Boximator: Bring Fine-grained Motion Controllability to Video Synthesis | Bytedance Research

We propose Boximator, a method that significantly enhances the motion richness and controllability of video synthesis by introducing novel constraint forms and training techniques to video diffusion models. Here in this video, you will learn about: 1. 00:00 Why we need motion control beyond text 2. 00:56 What Boximator do in motion control 3. 02:04 How Boximator do in motion control 3.1 02:02 the general idea 3.2 03:00 model design 3.3 03:58 training challenge 3.4 05:37 data challenge 4. 05:54 Empirical results 5. 07:02 Use Boximator to do magical things We have published our paper on arXiv: We also made a demo page: If you want to try Boximator or have any problems, don’t hesitate to contact us! Just email to: @ (^_^ )