国内的小伙伴可以看这 This repo contains the official implementation of the AAAI 2026 paper.
Image Aesthetics Assessment (IAA) evaluates visual quality through user-centered perceptual analysis and can guide various applications. Recent advances in Multimodal Large Language Models (MLLMs) have sparked interest in adapting them for IAA. However, two critical limitations persist in applying MLLMs to IAA:
- the tokenization strategy leads to insensitivity to scores
- the classification-based decoding mechanisms introduce score quantization errors
- Current MLLM-based IAA methods treat the task as coarse rating classification followed by probability-to-score mapping, which loses fine-grained information.
To address these challenges, we propose ROC4MLLM, offering complementary solutions from two perspectives:
- Representation: We separate scores from the word token space to avoid tokenizing scores as text. An independent position token bridges these spaces, improving the model’s sensitivity to score positions in text.
- Computation: We apply distinct loss functions for text and score predictions to enhance the model’s sensitivity to score gradients. Decoupling scores from text ensures effective supervision while preventing interference between scores and text in the loss computation.
Extensive experiments across five datasets demonstrate ROC4MLLM’s state-of-the-art performance without requiring additional training data. Additionally, ROC4MLLM’s plug-and-play design ensures seamless integration with existing MLLMs, boosting their IAA performance.
- Download the weight from Baidu Netdisk or HuggingFace.
- Or Download the Docker environment from Baidu Netdisk.
- Clone this repository and navigate to ROC4MLLM folder
git clone https://github.com/woshidandan/Assessing-Image-Aesthetics-via-Multimodal-Large-Language-Models.git
cd ROC4MLLM- Install Package
conda create -n roc4mllm python=3.10 -y
conda activate roc4mllm
pip install --upgrade pip
pip install -e .- Install additional packages for training cases
pip install -e ".[train]"
pip install flash-attn --no-build-isolation
from mplug_owl2.assessor import Assessment
from PIL import Image
assessment=Assessment(pretrained="Ricardo-M/ROC4MLLM")
images=["test_images/1_-10.jpg","test_images/1_-10.jpg"]
input_img=[]
for image in images:
img=Image.open(image).convert('RGB')
input_img.append(img)
answer=assessment(input_img,precision=4)
print(answer)Please refer to mPLUG-Owl2 for data preparation.
Notes: We have added a gt_score field. If you intend to use CE loss or EMD loss, the target field is required.
Below is an example of a data sample in AVA:
{
"image": "771257.jpg",
"gt_score": 3.463414634146341,
"conversations": [{"from": "human", "value": "<|image|>Could you evaluate the aesthetics of this image?"}, {"from": "gpt", "value": "The aesthetic rate of the image is [SCORE]. "}],
"target": [0.15853658536585366, 0.10975609756097561, 0.2073170731707317, 0.2926829268292683, 0.16463414634146342, 0.04878048780487805, 0.0, 0.0, 0.006097560975609756, 0.012195121951219513]
}Place your data file path in the DATA_FILE within scripts/finetune.sh. You also need to update the Image_root in the same script to point to the directory where your original images are stored.
Download the pretrained model checkpoints and update the LOAD in scripts/finetune.sh accordingly.
Run the following command to start training:
bash scripts/finetune.sh
You can modify min_score and max_score to define the score range in your dataset. Use l1_weight, ce_weight, and emd_weight to configure the loss functions and their respective weights for the score loss.
Important Note: If you use CE or EMD loss, ensure that the num_tokens matches the length of the target field in your training data.
@inproceedings{ma2026regression,
title={Regression over Classification: Assessing Image Aesthetics via Multimodal Large Language Models},
author={Ma, Xingyuan and He, Shuai and Ming, Anlong and Zhong, Haobin and Ma, Huadong},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={40},
number={10},
pages={7827--7835},
year={2026}
}
