VVR paper demos: verifier cleanly separates RL outputs from baselines

StellaLisy · x · 2026-09-29

Stella Lisy's team shares VVR generation samples: on a VVR prompt the verifier accepts both RLVVR outputs and rejects both baselines; on a DrawBench prompt, VVR-Easy renders the vase as a flat unshaded shape consistent with its lower aesthetic and HPSv2.1 scores, while adding GenEval2 restores shading and improves both metrics. This is the demo portion of their Verifiable Visual Rewards paper thread.

Related event: Verifiable Visual Rewards Boost SD3.5 Instruction Following from 2.8% to 28.3%(3 posts)→

Original post →

More from Multimodal

Multimodal channel →