You can deploy this model and start earning money today!

Thunderbolt215215/
ArtiMuse_AVA

Name: ArtiMuse_AVA
Availability: InStock

OpenGVLab/InternVL3-8B

finetune

apache-2.0

Views

No views yet

ArtiMuse: Fine-Grained Image Aesthetics Assessment with Joint Scoring and Expert-Level Understanding

[🌐 Project Page] [🚀 Online Demo] [💻 Code] [📄 Paper] [[🧩 Checkpoints: 🤗 Hugging Face | 🤖 ModelScope]]

🌟 Building upon on ArtiMuse, we introduce UniPercept, a comprehensive follow-up work that provides a meticulous study on perceptual-level image understanding. It spans Image Aesthetics Assessment (IAA), Image Quality Assessment (IQA), and Image Structure & Texture Assessment (ISTA) across both Visual Rating (VR) and Visual Question Answering (VQA) tasks.

arXiv

Website

<img alt="Code" src="https://img.shields.io/badge/GitHub-UniPercept-181717?logo=github&logoColor=white" height="25" />

> > > >

> > > > > >

> >

Shuo Cao, Nan Ma, Jiayang Li, Xiaohui Li, Lihao Shao, Kaiwen Zhu, Yu Zhou, Yuandong Pu, Jiarui Wu, Jiaquan Wang, Bo Qu, Wenhai Wang, Yu Qiao, Dajuin Yao†, Yihao Liu†

University of Science and Technology of China, Shanghai AI Laboratory, China Academy of Art, Peking University

† Corresponding Authors

📰 News & Updates

🚀 Dec 29, 2025 🔥 Building upon on ArtiMuse, we introduce UniPercept, a comprehensive follow-up work that provides a meticulous study on perceptual-level image understanding. It spans Image Aesthetics Assessment (IAA), Image Quality Assessment (IQA), and Image Structure & Texture Assessment (ISTA) across both Visual Rating (VR) and Visual Question Answering (VQA) tasks.
- Technical Report
- Project Page
- UniPercept-Bench: A comprehensive perceptual-level understanding benchmark for MLLMs, spanning IAA, IQA, and ISTA across VR and VQA tasks.
- UniPercept: A powerful baseline MLLM specialized for perceptual image understanding, optimized via Domain-Adaptive Pre-Training and Task-Aligned RL.
🚀 Dec 29, 2025 The test set of the ArtiMuse-10K Dataset is now available! 🚀
🚀 Sep 3, 2025
The Checkpoints and Evaluation Code of ArtiMuse are now available! 🚀
🚀 July 28, 2025
ArtiMuse was officially released at WAIC 2025, in the forum "Evolving with AI: The Iteration and Resilience of Artistic Creativity"
🚀 July 24, 2025
The Online Demo is now open for public access!
🚀 July 21, 2025
The Paper, Repository and Project Page are now live!

🔍 Abstract

The rapid advancement of educational applications, artistic creation, and AI-generated content (AIGC) technologies has substantially increased practical requirements for comprehensive Image Aesthetics Assessment (IAA), particularly demanding methods capable of delivering both quantitative scoring and professional understanding.

In this paper, we present:
(1) ArtiMuse, an innovative MLLM-based IAA model with Joint Scoring and Expert-Level Understanding capabilities;
(2) ArtiMuse-10K, the first expert-curated image aesthetic dataset comprising 10,000 images spanning 5 main categories and 15 subcategories, each annotated by professional experts with 8-dimensional attributes analysis and a holistic score.

💾 Dataset

The test set of the ArtiMuse-10K is available at ArtiMuse-10K.

The ArtiMuse-10K dataset is available for academic research. By requesting access, you agree to the following Terms of Use:

Non-commercial: Research use only. Commercial use is prohibited.
No Redistribution: Do not share or distribute the dataset to any third party.
Attribution: Properly cite or credit ArtiMuse in any resulting work.

To request the download link, please fill out the following application form: ArtiMuse-10K Access Request Form

📦 Checkpoints

All paper-version checkpoints share the same text pretraining process, but differ in their score finetuning datasets:

Checkpoint	Score Finetuning Dataset	Download	Notes
`ArtiMuse`	ArtiMuse-10K	🤗HF link 🤖MS link	Paper Version (Recommended)
`ArtiMuse_AVA`	AVA	🤗HF link 🤖MS link	Paper Version
`ArtiMuse_FLICKR-AES`	FLICKR-AES	🤗HF link 🤖MS link	Paper Version
`ArtiMuse_PARA`	PARA	🤗HF link 🤖MS link	Paper Version
`ArtiMuse_TAD66K`	TAD66K	🤗HF link 🤖MS link	Paper Version
`ArtiMuse_OnlineDemo`	ArtiMuse-10K & Internal Datasets	—	Surpasses paper versions thanks to additional internal datasets and advanced training; also supports fine-grained attribute scores. For access, please contact us for business collaboration.
`ArtiMuse-R1`	—	—	Next-generation model trained with GRPO, supporting CoT reasoning, delivering more accurate score predictions, and extending beyond IAA to handle a wider range of tasks.

⚙️ Setup

Clone this repository:

git clone https://github.com/thunderbolt215/ArtiMuse.git

Create a conda virtual environment and activate it: (please ensure that Python>=3.9).

conda create -n artimuse python=3.10
conda activate artimuse

Install dependencies using requirements.txt:

pip install -r requirements.txt

We recommend to use FlashAttention for acceleration:

pip install flash-attn --no-build-isolation

📊 Evaluation

1. Prepare Checkpoints

Download the pretrained checkpoints and place them under the checkpoints/ directory. The folder structure should look like:

ArtiMuse
└── checkpoints/
    ├── ArtiMuse
    ├── ArtiMuse_AVA
    ├── ArtiMuse_FLICKR-AES
    ├── ...

2. Evaluation on a Single Image

Run the following command to evaluate a single image:

bash

python src/eval/eval_image.py \
    --model_name ArtiMuse \
    --image_path example/test.jpg \
    --device cuda:0

Arguments
- --model_name: Name of the checkpoint to use (e.g., ArtiMuse, ArtiMuse_AVA).
- --image_path: Path to the input image.
- --device: Inference device, e.g., cuda:0.
Results are saved to:

results/image_results/{input_image_name}_{model_name}_eval.json

3. Evaluation on Benchmark Datasets

Download the test datasets and organize them under test_datasets/{dataset_name}/images/. The expected structure is:

ArtiMuse
└── test_datasets/
    ├── AVA
    │   ├── images/
    │   └── test.json
    ├── TAD66K
    ├── FLICKR-AES
    └── ...

images/: contains the test images.
test.json: provides the ground-truth scores (gt_score) for evaluation.

Run dataset-level evaluation with: