Tag
VIEScore2 introduces a unified evaluator for image generation and editing that predicts quality scores and defect locations on an NxN grid in a single model pass, using GRPO post-training with cell-level localization rewards. It outperforms general-purpose VLMs such as Gemini-3-Flash, achieving 0.601 overall SRCC and top-three results on five of six defect-localization benchmarks.