VIEScore2: Unified Image Evaluation with Spatially Grounded Explanations
VIEScore2 addresses the shortcoming that most automatic image evaluators return only a scalar quality score without identifying defective regions. It introduces a unified evaluator that represents an image as a sparse $16\times 16$ text grid and jointly predicts perceptual quality, semantic consistency, and defect cell coordinates in a single autoregressive pass. By framing localization as text generation, it enables directly verifiable post-training rewards and supports heterogeneous supervision across generation and editing tasks. The practical value lies in producing human-readable, spatially grounded feedback from one model without task-specific prediction heads.
VIEScore2 makes a credible, methodologically interesting step toward explainable image evaluation. Its text-native grid representation and GRPO-based post-training with a cell-level Dice reward yield measurable improvements over supervised fine-tuning, and the final model surpasses matched general-purpose VLM baselines in aggregate score correlation. As the abstract reports, "On the primary suite, VIEScore2 achieves an overall-score SRCC of 0.601, compared with 0.491 for Gemini-3-Flash, the strongest zero-shot general-purpose VLM baseline under matched inputs." Localization is competitive across six benchmarks, but it is not uniformly dominant.
The sparse grid representation is the paper’s strongest design decision: it unifies masks, heatmaps, and boxes into a common text format, allowing joint training on score-only, localization-only, and fully supervised examples without task-specific heads. The GRPO ablation is also convincing. The authors write, "Across three independent runs, GRPO improves grid IoU by 0.024–0.029 over the shared SFT checkpoint, whereas an additional SFT epoch gives 0.001," which cleanly isolates the benefit of the reinforcement stage from mere additional gradient steps.
The fixed grid fundamentally limits spatial fidelity, and the authors acknowledge this explicitly: "The fixed grid can miss small defects and only coarsely approximates irregular boundaries." They also note a concrete failure mode of the post-training stage: "GRPO can increase false alarms on some clean-image subsets, reflecting a trade-off between stronger defect localization and conservative clean-image recognition." Beyond representation limits, the entire validation rests on a single backbone (Qwen3-VL-8B), leaving transfer to other VLMs or scales unstudied. Comparisons with specialized spatial evaluators are further confounded by differing native outputs and input interfaces.
Under matched $\chi_K$ inputs, VIEScore2 leads all tested general-purpose VLMs on the primary 900-example suite with an overall SRCC of 0.601 versus 0.491 for Gemini-3-Flash (Table 2). However, localization is split across benchmarks: VIEScore2 wins on RichHF, PAL4VST, and SynthScars, yet trails GPT-5.6-sol on HAD and SDG on SDG-30K (Table 1). The reward ablation supports the authors’ key claim but also reveals fragility: they note that "The Dice-only variant reaches the same SRCC (Table 16), so the score reward does not measurably change scoring on the primary suite; its main effect is a small F1 gain (0.481 to 0.506)," suggesting the scoring head does not benefit much from the multi-objective reward.
Hyperparameters and data pipelines are documented with reasonable care in the appendices. For SFT, the authors specify that they use "LoRA on all linear layers (rank 32, scaling 16, dropout 0.05)" and for GRPO they use "8-bit AdamW with learning rate $10^{-5}$, one epoch, four responses per example, eight gradient-accumulation steps, and no KL penalty." The 38K-example training corpus is drawn from five public sources with detailed grid-conversion rules in Appendix A. Nevertheless, no public code release or model checkpoint is mentioned, and exact reproduction would require replicating the parameter-free parser templates (Appendix G) and the precise stratified sampling of 2,400 examples for GRPO. An independent group could reproduce the work, but missing artifacts would raise friction.
Existing synthetic image evaluators typically provide only a scalar quality score and do not identify the image regions that support it. We introduce VIEScore2, a unified evaluator for image generation and editing tasks with optional conditioning images. VIEScore2 represents an image as an N x N grid and jointly predicts quality scores and defect locations in a single model pass. Its text-native grid representation provides a common interface for heterogeneous spatial supervision and enables directly verifiable post-training objectives. We train on 38K examples spanning score-only, localization-only, and joint supervision across generation and editing tasks. Starting from supervised fine-tuning, we further apply GRPO to improve defect localization using rewards that combine cell-level Dice overlap, score accuracy, and output-format validity. A parameter-free parser converts the structured predictions into readable explanations. On the primary suite, VIEScore2 achieves an overall-score SRCC of 0.601, compared with 0.491 for Gemini-3-Flash, the strongest zero-shot general-purpose VLM baseline under matched inputs. For defect localization, VIEScore2 outperforms both general-purpose VLMs and specialized spatial evaluators on three of six benchmarks in per-image grid IoU and ranks among the top three on five, including datasets beyond its training sources.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.