A curated and continuously updated list of papers, datasets, and benchmarks for Geometry-Consistent Video World Models (GC-VWM) — generative video models that preserve stable scene geometry across viewpoints, time horizons, and camera trajectories.
Geometry-Consistent Video World Models: A Survey
Zida Yang*, Yihao Lu*, Liyang Wang*, Zeyu Zhang*†, Ling Shao, Hao Tang‡
*Equal contribution. †Project lead. ‡Corresponding author.
This survey will be regularly updated here. If you find this useful, please consider giving us a ⭐!
- Overview
- Taxonomy
- Comprehensive Method Comparison
- Methods by Category
- Datasets
- Benchmarks
- Related Surveys
- Citation
- Star History
Recent generative video models produce photorealistic clips but struggle to maintain consistent spatial structure under large camera motion or extended horizons. GC-VWM addresses this by integrating geometric reasoning — reprojection constraints, depth-aware warping, and spatial memory — into the generative process, enabling interactive exploration of generated worlds through precise camera control, reliable scene revisiting, and out-of-sight dynamics reasoning.
We organize the field into two primary paradigms:
| Paradigm | State Representation | Key Mechanism | Main Trade-off |
|---|---|---|---|
| Without 3D Cache | Latent state + 2D memory | Camera conditioning, geometric PE, RL/SFT alignment | Lightweight but limited geometric guarantees |
| With 3D Cache | Latent + 2D + 3D memory + world state | Explicit point cloud / surfel / 3DGS / hybrid patch | Strong consistency but higher compute & memory |
GC-VWM
├── Without 3D Cache
│ ├── Training-based Alignment (Geometry Forcing, Memory Forcing, LIVE)
│ ├── Camera-control Conditioning (CameraCtrl/II, SEVA, CamI2V, RealCam-I2V, EPiC)
│ ├── Geometry-aware Embeddings (ViewRope, UCPE, ReRoPE)
│ └── Latent Geometry Priors (Context as Memory, Diffusion as Shader)
└── With 3D Cache
├── Point Cloud / Surfel Memory (VMem, WorldMem, EvoWorld, Spatia)
├── Joint 4D Generation (WorldReel, VerseCrafter)
├── 3.5D Geometry Memory (GEN3C, NeoVerse)
├── Hierarchical / Compressed Memory (Infinite-World, RELIC)
├── Unified Camera–Memory (UCM, WorldStereo)
├── Hybrid Patch Memory (MosaicMem)
└── Dynamic 3D Cache / World State (LiveWorld, PERSIST)
44 methods surveyed, covering 2024–2026. Venues listed where confirmed.
| Method | Year | Venue | Paradigm | Geo. Rep. | Temp. Model | Cam. Anno. | Training Strategy |
|---|---|---|---|---|---|---|---|
| MotionCtrl | 2024 | SIGGRAPH | w/o 3D | Camera cond. | Diffusion | Yes | End-to-end |
| DynamiCrafter | 2024 | ECCV | w/o 3D | Image cond. | Diffusion | No | Temporal prior |
| GenWarp | 2024 | NeurIPS | w/o 3D | Latent warp | Diffusion | Yes | Cross-view attn. |
| CameraCtrl | 2024 | ICLR'25 | w/o 3D | Pose cond. | Diffusion | Yes | Plug-and-play |
| CamI2V | 2024 | arXiv | w/o 3D | Epipolar attn. | Diffusion | Yes | Fine-tuning |
| CameraCtrl-II | 2025 | ICCV | w/o 3D | Pose cond. | Diffusion | Yes | Plug-and-play |
| SEVA | 2025 | ICCV | w/o 3D | Implicit | Diffusion | Yes | End-to-end |
