Skip to content

GauHuman

Train a 3D human model in minutes and render it in real time

Visit official site

What GauHuman is

The model is described as a 3D human model built on Gaussian diffusion that trains in 1-2 minutes and renders in real time at up to 189 FPS.

What you can do with it

  • Encode Gaussian diffusion in canonical space
  • Transform 3D Gaussians into posed space with linear blend skinning
  • Refine pose and skinning with dedicated modules
  • Initialise and prune Gaussians from a 3D human prior

Application scenarios

  • Researchers use GauHuman for efficient 3D human modelling research
  • Game developers use GauHuman for fast character rendering
  • Animation teams create real-time 3D human animation with GauHuman

Model notes

  • NeRF-based implicit frameworks are described as needing hours of training and seconds per frame
  • Splitting and cloning of Gaussians are guided by KL divergence
  • A new merge operation is added for further speed-up
  • Pose and LBS refinement learns human detail at negligible compute cost

Pros & cons

✓ What we like

  • Fast training compared with hours for NeRF-based frameworks
  • Pose and LBS refinement learns detail at negligible compute cost

! What to watch out for

  • Gaussians are initialised and pruned from a 3D human prior
  • Rendering depends on transforming Gaussians into posed space via LBS

FAQ

How fast are training and rendering?

The material states 1-2 minutes of training and real-time rendering at up to 189 FPS.

How are Gaussians moved into posed space?

Linear blend skinning transforms 3D Gaussians from canonical space to posed space.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review