Skip to content

Human101

Rebuild a 3D human from a single-view video in 100 seconds

Visit official site

What Human101 is

The framework is described as a fast way to reconstruct a human body from a single view, training a 3D Gaussian model within 100 seconds and rendering images at 1024 resolution above 60 FPS. The material states it does so without storing Gaussian attributes for every frame.

What you can do with it

  • Reconstruct a high-fidelity 3D human from single-view video
  • Train a 3D Gaussian model in 100 seconds
  • Render images at 1024 resolution above 60 FPS
  • Use the end-to-end implementation code

How the pipeline works

  • Extract 2D human poses from a single-view video
  • Drive a 3D simulator with those poses to make matching 3D skeleton animation
  • Build a time-dependent 3D Gaussian model on the animation for real-time rendering

Application scenarios

  • Rebuild a dancer's 3D animation from a single-view video
  • Generate a digital self from a selfie video
  • Enable real-time human interaction in virtual reality scenes
  • Apply the reconstruction to visual content generation and digital human modeling

Pros & cons

✓ What we like

  • End-to-end implementation code is provided
  • Real-time rendering needs no stored per-frame attributes

! What to watch out for

  • Reconstruction starts from a single-view video as the input
  • Rendering depends on a time-dependent 3D Gaussian model built from animation

FAQ

How long does training take?

The material states a 3D Gaussian model is trained within 100 seconds.

What resolution and frame rate are stated?

The material states images at 1024 resolution rendered above 60 FPS.

Last reviewed: 2026-11-15

More AI video generation tools

View all →

How we review