StructAdapt is the companion open-source project to the paper
AdaConv: Structure-Aware Real-Time Style Transfer Algorithm and Its Rendering Implementation in Unity Engine.
This repository provides the engineering implementation of the method described in the paper — exporting the AdaConv style-transfer algorithm to the ONNX format and deploying it inside the Unity Engine.
It addresses the structural incompatibility between AdaConv’s dynamic kernel generation and static-graph inference frameworks by introducing a grouped-convolution vectorization strategy.
Three exportable model variants are derived from the same AdaConv checkpoint:
- Fully Dynamic Model – preserves both spatial and batch dimension dynamics.
- Batch-Dynamic Model – fixes spatial dimensions while keeping batch dynamics.
- Static Model – adopts a completely static computation graph.
All variants achieve real-time rendering of approximately 14.5 FPS on RTX 4060, and SSIM ≈ 0.67–0.68, demonstrating structural equivalence to the original AdaConv formulation.
For detailed methodology and experiments, please refer to the published paper.
This project requires Python ≥3.10.
All dependencies are listed in StyleAdapt-Core/config/requirements.txt.
cd StyleAdapt-Core/config
pip install -r requirements.txtKey libraries include:
- torch, onnx, onnxruntime-gpu
- tensorboard, PyYAML, scikit-image
- matplotlib, tqdm
The project was developed and verified under the following environment:
nvcc: NVIDIA (R) Cuda compiler driver
Build cuda_12.9.r12.9/compiler.35813241_0
Cuda compilation tools, release 12.9, V12.9.41
NVIDIA-SMI 581.29
Driver Version: 581.29
CUDA Version: 13.0
GPU: NVIDIA GeForce RTX 4060 Laptop GPU
Compute Mode: Default
VRAM: 8 GB
cuDNN Version: 9.10.2
Supported SM Architectures: up to 12.0Ensure your CUDA and cuDNN versions are consistent with these configurations to achieve optimal performance.
- Unity 2022.3.55f1c1
- NuGet Version:4.4.0
- Microsoft.ML.OnnxRuntime.Managed Version:1.20.0
- Platform: Windows 11 + RTX GPU (CUDA 12.9 / 13.0)
python train.py -c ./configs/lambda100.yaml -d ./data/raw -l ./logspython train.py -c ./configs/lambda100.yaml -d ./data/finetune --finetunetensorboard --logdir=./logs/tensorboard
python visualizer.pyExport different structural variants from the same AdaConv checkpoint:
python exporter.py --output model.onnx
python exporter.py --output model_static.onnx --static
python exporter.py --output model_dyn.onnx --dynamic
python exporter.py --output model_batch.onnx --dynamic-batchAll hyperparameters are defined in config/lambda100.yaml, including:
- Convolution kernel layout and dynamic mapping rules
- Loss coefficients (lambda_content, lambda_style, etc.)
- Dataset paths and normalization pipeline
- Training and inference batch parameters
You can duplicate and modify this YAML file to customize different training schemes.
If you find this repository useful in your research or projects, please cite:
@inproceedings{He2025,
title={AdaConv: Structure-Aware Real-Time Style Transfer Algorithm and Its Rendering Implementation in Unity Engine},
author={Bin He},
booktitle={Proceedings of the 2025 3rd International Conference on Image, Algorithms, and Artificial Intelligence (ICIAAI 2025)},
pages={1029--1042},
year={2025},
publisher={Atlantis Press},
doi={10.2991/978-94-6463-823-3_100},
url={https://doi.org/10.2991/978-94-6463-823-3_100},
issn={2352-538X}
}
This project is released under the MIT License.
See LICENSE for details.
- To reproduce the paper’s performance, ensure CUDA and cuDNN versions match those listed above.
- ONNX export supports both static and dynamic kernel graphs; static mode is recommended for Unity runtime integration.
- The project is optimized for RTX 4060 series GPUs but remains portable to other Ampere and Ada Lovelace architectures.