Report
UGG gives one model two jobs—video generation and 3D reconstruction, its researchers say
The team says UGG jointly models video, depth and camera motion to fill in missing scene information.
TLDR
According to UGG’s project page, the same model can predict depth and camera geometry from video, generate unseen viewpoints, or produce video, depth and camera motion from text. Rather than separating scene generation from reconstruction, it models them together. The researchers also report that searching multiple possible scene completions at inference time improves geometric predictions from sparse observations.
Combined views
25.2K
3 Sources, first seen ago

Comments