DeepCache: Accelerating Diffusion Models for Free

(Results on Stable Diffusion v1.5. Left: 50 PLMS steps. Right: 2.3x acceleration upon 50 PLMS steps)

DeepCache: Accelerating Diffusion Models for Free
Xinyin Ma, Gongfan Fang, Xinchao Wang
Learning and Vision Lab, National University of Singapore
🥯[Arxiv]🎄[Project Page]

Why DeepCache

• 🚀 Training-free and almost lossless
• 🚀 Support Stable Diffusion, Stable Diffusion XL, Stable Video Diffusion, Stable Diffusion Pipeline / XL Pipeline for Inpainting, Stable Diffusion Img2Img Pipeline, DDPM
• 🚀 Compatible with sampling algorithms like DDIM and PLMS

Updates

• June 27, 2024: 🔥Our new work AsyncDiff enables parallel inference of diffusion models on multiple GPUs. Check our paper and code!

• June 5, 2024: 🔥Our new work, Learning-to-Cache, an improved version of DeepCache on DiT. Code and checkpoints are released.

• January 5, 2024: 💥A doc page for DeepCache has been added in Diffusers! Check here for more information. Many thanks to the Diffusers team!

• December 26, 2023: 🔥Update a plug-and-play implementation of DeepCache, no longer requiring any modifications of the diffuser's code! Check here for the detailed usage! Big thanks to @yuanshi9815 for contributing the code!

• December 25, 2023: A demo is available via Colab

• December 21, 2023: Release the code for Stable Video Diffusion and Text2Video-Zero. In the figure below, the upper row shows the original videos generated by SVD-XT, and the lower row is accelerated by DeepCache. For Text2Video-Zero, the results can be found here

  
  (1.7x acceleration of SVD-XT)

• December 20, 2023: Release the code for DDPM. See here for the experimental code and instructions.

• December 6, 2023: Release the code for Stable Diffusion XL. The results of the stabilityai/stable-diffusion-xl-base-1.0 are shown in the below figure, with the same prompts from the first figure.

  
  (2.6x acceleration of Stable Diffusion XL)

Introduction

We introduce DeepCache, a novel training-free and almost lossless paradigm that accelerates diffusion models from the perspective of model architecture. Utilizing the property of the U-Net, we reuse the high-level features while updating the low-level features in a very cheap way. DeepCache accelerates Stable Diffusion v1.5 by 2.3x with only a 0.05 decline in CLIP Score, and LDM-4-G(ImageNet) by 4.1x with a 0.22 decrease in FID.

Quick Start

Install

pip install DeepCache

Usage

import torch

# Loading the original pipeline
from diffusers import StableDiffusionPipeline
pipe = StableDiffusionPipeline.from_pretrained('runwayml/stable-diffusion-v1-5', torch_dtype=torch.float16).to("cuda:0")

# Import the DeepCacheSDHelper
from DeepCache import DeepCacheSDHelper
helper = DeepCacheSDHelper(pipe=pipe)
helper.set_params(
    cache_interval=3,
    cache_branch_id=0,
)
helper.enable()

# Generate Image
deepcache_image = pipe(
        prompt,
        output_type='pt'
).images[0]
helper.disable()

We here take the Stable Diffusion pipeline as an example. You can replace pipe with any variants of the Stable Diffusion pipeline, including choices like SDXL, SVD, and more. You can find examples in the script. The argument cache_branch_id specifies the selected skip branch. For the skip branches that are deeper, the model will engage them only during the caching steps, and exclude them during the retrieval steps. The argument cache_interval represents the interval for updating the cache.

A general script for SD

python main.py --model_type sdxl #Support [sdxl, sd1.5, sd2.1, svd, sd-inpaint, sdxl-inpaint, sd-img2img]

Experimental code for DeepCache

The above implementation does not require changes to the forward or __call__ functions in the Diffusers pipeline, and is, therefore, more general. The following section is the experimental code that can be used to reproduce the results in the paper. It is implemented one by one for different model structures and pipelines, and thus, may not work properly due to the update of diffusers.

Setup

pip install diffusers==0.24.0 transformers

Stable Diffusion XL

python stable_diffusion_xl.py --model stabilityai/stable-diffusion-xl-base-1.0

You can add --refine at the end of the command to activate the refiner model for SDXL.

Stable Diffusion v1.5

python stable_diffusion.py --model runwayml/stable-diffusion-v1-5

Stable Diffusion v2.1

python stable_diffusion.py --model stabilityai/stable-diffusion-2-1

Currently, our code supports the models that can be loaded by StableDiffusionPipeline. You can specify the model name by the argument --model, which by default, is runwayml/stable-diffusion-v1-5.

Stable Video Diffusion

python stable_video_diffusion.py

DDPM and LDM

Please check here for the experimental code of DDPM and LDM.

Visualization

Images in the upper line are the baselines and the images in the lower line are accelerated by DeepCache.

Stable Diffusion v1.5 (2.15x Acceleration)

LDM-4-G for ImageNet (6.96x Acceleration)

DDPM for LSUN Church & Bedroom (1.48x Acceleration)

Quantitative Results

ImageNet

Stable Diffusion v1.5

More results can be found in our paper

Other Implementations and Plugins

We sincerely thank the authors listed below who implemented DeepCache in plugins or other contexts.

• Diffusers: https://huggingface.co/docs/diffusers/main/en/optimization/deepcache
• OneDiff Integration: https://github.com/siliconflow/onediff?tab=readme-ov-file#easy-to-use by @SiliconFlow. OneDiff also has implementations for DeepCache on SVD, check this for details.
• Comfyui: https://gist.github.com/laksjdjf/435c512bc19636e9c9af4ee7bea9eb86 by @laksjdjf
• Colab & Gradio: https://github.com/camenduru/DeepCache-colab by @camenduru
• WebUI: https://github.com/AUTOMATIC1111/stable-diffusion-webui/pull/14210 by @aria1th
• SD.Next: https://github.com/vladmandic/automatic by @vladmandic
• MLX_deepcache for running SD on M1 Mac: https://github.com/Barry0121/mlx_deepcache by @Barry Xue

We warmly welcome contributions from everyone. Please feel free to reach out to us.

Bibtex

@inproceedings{ma2023deepcache,
  title={DeepCache: Accelerating Diffusion Models for Free},
  author={Ma, Xinyin and Fang, Gongfan and Wang, Xinchao},
  booktitle={The IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2024}
}