News

• [2025/11/07] Our paper is accepted by AAAI-26 and selected as Oral 🎉!
• [2025/08/06] Our paper is available on arXiv.
• [2025/08/14] We release our code on Github.

Setup

Setup Base Environment

Before you start, please create a fresh environment:

conda create -n st-inference python=3.10

Inference Environment

First, because of some dirty imports, we need to manually install a specific version of torch and torchvision.

pip install torch==2.6.0 torchvision==0.21.0

After that, we need to manually install an old package.

pip install basicsr==1.3.5 --no-build-isolation

The issue is that basicsr imports torch inside its build wheel......

Then, for another specialized metrics module:

pip install t2v_metrics==1.2.0

Additionally, install other packages in requirements.txt.

pip install -r requirements.txt

Finally, we need to manually install CLIP.

pip install git+https://github.com/openai/CLIP.git

You might still see some red warnings popping up. However, after all operations above, we should be having a clear environment for you to run inference with. You may notice that we installed torch==2.6.0 then torch==2.5.1 then torch==2.6.0. This is to mitigate an inherent bug in torch==2.5.1 where we cannot build basicsr with it. Nevertheless, we need to compile t2v_metrics to torch==2.5.1, but we can't run with torch==2.5.1, so we need to reinstall torch==2.6.0 in the end.

Evaluation Environment

Next, we need to checkout a new environment to run eval scripts. You can start by cloning the original environment:

conda create -n st-evaluation --clone st-inference
conda activate st-evaluation

After that, we need to install pyiqa.

pip install pyiqa

Now, you can run utils/eval.py to evaluate your results.

Downloading Weights

Downloading the humanparsing and openpose weights from this 🤗 Hugging Face link

Setup API Key

• Select the platform from which you want to call the API (e.g., Qwen, OpenRouter).

• Apply for an API key following the instructions on their website.

• Write the API key to your environment variables.

Setup Google Search Engine

• Create your own project in Google Cloud, and within that project, request an API key and simultaneously enable the Custom Search API service.

• Create and configure your Programmable Search Engine and remember your custom ID.

• For more detailed information, please refer to this document.

Inference

conda activate styletailor
python pipeline.py

Eval

conda activate styletailor_eval
cd /code/utils
python eval.py

Citation

@misc{ma2025styletailorpersonalizedfashionstyling,
      title={StyleTailor: Towards Personalized Fashion Styling via Hierarchical Negative Feedback}, 
      author={Hongbo Ma and Fei Shen and Hongbin Xu and Xiaoce Wang and Gang Xu and Jinkai Zheng and Liangqiong Qu and Ming Li},
      year={2025},
      eprint={2508.06555},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.06555}, 
}