Research map / Video / Video generation
MotionEnhancer: Leveraging Video Diffusion for Motion-Enhanced Vision-Language Models
Yifan Xu, Chao Zhang, Ruifei Ma, Fei Gao, Zhifei Yang, Jiaxing Qi, Zhipeng Chen
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 Motion-sensitive Head Selection (MHS) (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · MotionBench dev | Overall accuracy: 54.83 -> 56.60 | Improved |
| add Motion-sensitive Head Selection (MHS) (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · FAVOR-Bench close-ended | Overall accuracy: 44.83 -> 46.65 | Improved |
| add Motion-salient Text Token Identification (MTTI) (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · MotionBench dev | Overall accuracy: 54.83 -> 55.80 | Improved |
| add MTTI on top of MHS (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · MotionBench dev | Overall accuracy: 56.60 -> 57.04 | Improved |
| add MHS on top of MTTI (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · FAVOR-Bench close-ended | Overall accuracy: 45.47 -> 46.88 | Improved |
| use 5 DDIM steps instead of 2 (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · MotionBench dev | Overall accuracy: 52.76 -> 57.04 | Improved |
| use 5 DDIM steps instead of 2 (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · FAVOR-Bench close-ended | Overall accuracy: 24.01 -> 46.88 | Improved |
| use 10 DDIM steps instead of 5 (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · FAVOR-Bench close-ended | Overall accuracy: 46.88 -> 49.02 | Improved |
| use 10 DDIM steps instead of 5 (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · MotionBench dev | Overall accuracy: 57.04 -> 57.51 | Improved |
| select 50% heads/tokens instead of 25% (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · MotionBench dev | Overall accuracy: 55.33 -> 57.04 | Improved |
| select 50% heads/tokens instead of 25% (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · FAVOR-Bench close-ended | Overall accuracy: 47.12 -> 46.88 | Hurt |
| select 50% heads/tokens instead of 75% (ablation) | MotionEnhancer (Qwen2.5-VL-7B) · MotionBench dev | Overall accuracy: 54.88 -> 57.04 | Improved |
Closest papers by meaning
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.91)
- FAVOR-Bench: A Comprehensive Benchmark for Fine-Grained Video Motion Understanding (NeurIPS 2025 · similarity 0.89)
- Towards Understanding Camera Motions in Any Video (NeurIPS 2025 · similarity 0.89)
- Video Diffusion Models are Training-free Motion Interpreter and Controller (NeurIPS 2024 · similarity 0.88)
- HIMMEL: Hierarchical Interleaved Multi-stream Motion Encoding for Long Video Understanding (NeurIPS 2026 · similarity 0.88)
- Masked Diffusion Vision-Language Models for Temporal Action Localization (NeurIPS 2026 · 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.
