Three works by World Labs on 3D generated from an image

World Labs unveils World Tracing, Modality Forcing and Flex4DHuman, three new AI models generating 3D scenes, depth and 4D humans from single 2D inputs.

The Spatial Intelligence Laboratory highlights three research projects conducted by its interns, all centered around a single idea: generating 3D by leveraging large generative models and the 2D priors they have learned.

World Tracing reconstructs a complete 3D scene from a single image; for each pixel, a diffusion model stacks an ordered series of points, from the visible surface to the hidden geometry behind it, which reconstructs occluded areas without deviating from the image. Its authors claim surface accuracy on par with the best monocular depth estimators, and superior geometric fidelity to image-to-3D and video-to-4D generators.

Modality Forcing adapts text-to-image models to jointly reason about text, image, and depth: a single DiT backbone, trained with a distinct noise level per modality, switches between depth estimation, generation from depth, and joint RGB-D generation; when applied to FLUX.2-klein-9B, it would match dedicated depth models like MoGe-2 or PPD.

Flex4DHuman, finally, transforms a monocular video, even a simple phone capture, into a dynamic 4D human: a finetuned video diffusion model produces synchronized multiple views, distilled into 4D Gaussian splats and re-integrable into 3D worlds, with intended uses in AR/VR and gaming.

Each project is accompanied by a public paper, page, and demo.