VIEScore2: Unified Image Evaluation with Spatially Grounded Explanations
VIEScore2 scores generated or edited images and points to the grid cells behind the judgment.
The paper frames the model as an evaluator that returns both quality scores and localized defect explanations in one pass. It trains on 38K mixed examples across score-only, localization-only, and joint supervision, then uses GRPO to improve localization and format validity. On the authors’ primary suite, it reports 0.601 overall-score SRCC, ahead of Gemini-3-Flash at 0.491 under matched inputs. For defect localization, it beats general VLMs and specialized spatial evaluators on three of six benchmarks by per-image grid IoU, and ranks top three on five. HF Daily Papers' note
The paper frames the model as an evaluator that returns both quality scores and localized defect explanations in one pass. It trains on 38K mixed examples across score-only, localization-only, and joint supervision, then uses GRPO to improve localization and format validity. On the authors’ primary suite, it reports 0.601 overall-score SRCC, ahead of Gemini-3-Flash at 0.491 under matched inputs. For defect localization, it beats general VLMs and specialized spatial evaluators on three of six benchmarks by per-image grid IoU, and ranks top three on five. HF Daily Papers' note
score 4