📊 How does it compare? MotionFlow outperforms state-of-the-art methods like DMT, VMC, and others by balancing:
✅ Motion Fidelity (realistic motion)
✅ Temporal Consistency (smooth transitions)
✅ Text Alignment (prompt accuracy)

MotionFlow is a novel invert-then-generate framework that uses cross-attention maps to transfer motion dynamics across contexts. It maintains motion fidelity while enabling scene edits. No extra training needed.

🌟 Exciting news! Our research, MotionFlow, is out! 🚀

A training-free method for motion transfer using pre-trained video diffusion models. 🎥✨

Big thanks to Hidir Yesiltepe, Connor Dunlop, and my advisor Pinar Yanardag. More from our lab soon—stay tuned! 🧵👇