Naver's Seoul World Model Generates Video of Real City Streets Without Hallucinating Them

Researchers at NAVER AI Lab, KAIST, and Seoul National University published the Seoul World Model (SWM) this week — the first AI video generation system explicitly grounded in a real city rather than a synthetically imagined one. The difference matters more than it sounds.

Existing video world models generate visually plausible environments by inventing the content from scratch. They look realistic. They do not correspond to anywhere. SWM instead retrieves actual street-view images from Seoul as conditioning signals during generation, anchoring each output to the real geometry of a specific location. The result is a system that can generate multi-kilometer video trajectories through actual Seoul streets — and that generalizes to other cities it was never trained on.

The training data is specific: 1.2 million real panoramic street-view images captured across Seoul, plus 10,000 synthetic videos from a CARLA urban simulator covering 431,500 square meters of city area. The synthetic data solved a structural limitation — vehicle-mounted cameras capture forward views at sparse intervals, leaving gaps in coverage. The simulator filled those gaps with diverse camera trajectories that real street-view capture cannot provide.

The core technical mechanism is retrieval-augmented generation applied to video. Given GPS coordinates, camera movement instructions, and optional text prompts, SWM retrieves nearby street-view images and conditions the autoregressive video model on the geometry and appearance of those references. This grounds each generated segment to what that location actually looks like, preventing the spatial drift that plagues long-horizon video generation.

The paper introduces what the authors call a Virtual Lookahead Sink: during generation, the system continuously retrieves a street-view image from a future point on the planned trajectory and uses it as an anchor to stabilize the current chunk. Over hundreds of meters of generated trajectory, this prevents accumulated error from making the output drift into incoherence.

On benchmarks across Seoul, Busan, and Ann Arbor — the last of which the model had never seen during training — SWM outperformed six existing video world models on visual quality and temporal consistency. The generalization result is notable: a model trained on Korean street-view data held up on a Michigan city without any additional fine-tuning.

The text-prompting capability adds a practical dimension. Users can describe scenarios — waves flooding the street, Godzilla between skyscrapers, changed weather conditions — and SWM integrates these alterations while maintaining the spatial fidelity of the base location. This combination of real-world grounding and fantasy overlay is precisely what most synthetic world models lack.

The obvious applications are urban planning simulation, autonomous driving evaluation in realistic city environments, and training data generation for robotics. More speculatively, this is the foundation for maps that are not static images but dynamic, navigable simulations of real places. Google and Apple run street-view programs for similar reasons. SWM shows that the underlying data can now support interactive video generation, not just photo capture.

Code and model weights are planned for release. The arXiv paper is 2603.15583.