← World Models

Occupancy Generation

3D and 4D occupancy grids that encode geometry and semantics in voxel space — the substrate for scene generation, forecasting, and world simulation.

⌘K

Occupancy world models represent scenes as 3D or 4D voxel grids encoding geometry and semantics, enabling controllable scene generation, future occupancy forecasting, and autoregressive simulation for autonomous driving.

Scene Representors — 3D Occupancy Scene Generation 10+ papers

Model/MethodFull TitleVenueYear
SSDDiffusion Probabilistic Models for Scene-Scale 3D Categorical Data arXiv2023
SemCitySemantic Scene Generation with Triplane Diffusion CVPR2024
WoVoGenWorld Volume-Aware Diffusion for Controllable Multi-Camera Scene Generation ECCV2024
UrbanDiffUrban Scene Diffusion through Semantic Occupancy Map arXiv2024
DrivingSphereBuilding A High-Fidelity 4D World for Closed-Loop Simulation CVPR2025
UniSceneUnified Occupancy-Centric Driving Scene Generation CVPR2025
InfiniCubeUnbounded and Controllable Dynamic 3D Driving Scene Generation ICCV2025
Control-3D-SceneControllable 3D Outdoor Scene Generation via Scene Graphs ICCV2025
X-SceneLarge-Scale Driving Scene Generation with High Fidelity arXiv2025

Occupancy Forecasters — 4D Occupancy Prediction 26+ papers

Model/MethodFull TitleVenueYear
ViGTVisual Implicit Geometry Transformer for Autonomous Driving arXiv2026
Emergent-OccDifferentiable Raycasting for Self-supervised Occupancy Forecasting ECCV2022
FF4DPoint Cloud Forecasting as a Proxy for 4D Occupancy Forecasting CVPR2023
OccWorldLearning A 3D Occupancy World Model for Autonomous Driving ECCV2024
Cam4DOccBenchmark for Camera-Only 4D Occupancy Forecasting CVPR2024
DriveWorld4D Pre-Trained Scene Understanding via World Models for AD CVPR2024
OccSora4D Occupancy Generation Models as World Simulators for AD arXiv2024
UnOUnsupervised Occupancy Fields for Perception and Forecasting CVPR2024
OccLLaMAAn Occupancy-Language-Action Generative World Model for AD arXiv2024
DOMETaming Diffusion Model into High-Fidelity Controllable Occupancy World Model arXiv2024
GaussianADGaussian-Centric End-to-End Autonomous Driving arXiv2024
Drive-OccWorldVision-Centric 4D Occupancy Forecasting and Planning via World Models AAAI2025
PreWorldSemi-Supervised Vision-Centric 3D Occupancy World Model ICLR2025
OccProphetPushing Efficiency Frontier of Camera-Only 4D Occupancy Forecasting ICLR2025
RenderWorldWorld Model with Self-Supervised 3D Label ICRA2025
EfficientOCFSpatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting CVPR2025
DIODecomposable Implicit 4D Occupancy-Flow World Model CVPR2025
UniOccA Unified Benchmark for Occupancy Forecasting and Prediction ICCV2025
I²WorldIntra-Inter Tokenization for Efficient Dynamic 4D Scene Forecasting ICCV2025
T³FormerTemporal Triplane Transformers as Occupancy World Models arXiv2025
COMEAdding Scene-Centric Forecasting Control to Occupancy World Model arXiv2025

Autoregressive Simulators — Generative 4D Occupancy 9+ papers

Model/MethodFull TitleVenueYear
XCubeLarge-Scale 3D Generative Modeling using Sparse Voxel Hierarchies CVPR2024
PDDPyramid Diffusion for Fine 3D Large Scene Generation ECCV2024
OccSora4D Occupancy Generation Models as World Simulators arXiv2024
DynamicCityLarge-Scale 4D Occupancy Generation from Dynamic Scenes ICLR2025
DrivingSphereBuilding A High-Fidelity 4D World for Closed-Loop Simulation CVPR2025
InfiniCubeUnbounded Dynamic 3D Driving Scene Generation ICCV2025
X-SceneLarge-Scale Driving Scene Generation with High Fidelity arXiv2025
PrITTIPrimitive-Based Generation of Controllable and Editable 3D Semantic Scenes arXiv2025