ssssmark commited on
Commit
c4f02ec
·
verified ·
1 Parent(s): d1bc5a6

Create README.md

Browse files

# Unlocking the Essence of Beauty: Advanced Aesthetic Reasoning with Relative-Absolute Policy Optimization
<a href="https://arxiv.org/pdf/2509.21871" target="_blank">
<img alt="arXiv" src="https://img.shields.io/badge/arXiv-Aes--R1-red?logo=arxiv" height="25" />
</a>
<a href="https://huggingface.co/ssssmark/Aes-R1" target="_blank">
<img alt="HF Model: Aes-R1" src="https://img.shields.io/badge/%F0%9F%A4%97%20Model-Aes--R1-ffc107" height="25" />
</a>
<a href="https://huggingface.co/TianheWu/VisualQuality-R1-7B-preview" target="_blank">
<img alt="HF Dataset : Aes-CoT" src="https://img.shields.io/badge/%F0%9F%A4%97%20Dataset-Aes--CoT-ffc107" height="25" />
</a>
</div>





> A novel and effective reinforcement learning framework designed for Image Aesthetic Assessment and general open-ended preference evaluation.
# 🖥️Training
## Preparation
1. First download the IAA datasets and place them all in a single folder.
2. Construct your image-score dataset in the following format:
```json
{
"messages": [
{
"content": "prompt here",
"role": "user"
},
{
"content": "response here",
"role": "assistant"
}
],
"images": "image_path_1"
},
```
we provide an example dataset in `AesR1/data` folder.
3. Download the pre-trained model weights from [here](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) and place them in `AesR1/models`

## Cold-start
We use [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) to train the SFT model.

1. Clone the [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) repository and install the dependencies.

```bash
git clone https://github.com/hiyouga/LLaMA-Factory.git
conda create -n coldstart python=3.11.10
conda activate coldstart
cd LLaMA-Factory
pip install -e ".[torch,metrics]"
```
2. Put your cot dataset info in `LLaMA-Factory/data/dataset_info.json` and move `qwen_aescot.yaml` into `LLaMA-Factory/examples/train_full`
3. Run the following command to train the SFT model.

```bash
llamafactory-cli train examples/train_full/qwen_aescot.yaml
```

## RAPO
First setup the environment for RAPO training.
```bash
conda create -n rapo python=3.11.10
conda activate rapo
bash setup.sh
```
After modification, run the following command to train the RAPO model.
```bash
# For single node training
bash train/rapo/src/open-r1-multimodal/run_scripts/Aes/aes_onenode.sh

# For multi node training
bash train/rapo/src/open-r1-multimodal/run_scripts/Aes/aes_multinode.sh
```

# Inference
After training, you can inference the model by using the scripts in LLaMA-Factory.

```bash
#Install vllm
pip install vllm

#Infer
python scripts/vllm_infer.py \
--model_name_or_path [path/to/your/model] \
--dataset [dataset_name] \
--template qwen2_vl \
--save_name result.jsonl \
--temperature 0.6 \
```


# 📚 Citation
If you find this repo useful, please consider citing our paper as follows:
```


@misc
{liu2025unlockingessencebeautyadvanced,
title={Unlocking the Essence of Beauty: Advanced Aesthetic Reasoning with Relative-Absolute Policy Optimization},
author={Boyang Liu and Yifan Hu and Senjie Jin and Shihan Dou and Gonglei Shi and Jie Shao and Tao Gui and Xuanjing Huang},
year={2025},
eprint={2509.21871},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2509.21871},
}
```

Files changed (1) hide show
  1. README.md +11 -0
README.md ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ datasets:
4
+ - ssssmark/AesCoT
5
+ metrics:
6
+ - spearmanr
7
+ - pearsonr
8
+ base_model:
9
+ - Qwen/Qwen2.5-VL-7B-Instruct
10
+ pipeline_tag: reinforcement-learning
11
+ ---