World Tracing: Generative Pixel-Aligned Geometry Beyond the Visible
Researchers have introduced World Tracing, a novel generative representation for 3D geometry that aims to overcome the limitations of existing image-to-3D methods. This approach predicts 3D points aligned with input pixels while also generating geometry for occluded surfaces. The system, instantiated as a diffusion transformer called WT-DiT, treats multiple geometry layers as separate tokens and is trained using pixel-space flow matching. World Tracing demonstrates superior performance in both visible-surface reconstruction and complete geometry generation, enabling applications like text-driven 3D scene editing and novel-view video synthesis. AI
IMPACT This new method for generating 3D geometry could enhance applications in virtual reality, gaming, and content creation by improving the fidelity and completeness of 3D models derived from images.