What if a world simulation model could render not an imagined environment but a city that actually exists? Prior generative world models synthesize visually plausible yet artificial environments by imagining all content. We present Seoul World Model (SWM), a city-scale world model grounded in the real city of Seoul. SWM anchors autoregressive video generation through retrieval-augmented conditioning on nearby street-view images. However, this design introduces several challenges, including temporal misalignment between retrieved references and the dynamic target scene, limited trajectory diversity and data sparsity from vehicle-mounted captures at sparse intervals. We address these challenges through cross-temporal pairing, a large-scale synthetic dataset enabling diverse camera trajectories, and a view interpolation pipeline that synthesizes coherent training videos from sparse street-view images. We further introduce a Virtual Lookahead Sink to stabilize long-horizon generation by continuously re-grounding each chunk to a retrieved image at a future location. We evaluate SWM against recent video world models across three cities: Seoul, Busan, and Ann Arbor. SWM outperforms existing methods in generating spatially faithful, temporally consistent, long-horizon videos grounded in actual urban environments over trajectories reaching hundreds of meters, while supporting diverse camera movements and text-prompted scenario variations.
Figures & tables
Figure 1: Seoul World Model (SWM) generates videos over a kilometer grounded in a real city. A camera trajectory placed on a map produces continuous dynamic video depicting actual surroundings along the route. Users can further reshape the scene through text prompts, enabling imaginative scenarios.
Figure 2: Data overview: (a) real street-view and (b) synthetic datasets. The satellite-view map (leftmost column) shows the target video trajectory, with green and red indicating the start and end points, and yellow indicating reference view locations selected via cross-temporal pairing.
Figure 3: View interpolation pipeline: (a) Keyframe conditioning via channel concatenation, and (b) Keyframe conditioning with intermittent freeze-frame strategy ( Ours ). 4F and 1F denote latents derived from four frames and one frame, respectively, prior to the 4 × temporal compression of the 3D VAE.
Figure 4: Model overview. Given a start location, SWM autoregressively generates video grounded in a real city based on text prompt P(i) , and target camera trajectory C(i) , retrieving the relevant street-view images from a geo-indexed database. These retrieved images provide a Virtual Lookahead Sink for long-horizon stability and serve as conditioning for geometric referencing and semantic referencing to ground the generation in real-world geometry and appearance. The model generates each chunk autoregressively conditioned on self-generated history.
Figure 5
Figure 7: Qualitative results. Grounded in a real-world city, SWM can (a) generate high-fidelity videos across diverse scenarios from user-guided text prompts, (b) remain robust to diverse camera trajectories, and (c) reduce error accumulation, enabling long-horizon generation over several kilometers.
Figure 8: Qualitative comparisons with other methods. Across Busan- and Ann-Arbor-City-Bench, SWM produces physically grounded videos consistent with real urban structure and camera motion. Other world models, not originally designed for real-world grounding, tend to struggle with scene coherence, trajectory alignment, or long-horizon stability in this setting.
Method
FID ↓
FVD ↓
Img.Q. ↑
RotErr ↓
TransErr ↓
mPSNR ↑
mLPIPS ↓
Aether [ 63 ]
141.24/132.77
1096.50/1214.84
0.55/0.51
0.030/ 0.078
0.083/ 0.192
11.10/13.03
0.671/0.635
DeepVerse [ 7 ]
130.32/182.95
892.63/1524.97
0.53/0.46
0.062/0.251
0.103/0.469
12.20/13.43
0.679/0.727
Yume1.5 [ 32 ]
54.82/85.62
425.24/993.62
0.73/ 0.61
0.153/0.326
0.104/0.271
12.09/14.15
0.667/0.623
HY-World1.5 [ 20 ]
49.63/67.02
544.04/864.76
0.78 /0.54
0.044/0.193
0.079/0.221
11.87/ 14.26
0.588/0.575
FantasyWorld [ 8 ]
83.51/67.72
783.11/917.57
0.63/0.49
0.056/0.215
0.141/0.302
10.01/11.97
0.654/0.592
Lingbot [ 46 ]
62.14/57.99
717.44/1039.50
0.75/0.60
0.081/0.269
0.073/0.239
10.48/12.51
0.645/0.641
Table 1: Quantitative comparison with other methods. We evaluate visual and temporal fidelity, camera-following accuracy, and 3D adherence. Values are reported as Busan-City-Bench / Ann-Arbor-City-Bench.
