---
license: apache-2.0
base_model: Qwen/Qwen3-VL-8B-Instruct
pipeline_tag: image-text-to-text
library_name: transformers
tags:
  - image-evaluation
  - image-quality-assessment
  - defect-localization
  - grounding
  - grpo
  - reinforcement-learning
---

# VIEScore2

A unified evaluator for generated and edited images. In a single
autoregressive pass the model emits:

- a **quality score** — dual-axis perceptual quality (`pq:`) and semantic
  consistency (`sc:`),
- a **defect grid** — sparse cells on a text-native 16×16 grid localizing
  problems, optionally split into `artifact:` / `misalign:` channels, and
- a **verifiable explanation**, rendered deterministically from the score and
  grid so the text can never contradict the spatial evidence.

Because every output element is a discrete text token set, predictions are
exactly checkable against ground truth. Training is supervised fine-tuning
followed by **GRPO with a verifiable cell-level F_β reward** on the defect
grid.

## Repository contents

**Naming convention:** `VIEScore2` denotes the full model after GRPO;
ablated variants are marked by what they lack, never by added suffixes.

| path | contents |
|---|---|
| `/` (root) | **VIEScore2** — the full post-GRPO checkpoint (drop-in `Qwen3VLForConditionalGeneration`) |
| `wo-grpo/` | **VIEScore2 (w/o GRPO)** — the SFT-stage LoRA adapter on `Qwen/Qwen3-VL-8B-Instruct` |

## Usage

```python
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

model = Qwen3VLForConditionalGeneration.from_pretrained(
    "Allenda/VIEScore2", dtype="bfloat16", device_map="auto")
processor = AutoProcessor.from_pretrained("Allenda/VIEScore2")
```

Evaluation prompts, the frozen protocol, the deterministic explanation
renderer, and all benchmark converters and baseline harnesses live in the
accompanying code release (see the paper).

## Evaluation

On a 1,300-example multi-source suite (RichHF, PAL4VST, EvalMuse,
ImagenWorld, COCO; frozen protocol v3.1): localization cell-F1 **0.506**,
problem-sample grid IoU **0.324**, overall-score SRCC **0.601**. After mapping
all methods to a shared 16×16 grid, VIEScore2 is the only evaluator in the
top two of cell-F1 or grid-IoU on all six external localization benchmarks
(RichHF, AbHuman, HAD, SynthScars, PAL4VST, SDG-30K). See the paper and code
repository for the full tables and protocol details.

## 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}, 
}
```