VIEScore2: Unified Image Evaluation with Spatially Grounded Explanations

VIEScore2. Unified Image Evaluation with Spatially Grounded Explanations. By Xianda Du and seven others, from the University of Waterloo, NVIDIA, and two other institutions. Published on the first of October. Its code repository is worked out: about fifty-five source files, with a readme and a licence.

Existing evaluators of synthetic images typically return a single scalar quality score, and do not identify the regions of the image that support it. The number says how good an image is without saying where it fell short. A score of that kind can rank images, but it cannot point at the part of an image that brought the score down, and so it cannot explain itself.

VIEScore2 is one evaluator for both image generation and image editing, with optional conditioning images. Unified here means one model covering both tasks, with or without an image to condition on. It represents an image as an N by N grid and, in a single model pass, jointly predicts quality scores and the locations of defects. Because the grid representation is native to text, it gives heterogeneous kinds of spatial supervision a common interface, and it makes post-training objectives directly verifiable.

Training uses thirty-eight thousand examples. Some carry scores only, some localization only, and some both, across generation and editing tasks. After supervised fine-tuning, GRPO is applied to improve defect localization, with rewards that combine cell-level Dice overlap, score accuracy, and the validity of the output format. The text-native grid is what makes those rewards checkable. Predicted cells can be compared with the true defect cells directly, which is what the Dice overlap measures, alongside whether the score is right and whether the output can be parsed at all. A parameter-free parser then turns the structured predictions into readable explanations.

On the primary suite, VIEScore2 reaches an overall-score SRCC of zero point six zero one. The strongest zero-shot general-purpose VLM baseline under matched inputs, Gemini-3-Flash, reaches zero point four nine one. For defect localization, measured by per-image grid IoU, VIEScore2 outperforms both general-purpose VLMs and specialized spatial evaluators on three of six benchmarks, and ranks among the top three on five of them, including datasets beyond its training sources. Two abilities are measured separately here, scoring and localizing, and on the second the comparison includes evaluators built specifically for spatial judgments.

The grid holds the design together. Writing location as text lets one model score and point in the same pass, lets different kinds of labels train it, lets a reward check its answers cell by cell, and lets a plain parser explain them.