Skip to content

PSHuman

Rebuild a lifelike 3D human model from a single image

Visit official site

What PSHuman is

The framework is described as using multi-view diffusion models and explicit reconstruction to rebuild a lifelike 3D human model from a single image. A cross-scale diffusion model jointly models global full-body shape and local face features, while SMPL-X body priors strengthen cross-view body shape consistency.

What you can do with it

  • Rebuild a lifelike 3D human model from one image
  • Model full-body shape and local face features with cross-scale multi-view diffusion
  • Recover a textured human mesh from multi-view normal and colour images through differentiable rasterisation
  • Keep body shape consistent across views with SMPL-X body priors

Application scenarios

  • A game developer quickly generates lifelike game character models
  • An animation maker builds high-precision 3D human models for an animated film
  • A virtual reality company creates realistic virtual people for a VR experience

Reconstruction steps

  • Prepare a full-body person image
  • Feed the image and a predicted SMPL-X model into the framework
  • Generate six-view global full-body images and local face images with the multi-view diffusion model
  • Deform and rebuild the SMPL-X model through explicit human sculpting using the generated normal and colour maps

Pros & cons

✓ What we like

  • Handles complex self-occlusion and avoids geometric distortion in face detail
  • New views keep rich detail and identity features
  • Also handles non-real figures such as anime characters

! What to watch out for

  • The flow takes a full-body image and a predicted SMPL-X model as input
  • Reconstruction depends on the generated multi-view normal and colour images

FAQ

What does it take as input?

The material describes a single full-body image plus a predicted SMPL-X model in the reconstruction flow.

What subjects can it rebuild?

The material states it applies to real people and also non-real figures such as anime characters.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review