STAR: Incremental Predictive Quantization for Style-conditioned Test-time Adaptive Refinement

This is the official PyTorch implementation for the paper "STAR: Style-conditioned Test-time Adaptive Refinement".

📖 Introduction

STAR is a plug-and-play, test-time refinement framework that enhances existing artistic style transfer pipelines using an expandable target-domain style memory, achieving high-fidelity and scalable stylization without modifying the frozen base generator.

🚀 News & Updates

• [2026.05] The official codebase of STAR is released!

🛠️ Installation

1. Clone the repository

git clone https://github.com/IIP-Lab-XDU/STAR.git
cd STAR

2. Environment Setup

We provide an environment.yaml file for easy reproduction. We recommend using Conda:

conda env create -f environment.yaml
conda activate STAR

📂 Datasets

Face Style Transfer Datasets

Our face stylization experiments are conducted on the following publicly available datasets:

• MetFaces https://github.com/NVlabs/metfaces-dataset

• APDrawing https://github.com/yiranran/APDrawingGAN2

• FS2K https://github.com/DengPingFan/FS2K

Scene Style Transfer Datasets

For scene stylization, the artistic style images are collected from WikiArt:

• WikiArt https://www.wikiart.org/

🏋️ Training

You can start training the STAR model by running the provided shell script. Configuration details can be modified in the configs/ directory.

bash train.sh

🧪 Testing & Evaluation

To test the model with test-time adaptive refinement:

bash test.sh