---
license: apache-2.0
library_name: transformers
pipeline_tag: image-text-to-text
tags:
  - image-quality-assessment
  - iqa
  - iaa
  - vqa
  - aesthetics
  - video-quality-assessment
  - q-align
  - qwen3.5
  - vision-language-model
  - multimodal
base_model:
  - Qwen/Qwen3.5-VL
language:
  - en
metrics:
  - srcc
  - plcc
---

<div align="center">

# Q-ReAlign — Pro (9B)

**Lightweight, human-aligned multimodal quality judge built on a modern Qwen3.5 vision-language backbone.**

*Q-Align-level performance with roughly half the parameters — the largest, highest-fidelity variant of the family.*

[GitHub](https://github.com/Q-Future/Q-ReAlign) · [Method](https://github.com/Q-Future/Q-ReAlign/blob/main/docs/METHOD.md) · [Adapting guide](https://github.com/Q-Future/Q-ReAlign/blob/main/docs/ADAPTING.md) · Mini (0.8B) · Lite (4B) · **Pro (9B)**

</div>

---

## What this is

Q-ReAlign scores the **perceptual quality / aesthetic appeal** of an image or video the way Q-Align does: the model is asked to rate quality, and the probability mass it places on the discrete words **`excellent / good / fair / poor / bad`** is collapsed — via a fixed weighting `[1.0, 0.75, 0.5, 0.25, 0.0]` — into a single scalar in `[0, 1]`.

**Pro (9B)** is the flagship of three sizes (**Mini 0.8B · Lite 4B · Pro 9B**). All three match or beat the original Q-Align across seven QA benchmarks; quality scales cleanly with size, and Pro sits at the top.

- **Backbone:** Qwen3.5-VL (`model_type: qwen3_5`), hybrid linear/full attention text tower + SigLIP-style vision encoder
- **Tasks:** IQA (image quality) · IAA (image aesthetics) · VQA (video quality) — the unified ONE-ALIGN setting
- **Training:** full-parameter SFT in bf16 via [ms-swift](https://github.com/modelscope/ms-swift), vision tower + projector trainable
- **Precision:** bfloat16 · **dtype** `auto`

## Results

Per-dataset **SRCC / PLCC** on seven QA benchmarks. Pro (9B) reaches **avg SRCC 0.896 vs. Q-Align's 0.869**.

| Model | KonIQ | SPAQ | KADID | AGI | LIVE | AVA | LSVQ | **Avg.** |
|---|---|---|---|---|---|---|---|---|
| Q-Align | 0.942 / 0.944 | 0.932 / 0.933 | 0.912 / 0.920 | 0.738 / 0.781 | 0.897 / 0.870 | 0.798 / 0.796 | 0.867 / 0.866 | 0.869 / 0.873 |
| **Pro (9B)** | **0.950 / 0.952** | **0.935 / 0.937** | **0.934 / 0.939** | **0.843 / 0.885** | **0.902 / 0.876** | **0.832 / 0.828** | **0.883 / 0.884** | **0.896 / 0.900** |

<sub>Each cell is SRCC / PLCC. Numbers are the full evaluation sets (KonIQ, SPAQ, KADID, AGI, LIVE, AVA, LSVQ).</sub>

## Quick start

```python
import torch
from PIL import Image
from transformers import AutoModelForImageTextToText, AutoProcessor
# transformers >= 5.2.0 for Qwen3.5 support

CKPT, IMAGE = "q-future/Q-ReAlign-Pro-9B", "photo.jpg"
LEVELS  = ["excellent", "good", "fair", "poor", "bad"]
WEIGHTS = [1.0, 0.75, 0.5, 0.25, 0.0]

device = "cuda" if torch.cuda.is_available() else "cpu"
processor = AutoProcessor.from_pretrained(CKPT)
model = AutoModelForImageTextToText.from_pretrained(CKPT, dtype="auto").to(device).eval()

messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": "How would you rate the quality of this image?"},
]}]
text = processor.apply_chat_template(messages, add_generation_prompt=True) + "The quality of the image is"
inputs = processor(text=[text], images=[Image.open(IMAGE).convert("RGB")], return_tensors="pt").to(device)

ids = [processor.tokenizer(" " + w, add_special_tokens=False).input_ids[0] for w in LEVELS]
probs = model(**inputs).logits[0, -1, ids].softmax(-1)
score = (probs * torch.tensor(WEIGHTS, device=device)).sum().item()
print(f"quality score: {score:.4f}")   # 0 (worst) .. 1 (best)
```

