Tool-IQA-8B

🧰 Tool-IQA: Augmenting Image Quality Assessment with Simple Tools

Tool-IQA is a tool-augmented Vision-Language Model workflow for no-reference Image Quality Assessment (NR-IQA). Instead of giving a score from a single glance, Tool-IQA observes quality attributes, inspects uncertain visual evidence, and predicts a calibrated quality score.

🚧 Current release: inference scripts + model weights. Training code and scheduler scripts will be released later.

✨ Highlights

• 🧠 Inspection-based IQA instead of one-shot scoring.
• 🔍 Lightweight visual inspection for local details and visibility cues.
• 📝 Rubric-guided reasoning before final score prediction.
• ⚡ vLLM-based asynchronous benchmark inference.

✅ Release Status

• README and paper figures
• Tool-augmented inference: infer/tool_infer.py
• Inference-time visual inspection tools
• Checkpoint loading via TARGET_MODEL_NAME
• Public training code
• Training scheduler / launch scripts
• Full training data preparation pipeline

The GRPO/BATE training pipeline is not included in this release and will be added later.

🧭 Method Overview

Tool-IQA follows three compact steps:

1. Rubric observation: summarize perceptual attributes such as sharpness, noise, artifacts, exposure, color, and subject visibility.
2. Visual inspection: gather additional local or visibility-aware evidence when the initial observation is uncertain.
3. Final quantification: synthesize all evidence and output a score in [1, 5].

⚙️ Installation

conda create -n tool-iqa python=3.12
conda activate tool-iqa
pip install numpy pillow tqdm scipy transformers vllm

🚀 Inference

Prepare the checkpoint and data first. See data/README.md for the expected dataset layout.

export TARGET_MODEL_NAME=/path/to/Tool-IQA-weights
python infer/tool_infer.py

Results are saved to:

results/toolIQA/<dataset>/<model_name>/results.txt

Each row records the image name, predicted score, ground-truth score, score history, reasoning history, inspection history, and error message if any.

🙏 Acknowledgement

This project is built based on ms-swift. We thank the MS Swift team for their open-source framework.

📬 Contact

For questions, please contact [email protected].