SADA: Stability-guided Adaptive Diffusion Acceleration

Fig. 1. Accelerating Flux, SDXL, SD-2 by 2.02×,1.86×,1.80× with Stability-guided Adaptive Diffusion Acceleration with 50 inference steps.

🔨 Installation

SADA plugs straight into any project built on HuggingFace Diffusers🤗. To start with a new environment, set up and running in two quick steps:

1. Create and activate a new conda environment:

git clone https://github.com/Ting-Justin-Jiang/sada-icml.git
conda create -n sada python=3.10
conda activate sada

2. Install the required packages:

pip install -r requirements.txt

---

🚀 Quickstart

We provide the following demos to test SADA with SD-2, SD-XL, and Flux architecture. Simply run:

python sd_demo.py 

python xl_demo.py 

python flux_demo.py 

---

with --solver {dpm|euler}， --prompt, and --seed

For any 🤗diffuser-based environment, SADA could be applied and enabled by a single configuration call 🔥🔥🔥:

patch.apply_patch(pipe,
                    sx=3, sy=3,
                    max_downsample=1,
                    acc_range=(10, 47),

                    lagrange_int=4,
                    lagrange_step=24,
                    lagrange_term=4,

                    max_fix=1024 * 5,
                    max_interval=4
                  )

Finetuning: If you have a LoRA checkpoint, uncomment the relevant lines in the demo scripts and set lora_path to your file. You can also swap the default pretrained models for any fine‑tuned variants sharing the same backbone.

---

Fig. 2. Overview of SADA pipeline.

📕 Citation

If you find this work useful, please cite our paper:

@inproceedings{jiang2025sada,
  title     = {SADA: Stability-guided Adaptive Diffusion Acceleration},
  author    = {Ting Jiang and Yixiao Wang and Hancheng Ye and Zishan Shao and Jingwei Sun and Jingyang Zhang and Zekai Chen and Jianyi Zhang and Yiran Chen and Hai Li},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning},
  year      = {2025}
}

---

🍾 Acknowledgement

SADA codebase is build upon the excellent work of Huggingface Diffuser and ToMeSD