20 | Improved |
| add noise alignment (ablation) | IMR-only SHIFT (A0) · SVD ablation | VBench Motion: 91.01 -> 87.18 | Hurt |
| add adversarial reward model updates (ablation) | A1 · SVD ablation | FVD: 472.20 -> 451.20 | Improved |
| add adversarial reward model updates (ablation) | A1 · SVD ablation | VBench Overall: 84.08 -> 84.47 | Improved |
| add long-term motion reward (ablation) | A1 · SVD ablation | VBench Motion: 87.18 -> 89.31 | Improved |
| add long-term motion reward (ablation) | A1 · SVD ablation | Motion Score: 3.75 -> 4.14 | Improved |
| add adversarial reward model to A3 (ablation) | A3 · SVD ablation | FVD: 460.00 -> 401.76 | Improved |
| use ViT-Base instead of ResNet-34 IMR encoder (ablation) | IMR-only SHIFT · SVD ablation (epoch 6) | FVD: 636.2 -> 502.6 | Improved |
| SHIFT versus SFT fine-tuning | SVD · VBench-I2V SVD | Motion Score: 2.69 -> 4.40 | Improved |
| SHIFT versus FlowGRPO training cost | SVD · VBench-I2V SVD | Training time: 19.35x -> 2.46x | Improved |
| SHIFT versus SFT on Wan2.2 | Wan2.2-TI2V · WISA-80K deformation | Motion Score: 1.94 -> 2.93 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Track4Gen — Track4Gen: Teaching Video Diffusion Models to Track Points Improves Video Generation (CVPR 2025) · better on 3 of 4 shared results
- DenseDPO — DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models (NeurIPS 2025) · better on 4 of 4 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
- Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation (NeurIPS 2024 · similarity 0.89)
- FlowMo: Variance-Based Flow Guidance for Coherent Motion in Video Generation (NeurIPS 2025 · similarity 0.88)
- Identifying and Solving Conditional Image Leakage in Image-to-Video Diffusion Model (NeurIPS 2024 · similarity 0.88)
- Training-free Camera Control for Video Generation (ICLR 2025 · similarity 0.88)
- Emergent Temporal Correspondences from Video Diffusion Transformers (NeurIPS 2025 · similarity 0.88)
- EDEN: Enhanced Diffusion for High-quality Large-motion Video Frame Interpolation (CVPR 2025 · similarity 0.88)
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.
