SA-IQA-model

SA-IQA Model

SA-IQA is a multimodal image quality assessment model released with “Beyond Pixels: Benchmarking and Reward-Based Assessing Framework for Visual Spatial Aesthetics.”

Accepted to CVPRW 2026.

GitHub | CVF Open Access | arXiv | Dataset

Usage Note: This repository is released as a full model bundle and is not intended to be loaded directly from the repository root. For inference with the SA-IQA codebase, download the full repository and use SA-IQA-model/sa-iqa-prompt4 as the model path.

The released final checkpoint is sa-iqa-prompt4, a fine-tuned model based on Ovis2.5-9B for assessing interior-image spatial aesthetics.

Hugging Face Release Layout

This Hugging Face repository is released as a full model bundle. Download the whole repository to ./SA-IQA-model when using it with the SA-IQA codebase.

The sa-iqa-prompt4/ directory is the released final fine-tuned checkpoint for inference. The Ovis2.5-9B/ directory is the bundled base model copy used by tools/train_sft.sh for training and reproducibility.

Because this repository contains two model directories, automatic loading from the repository root is not expected to work. Load the fine-tuned checkpoint from SA-IQA-model/sa-iqa-prompt4, or pass that path through the SA-IQA inference script with --model_path.

Model Details

Model Description

• Model type: multimodal vision-language model for image quality assessment
• Base model: Ovis2.5-9B
• Fine-tuned checkpoint: sa-iqa-prompt4
• Input: image plus a dimension-specific text prompt
• Output: textual quality label and token log-probabilities used to compute a continuous score
• Dimensions: distortion, harmony, layout, lighting

Intended Use

SA-IQA is intended for research, evaluation, and application use, including:

• spatial aesthetic assessment of interior images
• image quality benchmarking on SA-BENCH
• reward-model research for image generation and best-of-N selection
• comparison of prompt variants for spatial aesthetic assessment

Out-of-Scope Use

The model is not intended for:

• universal aesthetic judgment outside the interior-scene domain
• safety-critical or legally binding decision making

Usage

Use the SA-IQA inference script from the code repository:

python tools/infer.py --prompt_version 4 --mode all --dimension lighting

When running from the release bundle root, the default model path is: SA-IQA-model/sa-iqa-prompt4.

If you downloaded this Hugging Face repository to another local path, pass the nested sa-iqa-prompt4 checkpoint path through --model_path.

Release Bundle Structure

SA-IQA-model/
├── LICENSE
├── README.md
├── Ovis2.5-9B/                  # Base model used by training scripts
│   ├── LICENSE
│   ├── NOTICE
│   ├── config.json
│   ├── modeling_ovis2_5.py
│   ├── model-00001-of-00004.safetensors
│   ├── model-00002-of-00004.safetensors
│   ├── model-00003-of-00004.safetensors
│   ├── model-00004-of-00004.safetensors
│   └── ...
└── sa-iqa-prompt4/              # Fine-tuned checkpoint used for inference
    ├── config.json
    ├── modeling_ovis2_5.py
    ├── model-00001-of-00004.safetensors
    ├── model-00002-of-00004.safetensors
    ├── model-00003-of-00004.safetensors
    ├── model-00004-of-00004.safetensors
    └── ...

Training Data

The model is fine-tuned and evaluated on SA-BENCH, a 17,768-example benchmark for spatial aesthetics in interior scenes.

Limitations

• The model is designed for interior images and may not generalize to other image domains.
• Predictions are based on the SA-BENCH annotation protocol and prompt design.
• The output should be treated as an assessment signal, not as a definitive human aesthetic judgment.

License

The released SA-IQA model weights are licensed under the Apache License 2.0. See LICENSE for the full license text.

This model is fine-tuned from Ovis2.5-9B, which is also released under the Apache License 2.0. When redistributing or modifying this model, retain attribution and relevant notices from the base model:

• Ovis2.5-9B/LICENSE
• Ovis2.5-9B/NOTICE

Citation

If you use this model, please cite:

@InProceedings{Gao_2026_CVPR,
    author    = {Gao, Yuan and Song, Jin and Fei, Yiyun and Li, Gongzhe and Yang, Ruigao},
    title     = {Beyond Pixels: Benchmarking and Reward-Based Assessing Framework for Visual Spatial Aesthetics},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2026},
    pages     = {1819-1829}
}