Flex4DHuman: Flexible Multi-view Video Diffusion for 4D Human Reconstruction
Researchers have developed Flex4DHuman, a novel diffusion model capable of reconstructing dynamic 4D human models from monocular or sparse multi-view videos. This model, built upon the Wan 2.1 1.3B text-to-video architecture, does not require explicit geometry priors like skeletons or depth maps. Instead, it uses relative camera pose conditioning and a unique five-axis positional encoding to generate synchronized dense multi-view videos, which can then be used with 4D Gaussian Splatting for detailed 4D reconstruction. Flex4DHuman has demonstrated superior performance on benchmark datasets and shows potential for applications in gaming, AR/VR, and video re-shooting. AI
IMPACT Enables scalable 4D content creation from casual videos, potentially impacting AR/VR and gaming industries.