Unbounded 3D world generation is emerging as a foundational task for scene modeling in computer vision, graphics, and robotics. In this work, we present WorldFlow3D, a novel method capable of generating unbounded 3D worlds. Building upon a foundational property of flow matching - namely, defining a path of transport between two data distributions - we model 3D generation more generally as a problem of flowing through 3D data distributions, not limited to conditional denoising. We find that our latent-free flow approach generates causal and accurate 3D structure, and can use this as an intermediate distribution to guide the generation of more complex structure and high-quality texture - all while converging more rapidly than existing methods. We enable controllability over generated scenes with vectorized scene layout conditions for geometric structure control and visual texture control through scene attributes. We confirm the effectiveness of WorldFlow3D on both real outdoor driving scenes and synthetic indoor scenes, validating cross-domain generalizability and high-quality generation on real data distributions. We confirm favorable scene generation fidelity over approaches in all tested settings for unbounded scene generation. For more, see https://light.princeton.edu/worldflow3d.
Figures & tables
XCube [ 31 ]
BlockFusion [ 42 ]
LidarDM [ 46 ]
LT3SD [ 24 ]
InfiniCube [ 23 ]
WorldGrow [ 18 ]
XScene [ 44 ]
Ours
Unbounded
✗
✓
✗
✓
✓
✓
✓
✓
Controllable
✗
✗
✗
✗
✓
✗
✗
✓
Cross-Domain
✓
✓
✗
✗
✗
✗
✗
✓
Texture
✗
✗
✗
✗
✓
✓
✓
✓
Table 1: Summary of recent 3D scene generation methods. Ours is the only approach satisfying all desirable criteria.
Figure 2: WorldFlow3D decomposes generation into a sequence of independent flows over progressively richer representations — transporting from noise, through coarse geometry into fine geometry, and visual appearance ( Sec. 3.1 ) . All flows operate directly in raw volumetric space ( Sec. 3.2 ), enabling a latent-free, hierarchical scene generation procedure. Generation is controlled by a vectorized geometric layout and discrete scene attributes, giving consistent structural and semantic control at every level ( Sec. 3.4 ).
Figure 3: Feather weighted velocity averaging in overlapping chunk regions significantly improves the generated geometry for unbounded generations, as shown above.
Figure 4: Qualitative comparison on outdoor scene generation with WorldFlow3D and baseline methods trained on the Waymo [ 37 ] dataset. We showcase scenes generated at moderate scales, and closer-up views of specific details including buildings and vehicles. We obtain high-quality, realistic geometry and smooth surfaces with a good amount of detail, as viewed from coherent building structure, smooth road surfaces, and distinct vehicle geometry.
Figure 5: Qualitative comparison on indoor scene generation with WorldFlow3D and baseline methods trained on the 3D-FRONT [ 12 ] dataset. We showcase generations of regions including (potentially multiple) rooms with various objects. Our generations are high-fidelity and contain smooth surfaces and realistic geometry.
Method
COV ↑
MMD ↓
1-NNA ↓
JSD ↓
FDC↓
CD
EMD
CD
EMD
CD
EMD
Waymo Uncond.
XCube [ 31 ]
27.50
24.90
19.75
2.99
95.85
85.35
0.480
214.08
LidarDM [ 46 ]
15.30
15.40
29.74
3.60
99.40
98.10
0.564
232.49
LT3SD [ 24 ]
20.00
16.10
34.33
3.96
95.85
85.35
0.524
76.04
WorldFlow3D
33.00
32.70
16.57
2.81
89.15
70.15
0.490
74.83
Waymo Cond.
LidarDM [ 46 ]
12.00
11.30
35.46
3.81
99.60
99.20
0.590
215.49
Table 2: Quantitative Evaluation for Outdoor 3D Scene Generation on the Waymo Open Dataset [ 37 ] for WorldFlow3D and existing approaches. We show results for unconditional generation in the first section and conditional in the second. The best results for each metric are in bold ; second-best are underlined . We evaluate distribution coverage and alignment (COV, JSD), generation fidelity (MMD and 1-NNA), and feature-based distributional distance ( FDC ). We report metrics on large-scale scene sizes of 96m×96m .
Method
COV ↑
MMD ↓
1-NNA ↓
JSD ↓
FDC↓
CD
EMD
CD
EMD
CD
EMD
3D-Front [ 12 ] Uncond.
BlockFusion [ 42 ]
29.00
28.40
0.054
0.221
91.50
90.65
0.380
165.46
LT3SD [ 24 ]
23.40
26.20
0.056
0.223
93.00
91.60
0.230
44.97
WorldGrow [ 18 ]
34.70
32.80
0.053
0.221
88.05
88.65
0.272
172.58
WorldFlow3D
38.30
38.10
0.039
0.195
74.75
75.80
0.164
36.45
Table 3: Quantitative Evaluation for Indoor 3D Scene Generation on the synthetic 3D-FRONT [ 12 ] dataset. The best results for each metric are in bold ; second-best are underlined . We report metrics on small-scale scenes [ 24 , 42 ] of 2m×2m .