| RealCam-I2V | 2025 | ICCV | w/o 3D | Metric depth | Diffusion | Yes | Metric-scale align. |
| EPiC | 2025 | arXiv | w/o 3D | Anchor video | Diffusion | No | Anchor-ControlNet |
| Geo. Forcing | 2025 | ICLR'26 | w/o 3D | Latent align. | Diffusion | Yes | Geo. feature align. |
| Mem. Forcing | 2025 | arXiv | w/o 3D | Latent memory | Diffusion | Yes | Memory supervision |
| DaS | 2025 | SIGGRAPH | w/o 3D | 3D tracking | Diffusion | Yes | Multi-task unified |
| Context as Mem. | 2025 | SIG Asia | w/o 3D | Latent retrieval | Diffusion | Partial | Memory retrieval |
| PanoWAN | 2025 | arXiv | w/o 3D | Spherical | Diffusion | Yes | Pano. alignment |
| CamCloneMaster | 2025 | SIG Asia | w/o 3D | Ref. video | Diffusion | No | Motion transfer |
| WorldPack | 2025 | arXiv | w/o 3D | Traj. memory | Diffusion | Yes | Memory comp. |
| ViewRope | 2026 | arXiv | w/o 3D | Ray-based PE | Transformer | Yes | PE injection |
| UCPE | 2025 | arXiv | w/o 3D | Ray encoding | Diffusion | Yes | PE injection |
| LIVE | 2026 | arXiv | w/o 3D | Cycle cons. | Diffusion | No | Cycle-consistency |
| AnchorWeave | 2026 | arXiv | w/o 3D | Local mem. | Diffusion | Yes | Anchor retrieval |
| ViewCrafter | 2024 | arXiv | w/ 3D | Point cloud | Diffusion | Yes | View cond. |
| GEN3C | 2025 | CVPR | w/ 3D | Depth→PC | Diffusion | Yes | 3D cache cond. |
| VMem | 2025 | ICCV | w/ 3D | Surfel memory | AR+retrieval | Yes | Plug-and-play mem. |
| WorldMem | 2025 | NeurIPS'25 | w/ 3D | State memory | Diffusion | Yes | State-aware attn. |
| EvoWorld | 2025 | arXiv | w/ 3D | Point cloud | Diffusion | Yes | 3D reproj. guide |
| Spatia | 2025 | CVPR'26 | w/ 3D | SLAM PC | Diffusion | Yes | SLAM update loop |
| WorldReel | 2025 | arXiv | w/ 3D | Pointmap+flow | Diffusion | Yes | Joint 4D gen. |
| WVD | 2025 | CVPR | w/ 3D | XYZ images | Diffusion | Yes | Co-generation |
| Geo4D | 2025 | ICCV | w/ 3D | Point/Ray map | Diffusion | Yes | Multi-modal align. |
| UniFuture | 2025 | arXiv | w/ 3D | Hier. scene | Diffusion | Yes | Hierarchical |
| TrajectoryCrafter | 2025 | ICCV | w/ 3D | Temp. PC | Diffusion | Yes | PC-guided |
| DeepVerse | 2025 | arXiv | w/ 3D | 4D Auto-reg. | AR | Yes | Geo. conditioning |
| Aether | 2025 | arXiv | w/ 3D | 4D Recon. | Diffusion | Yes | Action-conditioned |
| VerseCrafter | 2026 | arXiv | w/ 3D | 3DGS traj. | Diffusion | Yes | 4D control render. |
| NeoVerse | 2026 | arXiv | w/ 3D | 4D world | Recon. | No | Pose-free 4D |
| ReCapture | 2025 | CVPR | w/ 3D | Depth recon. | Diffusion | Yes | Masked fine-tuning |
| ReCamMaster | 2025 | ICCV | w/ 3D | Video cond. | Diffusion | Yes | Video conditioning |
| RELIC | 2025 | arXiv | w/ 3D | Compressed KV | AR diffusion | Yes | Self-forcing distill. |
| Infinite-World | 2026 | arXiv | w/ 3D | Hier. memory | AR diffusion | No | Pose-free HPMC |
| LiveWorld | 2026 | arXiv | w/ 3D | BG+entities | State evol. | Yes | Monitor mechanism |
| PERSIST | 2026 | arXiv | w/ 3D | Latent 3D scene | Scene evol. | Yes | 3D scene sim. |
| UCM | 2026 | arXiv | w/ 3D | PE warping | Dual-stream DiT | Yes | Time-aware PE warp |
