Hi Cosmos team,
I have a question about the Driving Transfer setting in the Cosmos 3 technical report.
The report says the MADS driving data has 7 synchronized camera views and per-camera world-scenario-map controls.
Was Cosmos 3 trained to generate all views jointly, or is Driving Transfer trained/inferred as single-view control-video -> RGB-video pairs?
Also, does the current open-source Cosmos 3 release support world-scenario-map / driving transfer control inputs at inference time? If not, is this planned for a future release?
Thanks!
Hi Cosmos team,
I have a question about the Driving Transfer setting in the Cosmos 3 technical report.
The report says the MADS driving data has 7 synchronized camera views and per-camera world-scenario-map controls.
Was Cosmos 3 trained to generate all views jointly, or is Driving Transfer trained/inferred as single-view control-video -> RGB-video pairs?
Also, does the current open-source Cosmos 3 release support world-scenario-map / driving transfer control inputs at inference time? If not, is this planned for a future release?
Thanks!