PuzzleAvatar reconstructs textured 3D clothed humans from unconstrained personal photo collections. It targets researchers and developers in computer vision and graphics, enabling avatar creation without requiring specific poses, camera calibration, or predefined body mesh templates.

How It Works

The approach leverages a novel pipeline that bypasses traditional constraints like human pose estimation, camera pose, or explicit geometric cues. It integrates components like Grounded-SAM, PuzzleBooth, and SDS, allowing for the assembly of avatars from arbitrary images, offering flexibility and robustness against common reconstruction challenges.

Quick Start & Requirements

• Install: Set paths in scripts/env.sh and follow installation instructions.
• Run: bash scripts/run.sh <data_path> <results_path> [optional_name] (e.g., bash scripts/run.sh data/human/yuliang results/human/yuliang human_yuliang).
• Prerequisites: Requires specific package setup. The process takes approximately 4 hours.
• Dataset: Download 194GB dataset via bash scripts/fetch_data.sh.
• Links: PuzzleAvatar Website, ICON PRT Fix

Highlighted Details

• Reconstructs avatars from unconstrained photo collections.
• Does not require human body pose, camera pose, or geometric cues.
• Supports any A-posed body mesh template, avoiding reliance on SMPL or similar models.
• Integrates with Grounded-SAM, GPT-4V(ision), and Stable Diffusion.

Maintenance & Community

The project is associated with SIGGRAPH Asia 2024 and has released code and datasets. It acknowledges contributions from various foundational projects and welcomes community input.

Licensing & Compatibility

The code and models are available for non-commercial scientific research purposes. Commercial licensing inquiries should be directed to ps-licensing@tue.mpg.de.

Limitations & Caveats

The rendering process may get stuck, requiring specific changes for PRT computation as noted in the ICON repository. The project is primarily built upon TeCH, which should also be cited.