VIEScore2: Unified Image Evaluation with Spatially Grounded Explanations

Xianda Du^*♠, Max Ku^*♠, Weiming Ren^♠, Zhi Rui Tam^♡, Chunlin Ren^♢, Ping Nie^♠, Min-Hung Chen^♣, Wenhu Chen^♠

^♠University of Waterloo ^♣NVIDIA ^♡National Taiwan University ^♢Nanyang Technological University
^*Equal contribution {a32du, m3ku, wenhu.chen}@uwaterloo.ca

arXiv Code 🤗 Model

An evaluation example with VIEScore2: inputs, evaluation instruction, predicted scores and defect grid, and the parsed explanation — **An evaluation example with VIEScore2.** The model predicts perceptual quality (PQ), semantic consistency (SC), and a 16×16 defect grid, where red and amber denote visual artifacts and semantic misalignments. A parameter-free parser converts the predictions into natural language.

TL;DR

Most image evaluators give you a single score but don't show where the image goes wrong. VIEScore2 is one model that scores generated and edited images and marks the defective regions on a 16×16 grid in a single pass. A simple rule-based parser then turns those predictions into a readable explanation.

Because the grid is plain text, it can be trained with SFT and then GRPO using rewards checked directly against ground truth. The result agrees with human ratings better than general-purpose VLMs such as GPT and Gemini under matched inputs, and is competitive with specialized spatial evaluators at localizing defects.

0.601

overall-score SRCC on the primary suite
(best VLM baseline: 0.491)

3 / 6

localization benchmarks ranked first in grid IoU
(top three on 5 / 6)

38K

training examples from 5 sources,
generation and editing tasks

Unified scoring, localization & explanation

One model handles generation and editing with zero, one, or multiple conditioning images. It outputs scores and defect locations without task-specific heads; a parameter-free parser turns them into spatially grounded feedback.

Verifiable spatial post-training

Defects are sparse cells on a text-native grid, so outputs are directly parseable. GRPO with a cell-level Dice reward optimizes localization overlap without a learned reward model.

Heterogeneous supervision

Scores, masks, and heatmaps from five sources are normalized into one format, keeping only the supervision each example actually has.

Method

Masks and heatmaps from ImagenWorld, RichHF-18K, PAL4VST and COCO, together with score-only data from EvalMuse-40K, are converted into one training format: scores on a 0–10 scale and defect cells on an N×N grid, written as text (e.g. r5: 8,9 marks columns 8–9 of row 5). Qwen3-VL-8B is first trained with supervised fine-tuning, then with GRPO using verifiable rewards for cell-level Dice overlap, score accuracy, and output format.

At inference, a parameter-free parser groups the predicted defect cells, describes their locations, and links artifacts to perceptual quality and misalignments to semantic consistency, so explanations stay faithful to the predictions by construction.

Overview of VIEScore2: data preparation, two-stage SFT and GRPO training, and inference — **Overview of VIEScore2:** data preparation, SFT and GRPO training, and inference. GRPO combines Dice, score, and format rewards. The parameter-free parser converts predicted scores and defect grids into explanations. GT denotes ground truth.

VIEScore2 in Action

A held-out RichHF example, end to end: the model's raw output lists scores and defect cells row by row (! marks cells with high annotation coverage), and the parameter-free parser turns it into the explanation on the right.

A RichHF example showing the generated image, ground-truth grid, predicted grid, raw model output, and the parser's explanation — A RichHF example with annotations, predictions, raw output, and the parameter-free parser's explanation. Red denotes artifacts and blue misalignment.

The same model handles all six ImagenWorld tasks: text-to-image generation (TIG), text-guided editing (TIE), and generation/editing with one (SRIG/SRIE) or multiple (MRIG/MRIE) conditioning images.

TIG · TIE · SRIG
SRIE · MRIG · MRIE

TIG, TIE, and SRIG examples with inputs, ground-truth and predicted defect grids — TIG, TIE, and SRIG examples. Orange cells mark defects.

Strong Localization across Benchmarks

All spatial predictions are compared on a shared 16×16 grid. VIEScore2 ranks first in per-image grid IoU on RichHF (0.299), PAL4VST (0.335), and SynthScars (0.234), second on HAD, and third on AbHuman, and it also leads in SynthScars F₁ (0.368).