| WorldStereo | 2026 | arXiv | w/ 3D | Dual geo. mem. | Diffusion | Yes | Global+spatial mem. |
| MosaicMem | 2026 | arXiv | w/ 3D | Hybrid patch 3D | Diffusion | Yes | Patch lift+compose |
- Geometry Forcing — Aligning Geometric Representations for Video Diffusion Models (ICLR 2026) [paper]
- Memory Forcing — Memory Forcing for Geometry-Consistent Video World Models [paper]
- LIVE — Long-horizon Video Generation via Cycle-Consistency Objective [paper]
- Self Forcing — Self-Forcing: Bridging the Train-Test Gap in Autoregressive Video Generation [paper]
- Rolling Forcing — Rolling Diffusion for Real-Time Long Video Generation [paper]
- Causal Forcing — Causal Autoregressive Diffusion Distillation [paper]
- CameraCtrl — Camera Motion Conditioned Video Diffusion (ICLR 2025) [paper] [code]
- CameraCtrl-II — Dynamic Scene Exploration (ICCV 2025) [paper] [project]
- CamI2V — Epipolar-Geometry-Aware I2V [paper]
- SEVA — Stable Virtual Camera (ICCV 2025) [paper]
- RealCam-I2V — Metric-Scale Camera Control (ICCV 2025) [paper]
- EPiC — Efficient Camera-Control Learning with Anchor Videos [paper]
- MotionCtrl — Unified Camera and Object Motion Control (SIGGRAPH 2024) [paper]
- CamCloneMaster — Reference-Video-Driven Camera Control (SIG Asia 2025) [paper]
- ViewRope — Geometry-Aware Rotary Position Encoding [paper]
- UCPE — Unified Camera Positional Encoding [paper]
- ReRoPE — Repurposing Rotary Embeddings for Camera Pose [paper]
- Context as Memory — Memory Retrieval for Spatial Coherence (SIG Asia 2025) [paper]
- Diffusion as Shader — 3D-Tracking-Guided Video Diffusion (SIGGRAPH 2025) [paper]
- Epipolar-DPO — Epipolar-Geometry Reward for Diffusion [paper]
- VGGRPO — Latent-Space Geometry Alignment via GRPO [paper]
- VMem — Surfel-Indexed View Memory (ICCV 2025 Highlight) [paper] [code] [project]
- WorldMem — Long-term Consistent World Simulation with Memory (NeurIPS 2025) [paper]
- EvoWorld — Evolving 3D Memory for Panoramic Video Generation [paper]
- Spatia — SLAM-Updated Point Cloud Memory (CVPR 2026) [paper]
- Voyager — World-Consistent 3D Point-Cloud Generation [paper]
- WorldReel — Joint 4D RGB + Geometry Generation [paper]
- VerseCrafter — 4D Geometric Control via 3DGS Trajectories [paper]
- WVD — World Video Diffusion with Co-Generated XYZ (CVPR 2025) [paper]
- Geo4D — Monocular 4D Reconstruction via Video Diffusion (ICCV 2025) [paper]
- GeometryCrafter — Multi-Modal Geometry from Video Priors [paper]
- TesserAct — Joint RGB-DN 4D World Representation [paper]
- GEN3C — 3D-Informed Video Generation (CVPR 2025 Highlight) [paper] [code] [project]
- NeoVerse — Pose-Free 4D World Models from Monocular Video [paper]
- TrajectoryCrafter — Camera Trajectory Guided Video Generation (ICCV 2025) [paper]
- Infinite-World — Pose-Free Hierarchical Memory Compressor [paper]
- RELIC — Compressed KV Memory for Real-Time Generation [paper]
- MemRoPE — Dual-Stream Memory Compression for Infinite Video [paper]
- UCM — Unified Camera-Memory via Time-Aware PE Warping [paper]
- WorldStereo — Dual Geometric Memories for Camera-Guided Generation [paper]
