Reaction
RVM reportedly scores 84.13 on VBench Overall for Wan2.1-1.3B
A post says RVM defines its target from the final sample rather than noisy intermediate transitions.
TLDR
A thread says RVM defines its target from the final sample, unlike trajectory-based updates that reinforce noisy intermediate transitions. It reports an 84.13 VBench Overall score on Wan2.1-1.3B, versus 83.10 for the base with CFG. It cites estimates of about 5.8 GPU-hours per iteration for RVM versus about 14.5 for DanceGRPO at the same batch size; the 11.8x total-cost comparison uses different training configurations.
Combined views
—
1 Source, first seen ago
— likes— comments— saves— reposts
