Korean developers from Naver have demonstrated a project called Seoul World Model — which managed to incorporate a significant portion of the real Seoul. This model is a kind of virtual city analog capable of showcasing travel through its streets without the risk of destruction. Moreover, it can be “fed” various fantasies: for example, creating scenarios of floods, alien attacks, or even Godzilla.
All of this is implemented on the basis of Nvidia Cosmos Predict 2.5 — a DiT model with 2 billion parameters. Training was completed in just 24 hours on powerful H100 graphics processors of the latest generation. As a result, the model can display about 15 frames per second on a single H100, although no one has yet focused heavily on optimizing it for running on local machines. The code and weights have not been published yet, but they promise to do so quite soon.
They plan to release an overview and the models in the near future.
Created with n8n:
https://cutt.ly/n8n
Created with syllaby:
https://cutt.ly/syllaby