	Training sources				Additional datasets
	RichHF		PAL4VST		AbHuman		HAD		SynthScars		SDG-30K
Method	IoU	F₁	IoU	F₁	IoU	F₁	IoU	F₁	IoU	F₁	IoU	F₁
Qwen3-VL-8B	0.022	0.137	0.026	0.124	0.024	0.108	0.044	0.111	0.026	0.117	0.114	0.269
GPT-5.6-terra^†	0.096	0.218	0.118	0.239	0.171	0.217	0.148	0.251	0.162	0.281	0.058	0.217
GPT-5.6-sol^†	0.156	0.274	0.141	0.248	0.217	0.282	0.211	0.298	0.186	0.290	0.078	0.258
Gemini-3-Flash^†	0.047	0.082	0.048	0.068	0.057	0.079	0.075	0.119	0.062	0.101	0.031	0.068
Claude Opus 5.5^†	0.167	0.285	0.140	0.225	0.146	0.308	0.142	0.235	0.196	0.358	0.084	0.216
PAL	0.014	0.035	0.306	0.535	0.009	0.016	0.009	0.038	0.027	0.046	0.007	0.014
SegFormer-b0	0.274	0.450	0.146	0.237	0.130	0.246	0.156	0.286	0.170	0.283	0.103	0.260
RAHF	0.285	0.474	0.105	0.169	0.141	0.260	0.164	0.281	0.172	0.294	0.068	0.164
ImageDoctor	0.284	0.471	0.100	0.167	0.144	0.293	0.158	0.275	0.169	0.305	0.086	0.207
LEGION^§	0.086	0.176	0.074	0.130	0.174	0.152	0.191	0.324	0.218	0.363	0.038	0.120
SDG	0.140	0.243	0.115	0.196	0.253	0.292	0.154	0.228	0.222	0.330	0.137	0.353
VIEScore2	0.299	0.466	0.335	0.439	0.215	0.194	0.197	0.314	0.234	0.368	0.102	0.272

Defect localization on six test sets using a shared 16×16 grid. IoU averages per-image overlap; F₁ pools cell counts. ^†API models use fixed prompts and temperature 0. ^§LEGION uses its released intermediate checkpoint. Bold: best; underline: second best.

Qualitative defect-localization examples on six benchmarks comparing VIEScore2, GPT-5.6-sol, Opus 5.5 and ImageDoctor — Selected defect-localization examples across six benchmarks. Red cells denote predictions and green outlines denote ground-truth defect regions. Numbers report per-image grid IoU.

Quality Scoring across Tasks

With matched inputs (generated image, prompt, and any conditioning images), VIEScore2 achieves the highest aggregate overall-score SRCC. ImagenWorld covers text-to-image generation (TIG), text-guided editing (TIE), and generation/editing with one (SRIG/SRIE) or multiple conditioning images (MRIG/MRIE). In joint evaluation, VIEScore2 also reaches 0.324 grid IoU and 0.506 F₁ with PQ/SC SRCC of 0.558/0.564, ahead of Qwen3-VL-8B and GPT-5.6-terra on every metric.

Method	All	RichHF	EvalMuse	TIG	TIE	SRIG	SRIE	MRIG	MRIE
Qwen3-VL-8B judge^‡	0.373	0.456	0.473	0.570	0.066	0.128	0.393	0.537	−0.290
GPT-5.6-terra^†	0.402	0.440	0.567	0.640	0.493	0.258	0.110	0.523	0.539
GPT-5.6-sol^†	0.437	0.470	0.655	0.623	0.590	0.275	0.253	0.539	0.684
Gemini-3-Flash^†	0.491	0.515	0.535	0.531	0.686	0.356	0.404	0.418	0.677
VIEScore2	0.601	0.692	0.803	0.589	0.429	0.375	0.403	0.509	0.417

Overall-score SRCC on the primary suite (900 examples) with matched inputs. ^†API correlations use successfully parsed scores. ^‡Qwen3-VL-8B uses our evaluation instructions without fine-tuning.

Analysis

GRPO beats more SFT. Across three runs, GRPO improves grid IoU by 0.024–0.029 over the shared SFT checkpoint; an extra SFT epoch gives only 0.001.
The Dice reward does the work. Dice alone raises grid IoU from 0.296 to 0.320; the full reward reaches 0.324 IoU and 0.506 F₁.
Conditioning images help. On the same 250 conditional examples, giving the model the source/reference images raises overall-score SRCC from 0.420 to 0.451 and localization F₁ from 0.514 to 0.536, with everything else held fixed.
N=16 balances fidelity and length. Trained pixel-IoU is similar at N=12 and N=16 (0.235/0.231) and drops at N=32 (0.192). N=16 preserves annotations better than N=12 (0.615 vs. 0.535) while keeping 95th-percentile targets far shorter than N=32 (471 vs. 1,926 tokens).

Grid resolution versus annotation fidelity, output length, and trained localization. Annotation fidelity compares grids with native masks. Shading marks N=16.

Limitations

Coarse grid. A 16×16 grid can miss small defects and only roughly follows irregular boundaries.
Faithful is not the same as correct. The parser's explanations faithfully reflect the predicted scores and grid, but don't check whether those predictions are right or explain why a defect occurs.
False alarms. GRPO can increase false alarms on some clean-image subsets, trading conservative clean-image recognition for stronger localization.
One backbone. VIEScore2 is validated only with Qwen3-VL-8B; other backbones and scales remain untested.

Failure cases: small defects lost on the grid, missed misalignment, a false alarm on a clean image, and an incorrect prediction with its explanation — Failure cases: lost small defects, missed misalignment, false alarms, and an incorrect prediction with its explanation.

Citation

@misc{du2026viescore2unifiedimageevaluation,
      title={VIEScore2: Unified Image Evaluation with Spatially Grounded Explanations},
      author={Xianda Du and Max Ku and Weiming Ren and Zhi Rui Tam and Chunlin Ren and Ping Nie and Min-Hung Chen and Wenhu Chen},
      year={2026},
      eprint={2610.00994},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.00994},
}