SA-IQA is a multimodal image quality assessment model released with “Beyond Pixels: Benchmarking and Reward-Based Assessing Framework for Visual Spatial Aesthetics.”
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
text
1SA-IQA-model/
2├── LICENSE
3├── README.md
4├── Ovis2.5-9B/ # Base model used by training scripts
5│ ├── LICENSE
6│ ├── NOTICE
7│ ├── config.json
8│ ├── modeling_ovis2_5.py
9│ ├── model-00001-of-00004.safetensors
10│ ├── model-00002-of-00004.safetensors
11│ ├── model-00003-of-00004.safetensors
12│ ├── model-00004-of-00004.safetensors
13│ └── ...
14└── sa-iqa-prompt4/ # Fine-tuned checkpoint used for inference
15 ├── config.json
16 ├── modeling_ovis2_5.py
17 ├── model-00001-of-00004.safetensors
18 ├── model-00002-of-00004.safetensors
19 ├── model-00003-of-00004.safetensors
20 ├── model-00004-of-00004.safetensors
21 └── ...
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:
bibtex
1@InProceedings{Gao_2026_CVPR,
2 author = {Gao, Yuan and Song, Jin and Fei, Yiyun and Li, Gongzhe and Yang, Ruigao},
3 title = {Beyond Pixels: Benchmarking and Reward-Based Assessing Framework for Visual Spatial Aesthetics},
4 booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
5 month = {June},
6 year = {2026},
7 pages = {1819-1829}
8}
