Hao Zhang, Mohamed El Banani, Jen-Hao Cheng, Paul Zhang, Yi Hua, Ben Mildenhall, Christoph Lassner, Narendra Ahuja et al.
World Tracing is a generative pixel-aligned geometry representation that predicts an ordered stack of 3D points for each input pixel, including both the visible surface and occluded backside.
Existing image-to-3D methods trade off faithfulness and completeness: depth estimators align with input pixels but only predict visible surfaces, while image-to-3D generative models produce complete shapes but often misalign with the input.
World Tracing predicts an ordered stack of camera-space 3D points for each input pixel. The first layer represents the visible surface, and subsequent layers represent front-to-back intersections with occluded surfaces. We instantiate this with WT-DiT (World-Tracing Diffusion Transformer), which treats multiple geometry layers as separate denoising tokens coupled through factorized and global attention. It is trained with pixel-space flow matching and a mixed noise schedule that balances visible-surface reconstruction with occluded-geometry generation.
World Tracing achieves strong performance on visible-surface reconstruction and complete geometry generation across object, scene, and dynamic benchmarks, outperforming both depth predictors and image-to-3D generators. It also preserves 2D-to-3D correspondence, enabling text-driven 3D scene editing, geometry-conditioned novel-view video synthesis, and training-free integration with textured-mesh generators.