Katz is a high-performance serving system designed specifically for diffusion model workflows with multiple adapters. It dramatically improves inference efficiency while maintaining image quality.
- ControlNet-as-a-Service: Decouples ControlNets from the base model for independent scaling.
- Bounded Asynchronous LoRA Loading: Overlaps LoRA loading with base model execution for reduced latency.
- Latent Parallelism: Accelerates base model execution across multiple GPUs.
-
Performance Gains: Up to
$7.8 \times$ latency reduction and$1.7 \times$ throughput improvement.
Prompt: papercut -subject/scene-a shiba inu wearing a beret and black turtleneck, 4k, clean background
Negative prompt: low quality, bad quality, sketches, numbers, letters
This image was generated with 1 ControlNet with depth guidance and 1 LoRA for the papercut style. The depth reference image used for guidance is available here.
- NVIDIA GPUs (H800 recommended for best performance)
- CUDA 11.8+
- Python 3.10+
$ conda create -n katz python=3.10
$ conda activate katz
$ pip install -r requirements.txt
# Install our customized diffusers package
$ pushd ./diffusers-hf && pip install -e . && popd
# Install fast-kernel
$ pushd ./diffusers-hf/src/fast_kernel/ && git submodule update --init --recursive && pip install . && popd
Coming soon.
For detailed benchmarking instructions and reproducing our results, see the artifact evaluation guide.
We provide tools and datasets for analyzing real-world production traces in the trace directory.
Please cite our paper if it is helpful to your research.
@inproceedings{Katz2025,
title = {Katz: Efficient Workflow Serving for Diffusion Models with Many Adapters},
author = {Li, Suyi and Yang, Lingyun and Jiang, Xiaoxiao and Lu, Hanfeng and An, Dakai and Di, Zhipeng and Lu, Weiyi and Chen, Jiawei and Liu, Kan and Yu, Yinghao and Lan, Tao and Yang, Guodong and Qu, Lin and Zhang, Liping and Wang, Wei},
booktitle = {Proc. USENIX ATC},
year = {2025}
}
We thank the contributors of 🤗 Diffusers for their foundational work.
For questions and support, please open an issue or contact the authors.