Download
PIFuHD (Pixel-Aligned Implicit Function for 3D human reconstruction at high resolution) is a method and codebase to reconstruct high-fidelity 3D human meshes from a single image. It extends prior PIFu work by increasing resolution and detail, enabling fine geometry in cloth folds, hair, and subtle surface features. The method operates by learning an implicit occupancy / surface function conditioned on the image and camera projection; at inference time it queries dense points to reconstruct a mesh via marching cubes. It also uses a two-stage architecture: a coarse global model followed by local refinement patches to capture fine detail, balancing global consistency and local detail. The repo includes training pipelines, dataset loaders (for Multi-POP, etc.), and inference scripts for mesh output including depth maps for postprocessing. To help practical use, there are utilities for normal estimation, texture back-projection, mesh cleanup, and integration with rendering pipelines.
Features
- High-resolution human mesh reconstruction from a single 2D image
- Pixel-aligned implicit function representation conditioned on image projection
- Two-stage architecture for coarse global and fine local detail modeling
- Training and inference scripts with data loaders for human datasets
- Mesh extraction utilities (marching cubes), normals, texture projection, cleanup
- Support for challenging cases (thin parts, folds, fine geometry)
Project Samples
