NVIDIA 发了新方法 MOTIVE，帮你从海量视频里挑出对运动建模最有效的片段来微调，效果比基座好很多。
NVIDIA Research提出MOTIVE方法，用于识别视频模型训练中对运动建模真正重要的片段。MOTIVE通过重新加权训练信号，聚焦运动区域、淡化静态背景，并依据对运动的影响程度为每个片段打分。仅用少量高影响子集微调，即可提升时序动态质量，在VBench动态性指标上显著提升，并以74.1%的人类偏好胜率优于基线模型。该工作获ICML2026杰出论文荣誉提名。
When you train a video model for motion, only some clips actually matter. New from NVIDIA Research ...
When you train a video model for motion, only some clips actually matter. New from NVIDIA Research and recipient of an Outstanding Paper Honorable Mention at #ICML2026 : MOTIVE shows exactly which training clips improve motion. MOTIVE re-weights training signals toward moving regions and away from static backgrounds, then scores each clip by its influence on motion. You can then curate a small high-influence subset to fine-tune on and get better temporal dynamics. Because it attributes motion rather than appearance, MOTIVE curates data that boosts VBench dynamics and wins 74.1% human preference against the base model. Your browser does not support the video tag. 🔗 View on Twitter 💬 2 🔄 4 ❤️ 32 👀 3552 📊 7 ⚡