World Tracing: Generative Multilayer 3D Geometry Representation

Model Overview

World Tracing is a generative 3D geometry representation framework developed by researchers from World Labs and University of Illinois Urbana-Champaign (UIUC) (Hao Zhang, Mohamed El Banani, et al.).

It bridges the gap between monocular depth estimation (high pixel alignment, incomplete geometry) and 3D generative models (complete geometry, poor pixel alignment). Given a single 2D image or short video clip, World Tracing predicts an ordered stack of camera-space 3D points (X, Y, Z) along camera rays for every pixel. The first layer represents visible surfaces, while subsequent layers generate occluded geometry behind them.

---

Key Features

• Multilayer Pixel-Aligned Geometry: Predicts an ordered stack of 3D points per pixel, maintaining 1:1 alignment while capturing occluded surfaces.
• WT-DiT Architecture: Utilizes a Flow-Matching Diffusion Transformer that treats geometry layers as coupled denoising tokens with factorized attention.
• Unified Multi-Domain Support: Provides specialized variants for single objects (WT-O), scenes (WT-S), and dynamic video sequences.
• Zero-Shot Downstream Integration: Enables text-guided 3D scene editing, novel-view video synthesis, and integration with mesh generators.
• Faithfulness & Completeness Balance: Preserves high-precision visible surface depth accuracy while generating complete 3D structure.

---

Verified Project Links

• Project Website: https://haoz19.github.io/world-tracing-page/
• arXiv Paper: https://arxiv.org/abs/2606.13652
• GitHub Repository: https://github.com/haoz19/world-tracing
• Hugging Face Demo: https://huggingface.co/spaces/haoz19/world-tracing-demo

---

Benchmarks & Evaluation

• Objects & 3D-FRONT Benchmarks: Outperforms monocular depth and image-to-3D generative baselines in depth accuracy and occluded geometry reconstruction.