Ablation
COV ↑
MMD ↓
1-NNA ↓
FDC↓
GPU-hrs ↓
CD
EMD
CD
EMD
CD
EMD
Latent Diffusion
37.75
33.75
7.426
1.801
91.38
75.63
120.87
288
Latent Flow
37.00
32.75
6.755
1.778
90.88
74.88
112.35
288
Flow from Noise
33.75
33.75
6.887
1.747
92.25
81.25
107.59
144
WorldFlow3D ( p(f)=p(1) )
38.00
31.50
7.101
1.821
92.75
75.63
89.47
72
WorldFlow3D(full)
42.50
37.25
4.942
1.696
85.00
70.25
103.20
144
Table 4: Ablation Study over the core contributions of our method, comparing distribution coverage (COV), geometric fidelity (MMD and 1-NNA), visual texture quality ( FDC) , and training efficiency (GPU-hrs). We compare against traditional latent diffusion and latent flow approaches, followed by an ablation of our flow matching through distributions objective, and finally of our flow sequence hierarchy. We conduct our evaluation on the Waymo Open Dataset [ 37 ] , and the best results are bolded .
Figure 6: Ablation of Core Components. We provide qualitative results obtained by ablating the key components of our method. Latent diffusion and flow approaches produce structurally degenerate, non-realistic results, while flow from noise at finer distributions produces noisy outputs. Examples of this can be seen on the building walls, which are smooth with WorldFlow3D, and on vehicle details such as tires. Our hierarchical, latent-free approach is the only one that can produce high-quality, geometrically plausible results.
Figure 7: Visual Texture and Controllability. We report large-scale outdoor scenes with texture control via text attributes and geometry control via vector maps, yielding results conditioned on the same geometry to produce distinct environments.
Method
Rank
BT Score
95% CI
Win Rate
WorldFlow3D
1
0.692
[0.569, 0.813]
88%
XCube [ 31 ]
2
0.212
[0.118, 0.314]
63%
LT3SD [ 24 ]
3
0.085
[0.047, 0.131]
43%
LidarDM [ 46 ]
4
0.011
[0.004, 0.022]
6%
Table 5: User Study. We report Bradley-Terry (BT) [ 3 ] scores with 95% bootstrap confidence intervals and overall win rates.
Current 3D generation methods lack explicit local control: geometric adherence is often defined by a global control strength, and appearance cannot be specified locally. We present SpaceFlow, a training-free pipeline for locally controllable 3D generation from text descriptions and a collection of geometric primitives. Each primitive serves as a proxy for an object part and is assigned a local control level, enabling users to specify whether regions should strictly follow the input shape or allow generative completion. During structure generation, we enforce these spatial constraints within the generative flow process. For appearance synthesis, the generated structure is segmented and matched to the primitives. Each generated part is conditioned only on its assigned text or image cue, thereby limiting cross-part leakage. Regional geometry metrics demonstrate that SpaceFlow preserves the specified geometry in high-control regions and enables plausible shape variation in low-control areas. A user study further indicates that the resulting balance between geometric fidelity and generative freedom remains competitive in overall quality. When evaluating appearance on fixed geometry, text-conditioned routing achieves state-of-the-art prompt faithfulness and color/material accuracy. Qualitative results additionally show localized routing of image cues. The project page is available at SpaceFlow3D.github.io.
Neil De La Fuente, Joan Lafuente, Mukhammadali Sayfiddinov +5
Existing diffusion-based 3D scene generation methods primarily operate in 2D image/video latent spaces, which makes maintaining cross-view appearance and geometric consistency inherently challenging. To bridge this gap, we present OneWorld, a framework that performs diffusion directly within a coherent 3D representation space. Central to our approach is the 3D Unified Representation Autoencoder (3D-URAE); it leverages pretrained 3D foundation models and augments their geometry-centric nature by injecting appearance and distilling semantics into a unified 3D latent space. Furthermore, we introduce token-level Cross-View-Correspondence (CVC) consistency loss to explicitly enforce structural alignment across views, and propose Manifold-Drift Forcing (MDF) to mitigate train-inference exposure bias and shape a robust 3D manifold by mixing drifted and original representations. Comprehensive experiments demonstrate that OneWorld generates high-quality 3D scenes with superior cross-view consistency compared to state-of-the-art 2D-based methods. Our code will be available at https://github.com/SensenGao/OneWorld.
Sensen Gao, Zhaoqing Wang, Qihang Cao +5
Mohamed bin Zayed University of Artificial Intelligence · AISphere · Shanghai Jiao Tong University +3
Recent advances in video generation enable a new paradigm for 3D scene creation: generating camera-controlled videos that simulate scene walkthroughs, then lifting them to 3D via feed-forward reconstruction techniques. This generative reconstruction approach combines the visual fidelity and creative capacity of video models with 3D outputs ready for real-time rendering and simulation. Scaling to large, complex environments requires 3D-consistent video generation over long camera trajectories with large viewpoint changes and location revisits, a setting where current video models degrade quickly. Existing methods for long-horizon generation are fundamentally limited by two forms of degradation: spatial forgetting and temporal drifting. As exploration proceeds, previously observed regions fall outside the model's temporal context, forcing the model to hallucinate structures when revisited. Meanwhile, autoregressive generation accumulates small synthesis errors over time, gradually distorting scene appearance and geometry. We present Lyra 2.0, a framework for generating persistent, explorable 3D worlds at scale. To address spatial forgetting, we maintain per-frame 3D geometry and use it solely for information routing -- retrieving relevant past frames and establishing dense correspondences with the target viewpoints -- while relying on the generative prior for appearance synthesis. To address temporal drifting, we train with self-augmented histories that expose the model to its own degraded outputs, teaching it to correct drift rather than propagate it. Together, these enable substantially longer and 3D-consistent video trajectories, which we leverage to fine-tune feed-forward reconstruction models that reliably recover high-quality 3D scenes.