BREAKING
Motion 3-to-4 code released
How Motion 3-to-4 works
1
Monocular video in
↓
2
Static 3D shape
↓
3
Motion latent
↓
4
4D animated mesh
Feed-forward vs iterative
Prior methods
●
Iterative optimization
●
High compute cost
●
Limited generalization
Motion 3-to-4
New
●
Feed-forward inference
●
Seconds per video
●
Canonical mesh reference
Benchmarked vs L4GM, GVFD, V2M4
Open resources and constraints
1
GitHub code
↓
2
Hugging Face weights
↓
3
Needs Hunyuan3D-2.0
↓
4
GPU compute > 8.0
4D synthesis moves faster
AI NEWS BLITZ
A CVPR 2026-accepted method for 4D synthesis from monocular video is now public.