The score is the expected value of the level weights under the model's next-token
distribution over the five level words — no sampling, one forward pass.

### Aesthetics or video

Swap the prompt for the task:

- **Aesthetics (IAA):** *"How would you rate the aesthetics of this image?"* → stem *"The aesthetics of the image is"*
- **Video (VQA):** sample N frames (default 8) and pass them as the image sequence; prompt *"How would you rate the quality of this video?"* → stem *"The quality of the video is"*

## Model details

| | Pro (9B) |
|---|---|
| Architecture | `Qwen3_5ForConditionalGeneration` |
| Text hidden size | 4096 |
| Text layers | 32 (linear attention with full-attention every 4th layer) |
| Vision encoder depth | 27, hidden 1152, patch 16, spatial merge 2 |
| Vocab | 248320 |
| Context length | up to 262144 |
| Tensor dtype | bfloat16 |
| Shards | 4 × safetensors (~18.8 GB total) |

## Scoring contract

- **Level vocabulary:** `excellent, good, fair, poor, bad`
- **Weights:** `[1.0, 0.75, 0.5, 0.25, 0.0]`
- **Output:** scalar in `[0, 1]`, higher = better
- The five level tokens are matched with a **leading space** (`" excellent"`, …); keep that when porting to other tokenizers.

## Intended use & limitations

- **Use:** no-reference image/video quality assessment, aesthetic scoring, dataset
  curation, ranking and filtering generated media, reward signals for generative
  pipelines.
- **Out of scope:** safety/content moderation, factual or identity judgments,
  medical/forensic grading. Quality is perceptual and dataset-conditioned.
- Scores are calibrated to the training MOS distribution; absolute values are most
  meaningful **relative** to one another. Re-calibrate before mixing with other scales.

## Acknowledgements & citation

Built on the shoulders of **[Q-Align](https://github.com/Q-Future/Q-Align)** (the
discrete text-defined-levels method and ONE-ALIGN), **[ms-swift](https://github.com/modelscope/ms-swift)**
(training/inference backbone), and **[Qwen3.5-VL](https://github.com/QwenLM/Qwen3-VL)**
(the vision-language backbone). If you use this model, please also cite the originals:

```bibtex
@inproceedings{wu2024qalign,
  title     = {Q-Align: Teaching {LMM}s for Visual Scoring via Discrete Text-Defined Levels},
  author    = {Wu, Haoning and Zhang, Zicheng and Zhang, Weixia and Chen, Chaofeng and
               Liao, Liang and Li, Chunyi and Gao, Yixuan and Wang, Annan and Zhang, Erli and
               Sun, Wenxiu and Yan, Qiong and Min, Xiongkuo and Zhai, Guangtao and Lin, Weisi},
  booktitle = {Proceedings of the 41st International Conference on Machine Learning (ICML)},
  year      = {2024}
}

@inproceedings{swift2025,
  title     = {{SWIFT}: A Scalable Lightweight Infrastructure for Fine-Tuning},
  author    = {ModelScope Team},
  booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)},
  year      = {2025},
  note      = {\url{https://github.com/modelscope/ms-swift}}
}

@misc{qwen3_5,
  title        = {Qwen3.5: Towards Native Multimodal Agents},
  author       = {Qwen Team},
  year         = {2025},
  howpublished = {\url{https://github.com/QwenLM/Qwen3-VL}}
}
```