Diffuman4D

Why we included this project

Diffuman4D tackles a genuinely hard problem in human view synthesis: keeping a reconstructed person consistent across time when you only have a handful of synchronized cameras. It pairs 3D Gaussian splatting with a spatio-temporal diffusion model, so rendered frames stay stable instead of flickering as the viewpoint moves. The repo is practical to try, with a quick-start install, example data from the DNA-Rendering dataset, and an interactive browser demo, so you can test it on your own sparse-view captures without much setup. Researchers following recent human view synthesis work will also find the EasyVolcap-based pipeline clearly documented, and the re-annotated multi-view labels are a useful addition for training and evaluation.