Variant
FID ↓
FVD ↓
Img.Q. ↑
RotErr ↓
TransErr ↓
mPSNR ↑
mLPIPS ↓
Full model
28.43
301.76
0.78
0.020
0.015
14.56
0.392
w/o cross-temporal pairing
44.74
487.87
0.77
0.057
0.123
12.54
0.519
w/o synthetic data
27.74
365.24
0.78
0.021
0.020
13.52
0.427
w/o real street-view data
29.82
467.58
0.77
0.059
0.050
13.99
0.411
w/o geometric referencing
33.01
398.74
0.79
0.036
0.051
12.33
0.525
w/o semantic referencing
30.27
326.18
0.78
0.032
0.022
14.08
0.442
Table 2: Ablation on data strategy, conditioning strategy, and attention sink design. We evaluate visual and temporal fidelity, camera-following accuracy, and 3D adherence. Values are reported as Busan-City-Bench.
Figure 10
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 11: Coverage area. Spatial distribution of collected street-view data within the Seoul Metropolitan Area.
Method
PSNR ↑
SSIM ↑
LPIPS ↓
Channel Concatenation
22.52
0.628
0.245
Intermittent Freeze-Frame (Ours)
25.03
0.703
0.162
Appendix
Table 3: Quantitative comparison of view interpolation strategies. We compare the channel concatenation baseline with our Intermittent Freeze-Frame approach.
Figure 12: Qualitative comparison of view interpolation strategies. Channel concatenation and our Intermittent Freeze-Frame (IFF).
Figure 13: Effect of the number of retrieved references K .
Figure 14: Attention sink comparison under Self-Forcing. Sliding-window FID over time for different sink strategies in the SF configuration.
Method
FID ↓
FVD ↓
Img.Q. ↑
RotErr ↓
TransErr ↓
mPSNR ↑
mLPIPS ↓
VMem [ 25 ]
100.75
913.26
0.748
0.105
0.212
12.74
0.550
GEN3C [ 34 ]
45.72
416.94
0.732
0.030
0.082
14.16
0.524
SWM (TF)
28.43
301.76
0.781
0.020
0.015
14.56
0.392
Appendix
Table 4: Quantitative comparison with video generative models for static scenes on Busan-City-Bench.
Variant
FID ↓
FVD ↓
Img.Q. ↑
RotErr ↓
TransErr ↓
mPSNR ↑
mLPIPS ↓
Full model (VL Sink)
25.13
394.58
0.764
0.027
0.029
13.70
0.480
w/o any attention sink
37.37
550.81
0.751
0.026
0.041
12.94
0.575
w/ first-frame attention sink
30.85
440.65
0.767
0.045
0.044
13.08
0.534
w/ first-position attention sink
28.57
439.69
0.769
0.019
0.021
13.34
0.507
Appendix
Table 5: Long-horizon attention sink ablation on Busan-City-Bench (1,460 frames). Sequences are 4× longer than the standard benchmark.
Figure 15: Failure cases. Vehicles occasionally appear or disappear abruptly due to temporally inconsistent street-view sequences in the training data.
Figure 16: Additional qualitative ablation results. We show the effect of each component on generation quality on the Busan long-horizon benchmark. Specifically, (a) ours without VL sink, (b) ours with first-frame attention sink, (c) ours with first-position attention sink, (d) ours without geometric referencing, (e) ours without semantic referencing, and (f) ours without cross-temporal pairing.