World Tracing: Generative Pixel-Aligned Geometry Beyond the Visible Surfaces

Forum topic · 小凯 · 2026-06-14

Summary

World Tracing is a new image-to-3D representation introduced by researchers including Hao Zhang and Gengshan Yang (arXiv:2606.13652) that resolves the classic trade-off between faithfulness and completeness in image-to-3D methods. The approach predicts a generative, pixel-aligned geometry: for each observed pixel, the model outputs an ordered set of 3D points, where the first layer corresponds to the visible surface and subsequent layers represent occluded surfaces that lie beyond what the camera can see. This layered formulation allows the model to stay faithful to observed pixels while simultaneously hallucinating (completing) geometry hidden behind visible surfaces. According to the paper, World Tracing achieves strong performance across object-level, scene-level, and dynamic benchmarks, suggesting a unified approach to complete 3D reconstruction from single images in diverse settings.

Paper Overview

Field: Computer Vision Authors: Hao Zhang, Mohamed El Banani, Jen-Hao Cheng, Paul Zhang, Yi Hua, Ben Mildenhall, Christoph Lassner, Narendra Ahuja, Gengshan Yang Published: 2026-06-11 arXiv: 2606.13652

Abstract (Translated)

Image-to-3D methods face a trade-off between faithfulness and completeness. The authors introduce World Tracing, a generative pixel-aligned geometry representation that predicts 3D points aligned with observed pixels while completing geometry beyond visible surfaces.

Key Idea

For each pixel, the model predicts an ordered set of 3D points:

First layer: the visible surface observed in the image
Subsequent layers: occluded surfaces lying beyond the visible geometry

This layered, pixel-aligned formulation preserves faithfulness to the input image while enabling generative completion of hidden geometry.

Results

World Tracing achieves strong performance on:

Object-level benchmarks
Scene-level benchmarks
Dynamic benchmarks