AIAny
Icon for item

WorldSculpt: Generating Compositional Worlds from Grounded Videos

Generates compositional 3D scenes as collections of individual object meshes by conditioning a single-object 3D generative prior on multi-view posed observations. Key features include anchor-aligned canonical frames, multi-view DINOv3 feature lifting with an IBR-style fusion, and LoRA adaptation to complete heavily occluded objects; includes a large UE-MeshyScene benchmark.

Introduction

Most scene reconstruction systems either fuse everything into a single monolithic surface or leave occluded regions incomplete. WorldSculpt's core insight is that a strong single-object 3D generative prior can be grounded to multi-view, posed observations per object and then reused compositionally to produce large, cluttered scenes of individually addressable meshes—without scene-level retraining.

Key Findings
  • Object-centric generation: Each object is reconstructed as an individual mesh in an anchor-aligned canonical frame, then placed into a shared world frame via a canonical-to-world transform, enabling selection, manipulation, and simulation.
  • Multi-view grounding + generative completion: Per-view DINOv3 features are lifted into canonical voxel volumes and fused with an IBR-style, permutation-invariant aggregator; the aggregated 3D condition is injected into a frozen Pixal3D prior via zero-initialized projection layers while LoRA adapters let the prior exploit multi-view evidence to plausibly complete unobserved geometry.
  • No scene-level training: The generative prior is finetuned only on single-object canonical data yet generalizes at test time to scenes with hundreds of densely occluded objects.
  • Benchmark and evaluation: Introduces UE-MeshyScene, a photorealistic, large-scale benchmark with per-object ground-truth meshes and annotations; WorldSculpt outperforms prior methods, with larger gains as occlusion increases.
Who it's for and tradeoffs

Great fit if you need editable, per-object meshes from multi-view captures (games, AR/VR, simulation, robotics) and want plausible completion in heavily occluded clutter. Look elsewhere if you require exact, physically verified scene-level joint optimization or end-to-end scene learning—WorldSculpt relies on per-object anchoring, posed observations, and a strong single-object prior. It also requires reliable object localization and posed multi-view inputs; quality depends on the prior and the coverage of object views.

Where it fits

Positions between geometry-only multi-view reconstruction (which leaves occluded holes) and monolithic generative worlds (which are not object-separable). It offers a practical path to compositional scene assets by marrying generative priors with grounded multi-view conditioning.

Method highlights
  • Anchor-aligned canonical frames built per detected object
  • Per-view feature extraction (DINOv3) and lifting into canonical voxels
  • Permutation-invariant IBR-style feature fusion supporting variable inputs
  • Zero-init projection layers + LoRA to inject multi-view condition into a pretrained Pixal3D prior
  • Outputs individual meshes that are transformed back into the world frame for composition

This combination yields scalable compositional reconstruction in scenes with hundreds of objects and severe mutual occlusion while preserving direct usability for downstream rendering and simulation.

Information

  • Websitearxiv.org
  • AuthorsMuyao Niu, Jixuan He, Ruihan Yu, Lian Fu, Yonghao Yu, Zheng-Hui Huang, Yifan Zhan, Fengbo Lan, Yongtao Ge, Yinqiang Zheng
  • Published date2026/09/04

More Items

Generates synchronized spoken dialogue and explicit full-body co-speech motion (facial expressions, hands, upper- and lower-body) end-to-end from the same hidden states, replacing the speech-then-motion cascade. Trains with a scalable pseudo-labeling pipeline (422,856 ranked pairs) and supports real-time inference (RTF 0.78) while matching teacher motion metrics within ~2%.

Builds high-fidelity image generators by pairing a 6B Diffusion Transformer with a frozen LLaDA2.0-Mini vision-language module, relying on extensive image-only pretraining and mid-training; model weights, training code, and recipes are released.

Learns generalizable World Action Models for robotic manipulation by scaling causal egocentric video pretraining and grounding learned dynamics with heterogeneous robot trajectories. Key features: a three-stage curriculum (video pretraining, video-action mid-training with a unified action representation, and target-robot specialization) and a Slow–Fast dual-system for 30 Hz real-time action prediction.