Awesome

Stable Diffusion with Promptable Image Generation on M1 Mac's using DeepCache

Authors: Yiming Zheng (yiming.zheng@mail.utoronto.ca), Xinran Zhang (xinran.zhang@mail.utoronto.ca), Barry Xue (zexin.xue@mail.utoronto.ca)

Description

This repository implements a version of the DeepCache algorithm in MLX.

We based our implementation on these two official repositories from Apple and the DeepCache author.

• DeepCache: https://github.com/horseee/DeepCache/tree/master.
• MLX Stable Diffusion Example: https://github.com/ml-explore/mlx-examples/tree/main/stable_diffusion.

MLX is a DL framework developed by Apple to accelerate model inference on M1 architecture through the METAL API. An advantage MLX has over other frameworks is that MLX supports memory sharing between the M1 CPU and GPU, and the code evaluates lazily.

Our goal is to use this as a proof-of-concept to demonstrate how MLX can be combined with DeepCache to enable fast, on-device inference.

Running our Code

Setup

All dependencies are in the requirements.txt file located at the root directory.

You can install them with:

pip install -r requirments.txt

Structure

We proved the original MLX Stable Diffusion demonstration in mlx_stable_diffusion, and our implementation with DeepCache based on the official demo in mlx_deepcache.

Both directories follow the same structure and can be run with the same set of commands.

Example

To run the official text-to-image example with a prompt run the following command:

cd mlx_stable_diffusion
python txt2image.py "A photo of an astronaut riding a horse on Mars." --n_images 4 --n_rows 2

To run the same example with DeepCache, run these commands instead:

cd mlx_deepcache
python txt2image.py "A photo of an astronaut riding a horse on Mars." --n_images 4 --n_rows 2 --cache_interval 3

For further information on running another type of experiment, please refer to either mlx_stable_diffusion/README.md or mlx_deepcache/txt2image.py.

Currently, only Stable Diffusion is supported (run this model with the --model sd flag) for the DeepCache version, and SDXL Turbo is still under development.