VGC
A post-processing video-to-video model that corrects geometric inconsistencies in generated videos through learned synthetic paired supervision, without modifying the source generator.
Timeline (last 90 days)
VGC (Video Geometry Corrector) is a post-hoc video-to-video model trained on synthetic paired supervision that corrects geometric errors in generated videos without modifying the source generator.
When training VGC, optical flow is used as an auxiliary cue to help the model disentangle camera motion from object motion.
In both pose-conditioned and prompt-driven benchmarks, VGC improves the geometric consistency of multiple source generators, indicating that its correction capability is not tied to any particular generator.
After scaling the backbone to a larger size, VGC can extend toward more general video refinement.
As a post-processing module, VGC's correction capability is limited by how much information the source video retains; if the source generator has already broken down severely geometrically, the room for post-processing recovery is also limited.
The distribution gap between synthetic distortions and real generation distortions remains an issue that VGC methods need to continuously verify.
The research on VGC comes from rss (referring to the RSS conference or source).
All Facts (7)
VGC (Video Geometry Corrector) is a post-hoc video-to-video model trained on synthetic paired supervision that corrects geometric errors in generated videos without modifying the source generator.
50%UnverifiedWhen training VGC, optical flow is used as an auxiliary cue to help the model disentangle camera motion from object motion.
50%UnverifiedIn both pose-conditioned and prompt-driven benchmarks, VGC improves the geometric consistency of multiple source generators, indicating that its correction capability is not tied to any particular generator.
50%UnverifiedAfter scaling the backbone to a larger size, VGC can extend toward more general video refinement.
50%UnverifiedAs a post-processing module, VGC's correction capability is limited by how much information the source video retains; if the source generator has already broken down severely geometrically, the room for post-processing recovery is also limited.
50%UnverifiedThe distribution gap between synthetic distortions and real generation distortions remains an issue that VGC methods need to continuously verify.
50%UnverifiedThe research on VGC comes from rss (referring to the RSS conference or source).
50%