World Tracing: Generative Pixel-Aligned Geometry Beyond the Visible

World Tracing, developed by World Labs and the University of Illinois Urbana-Champaign, reframes image-to-3D as a pixel-aligned multilayer geometry problem: for each input pixel, the model predicts a stack of 6 camera-space XYZ intersections, with layer 0 reconstructing the visible surface and deeper layers completing occluded geometry beyond the visible. The system is instantiated as WT-DiT, a 1.7B-parameter flow-matching diffusion transformer built on a frozen MoGe ViT-L encoder, trained with an XYZ-only depth-filling objective, three-way attention factorization over layer-wise, ray-wise, and global tokens, and a mixed diffusion-time schedule that balances reconstruction and generative completion. On held-out object data, WT-O reaches 0.0149 MAE and 0.0243 RMSE on visible-surface depth while also improving complete geometry to 0.0213 L1, 0.00194 L2, and 0.898 F@0.05; on held-out 3D-FRONT scenes, WT-S posts 0.0102 MAE, 0.0215 RMSE, and 0.9867 F@0.05 at layer 0, plus 0.0216 All-L CD-L1 and 0.8951 All-L F@0.05 for full multilayer geometry. The model also extends to dynamics, achieving a 0.0105 mean CD-L2 across Obj.-Val, Truebone, and ActionBench, and its pixel-aligned representation enables training-free pose-aware textured-mesh generation, text-driven scene editing, and geometry-conditioned novel-view video synthesis. A mix-training ablation on real depth corpora further lifts WT-S to 0.0374 AbsRel on NYU and 0.0332 AbsRel on ETH3D-indoor, showing that the same multilayer architecture can absorb single-layer RGBD supervision without sacrificing occluded-geometry prediction.