- MosaicMem — Lifted 3D Patch Memory for Diffusion-Native Conditioning [paper]
- LiveWorld — Out-of-Sight Dynamics with Monitor-Based Evolution [paper]
- PERSIST — Persistent Latent 3D Scene Simulation [paper]
- DeepVerse — 4D Autoregressive World Model [paper]
| Dataset | Type | Scale | Pose | Depth | Metric | Primary Usage |
|---|---|---|---|---|---|---|
| RealEstate10K | Posed video | ~80K clips | ✓ | — | — | Camera-controlled generation, NVS |
| ACID | Aerial video | Thousands | ✓ | — | — | Outdoor camera control |
| RealCam-Vid | Posed video | ~105K | ✓ | — | ✓ | Metric-scale camera control |
| Princeton365 | Posed video | 365 videos | ✓ | — | ✓ | High-precision pose evaluation |
| MultiCamVideo | Synthetic multi-cam | 136K videos | ✓ | ✓ | ✓ | Multi-view consistency |
| Virtual KITTI 2 | Synthetic driving | 5 sequences | ✓ | ✓ | ✓ | Domain transfer |
| KITTI | Driving | 22 sequences | ✓ | ✓ | ✓ | Trajectory estimation |
| Waymo Open | Driving | 2,030 segments | ✓ | ✓ | ✓ | Driving world models |
| Argoverse 2 | Driving | 1,000 sequences | ✓ | ✓ | ✓ | Driving generation |
| CO3D | Object-centric | 18,619 videos | ✓ | ✓ | — | Object-level NVS |
| Ego4D | Egocentric | ~3.7K hrs | Partial | Partial | — | Long-horizon ego tasks |
| Benchmark | Year | Target | Metrics | Focus |
|---|---|---|---|---|
| Tanks & Temples | 2017 | 3D Recon. | Precision, Recall, F-score | Reconstruction accuracy |
| VBench | 2024 | Video Gen. | 16 dims: subject/bg consist., temporal | Disentangled video quality |
| WorldModelBench | 2025 | Video Gen. | Instruction, common sense, physics | Instruction & physics adherence |
| VBench-2.0 | 2025 | Video Gen. | 18 dims: 3D spatial, camera, physics | Prompt faithfulness |
| WorldScore | 2025 | 3D/4D Gen. | Ctrl., quality, dynamics | Unified 3D/4D/video eval |
| ViewBench | 2026 | Video Gen. | PSNR, SSIM, LPIPS, LCE | View consistency & loop-closure |
| StEvo-Bench | 2026 | Video Gen. | State progress, physics | State evolution ≠ observation |
| LiveBench | 2026 | Video Gen. | PSNR, SSIM, VQA-Acc | OOS dynamics & identity |
| Omni-WorldBench | 2026 | Video Gen. | AgenticScore | Interactive camera & loop |
| Survey | Year | Scope |
|---|---|---|
| Tewari et al. — Advances in Neural Rendering | 2022 | Neural Rendering |
| Wang et al. — World Models for Autonomous Driving | 2024 | General WM |
| Ding et al. — Understanding World Models | 2024 | General WM |
| Kong et al. — World Modeling for 3D/4D | 2025 | 3D/4D WM |
| Cho et al. — Simulating the Visual World | 2025 | Visual AI |
| Ours | 2026 | GC-VWM |
If you find this repository or our survey useful, please consider citing:
@article{yang2026gcvwm,
title={Geometry-Consistent Video World Models: A Survey},
author={Yang, Zida and Lu, Yihao and Wang, Liyang and Zhang, Zeyu and Tang, Hao},
journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
year={2026},
note={Under Review}
}We welcome contributions! Please feel free to submit a Pull Request to add new papers, correct errors, or suggest improvements.
Maintained by: Zida Yang and co-authors.
Last updated: April 2026