Research framework that turns posed images with instance masks and 3D boxes into per-object meshes for cluttered scenes.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
WorldSculpt is a research framework from Alaya Lab and The University of Tokyo that reconstructs cluttered scenes containing hundreds of objects into complete, individually placed object meshes. Its input is a set of RGB images or video frames with instance masks and 3D boxes (grounded video). Its output is a compositional mesh scene, with each object a separate mesh.
It adapts Pixal3D, a single-object generative prior, to take multiple occluded views. Per-view DINOv3 features are lifted into each object's canonical voxel grid and fused by a permutation-invariant, IBR-style aggregator, injected through zero-initialized layers and LoRA. Occluded regions are synthesized while observed regions stay faithful to the views. Because every generated mesh is complete, composing the scene needs only rigid transforms, with no cross-object fusion or per-object post-hoc optimization. The model is trained only on single canonical objects.
The authors also show the pipeline converting 3D Gaussian Splatting worlds generated by World Labs' Marble into compositional meshes. They introduce a benchmark, UE-MeshyScene, with six Unreal Engine 5 environments and exact ground truth. The ground-truth meshes are not released for copyright reasons, and evaluation requires contacting the authors. Code is reported released via a GitHub repository, but the repository URL and license were not in the provided material. The tool is aimed at researchers and 3D pipeline developers. It needs instance masks and 3D boxes as input, not raw images alone.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Turns multiple images or a video into a 3D scene with separate editable objects, and can convert Gaussian splat worlds into mesh scenes. Code released.
Closest neighbours by meaning, not just by category.
Research
Training-free research system that reconstructs simulation-ready 4D mesh scenes from a single monocular video.
Model
Research model that predicts layered, pixel-aligned 3D point clouds, including hidden geometry, from one image or video.
Model · World Labs
World Labs' world model that generates and reconstructs explorable 3D worlds from images, text, video or 3D input.
Model
Feed-forward 3D Gaussian Splatting model that reconstructs scenes from arbitrary input views, with or without camera poses.
Research
Research system that turns a single image into a 360° Gaussian scaffold and renders interactive video at 8 FPS.
Research
Research method that builds controllable, articulated 3D object models from keypoints, point clouds or depth video.
