🚀 DiffuseKronA (WACV-25 🎉)
webpage | paper | video | dataset | weights | demo (soon!)

💡 Highlight

✔️ Parameter Efficient: A minimum 35% reduction in parameters. By changing Kronecker factors, we can even achieve up to a 75% reduction with results comparable to LoRA-DreamBooth.
✔️ Enhanced Stability: Our method is more stable compared to LoRA-DreamBooth. Stability refers to variations in images generated across different learning rates and Kronecker factor/ranks, which makes LoRA-DreamBooth harder to fine-tune.
✔️ Text Alignment and Fidelity: On average, DiffusekronA captures better subject semantics and large contextual prompts.
✔️ Interpretability: Leverages the advantages of the Kronecker product to capture structured relationships in attention-weight matrices. More controllable decomposition makes DiffusekronA more interpretable.

⭐ Method Details

Overview of DiffuseKronA:
✨ Fine-tuning process involves optimizing the multi-head attention parameters (Q, K, V , and O) using Kronecker Adapter, elaborated in the subsequent blocks.
✨ During inference, newly trained parameters, denoted as θ, are integrated with the original weights Dϕ and images are synthesized using the updated personalized model D_ϕ+θ.
✨ We also present a schematic illustration of LoRA vs DiffuseKronA; LoRA is limited to one controllable parameter, the rank r; while the Kronecker product showcases enhanced interpretability by introducing two controllable parameters a₁ and a₂ (or equivalently b₁ and b₂). Furthermore, we also showcase the advantages of the proposed method.

🛠️ Installation Steps

1. Create conda environment

conda create -y -n diffusekrona python=3.11
conda activate diffusekrona

2. Package installation

pip install diffusers==0.21.0
pip install -r requirements.txt
pip install accelerator

3. Install CLIP

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

🔥 Quickstart

Note: For diffusers=0.21.0, you will get ImportError: cannot import name 'cached_download' from 'huggingface_hub' error. To solve it please remove the line from huggingface_hub import HfFolder, cached_download, hf_hub_download, model_info in dyanamic_models_utils.py script.

1. Clone the dataset and remove the *subject/generated subfolders

git clone https://github.com/diffusekrona/data && rm -rf data/.git
mkdir outputs
cd diffusekrona/
python format_datasets.py       # To format the dataset (NOT mandatory)

2. Finetune diffusekrona using script file

cd diffusekrona/                                        # RUN inside diffusekrona folder
CUDA_VISIBLE_DEVICES=$GPU_ID bash scripts/finetune_sdxl.sh      # Leveraging SDXL model
CUDA_VISIBLE_DEVICES=$GPU_ID bash scripts/finetune_sd.sh        # Leveraging SDXL model

3. Generate images from the finetuned weights (RUN inside diffusekrona folder)

CUDA_VISIBLE_DEVICES=$GPU_ID accelerate launch scripts/inference_sdxl.sh    # Leveraging SDXL model
CUDA_VISIBLE_DEVICES=$GPU_ID accelerate launch scripts/inference_sd.sh      # Leveraging SD model

Note: Specify a single GPU index only (e.g., CUDA_VISIBLE_DEVICES=0) and avoid listing multiple IDs.

🎖️ Results

Generation Results on Human Faces 🗿

Generation Results on Animal (Cat), Teddy Bear, and Shoes

Generation Results on Toy, Teddy Bear, and Anime Character

Generation Results on Anime Characters and Animal (Cat)

Generation Results on Car modifications and showcase 🚘

One-shot Image Generation Results on HuggingFace 🤗

For more results, please visit here.

🙏🏼 Acknowledgement

Our codebase is built on top of the HuggingFace Diffusers library, and we’re incredibly grateful for their amazing work!

✏️ Citation

If you think this project is helpful, please feel free to leave a star⭐️ and cite our paper:

@InProceedings{Marjit_2025_WACV,
    author    = {Marjit, Shyam and Singh, Harshit and Mathur, Nityanand and Paul, Sayak and Yu, Chia-Mu and Chen, Pin-Yu},
    title     = {DiffuseKronA: A Parameter Efficient Fine-Tuning Method for Personalized Diffusion Models},
    booktitle = {Proceedings of the Winter Conference on Applications of Computer Vision (WACV)},
    month     = {February},
    year      = {2025},
    pages     = {3529-3538}
}

✉️ Contact

Shyam Marjit: marjitshyam@gmail.com or shyam.marjit@iiitg.ac.in