Code for the paper [arXiv]

Transformed Low-rank Adaptation via Tensor Decomposition and Its Applications to Text-to-image Models (ICCV-25)

by Zerui Tao, Yuhta Takida, Naoki Murata, Qibin Zhao, and Yuki Mitsufuji

The code for subject-driven generation and controllable generation experiments are stored in ./subject-driven-generation and ./control respectively.

• Dataset: The dataset can be downloaded from here. Please store the dataset in ./subject-driven-generation/dataset/Customization.
• Experiment: cd ./subject-driven-generation && bash run_tlora.sh.

We follow the guideline of OFT for data processing, training, and evaluation. After downloading and processing the datasets, please run the following example scripts for exepriments:

• bash tt_control_build.sh: Build the model.
• bash tt_control_train.sh: Train the model with images and controlling signals.
• bash tt_control_test.sh: Generate test images given test controlling signals.
• python eval_{canny,landmark,segm}.sh: Evaluations for corresponding tasks.

Our codebase is implemented based on the following projects. Thanks for their contributions.

• DCO: https://github.com/kyungmnlee/dco
• SODA: https://github.com/phymhan/SODA-Diffusion
• OFT: https://github.com/zqiu24/oft

@inproceedings{tao2025transformed,
  title={Transformed Low-rank Adaptation via Tensor Decomposition and Its Applications to Text-to-image Models},
  author={Tao, Zerui and Takida, Yuhta and Murata, Naoki and Zhao, Qibin and Mitsufuji, Yuki},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  month={October},
  year={2025}
}