Interactive demo · ZeroGPU

World Tracing

Generative pixel-aligned geometry beyond the visible — a single image becomes a layered 3D world.

Hao Zhang · Mohamed El Banani · Jen-Hao Cheng · Paul Zhang · Yi Hua · Ben Mildenhall · Christoph Lassner · Narendra Ahuja · Gengshan Yang

World Labs · University of Illinois Urbana-Champaign

Upload a single image (Object / Scene tabs) or a short video clip (Dynamic tab) and click Generate. The model produces 6 pixel-aligned XYZ layers (a stack of camera-space points per pixel, not a depth map) — the visible surface plus the (partially) occluded geometry behind it. Layer 0 keeps the input colours; deeper layers fade to grey.

Step 1 · Choose a model

Which input do you have?

Pick the model that matches your input — the upload appears once you choose. You can switch back any time.

Object

A single object photo → layered per-pixel point cloud.

Best for: products, single items, characters, toys.

Scene

A scene photo → layered per-pixel point cloud at 840 px resolution.

Best for: indoor & outdoor scenes. Remove the sky for outdoor.

Dynamic

A short video clip → per-frame layered per-pixel point cloud.

Best for: one moving subject, clips ≤ 10 s.

01+

Object · Pro 🔒

Same object workflow, with a deeper 10-layer back-side reconstruction (r75g_10layer).

Not yet open: access to these weights will be opened up in a while. Sign in with Hugging Face to check status.

02+

Scene · Pro 🔒

Same scene workflow, with mix training with RGBD data — real RGBD captures mixed into training for better real-world robustness.

Not yet open: access to these weights will be opened up in a while. Sign in with Hugging Face to check status.

Textured Mesh

Image → textured GLB mesh (r75b → TRELLIS.2).

Coming soon on ZeroGPU — preview the roadmap.