10 KiB
Table of Contents
Tested on MacBook Air M2 with Docker Desktop for Mac with Apple Chip.
Setup
Directly on Apple silicon
For Mac M1/M2. Read more about Metal Performance Shaders (MPS) framework.
Prerequisites
Install the latest versions of macOS, Homebrew, Python, and Git.
brew install cmake protobuf rust
brew install --cask miniconda
conda init zsh && source ~/.zshrc # or bash and .bashrc
Set up
GITHUB_STABLE_DIFFUSION=https://github.com/santisbon/stable-diffusion.git
git clone $GITHUB_STABLE_DIFFUSION
cd stable-diffusion
mkdir -p models/ldm/stable-diffusion-v1/
Go to Hugging Face, and click "Access repository" to Download sd-v1-4.ckpt (~4 GB). You'll need to create an account but it's quick and free. Then set up the environment:
PATH_TO_CKPT="$HOME/Downloads" # or wherever you saved sd-v1-4.ckpt
ln -s "$PATH_TO_CKPT/sd-v1-4.ckpt" models/ldm/stable-diffusion-v1/model.ckpt
# When path exists, pip3 will (w)ipe.
# restrict the Conda environment to only use ARM packages. M1/M2 is ARM-based. You could also conda install nomkl.
PIP_EXISTS_ACTION=w
CONDA_SUBDIR=osx-arm64
conda env create -f environment-mac.yaml && conda activate ldm
You can verify you're in the virtual environment by looking at which executable you're getting:
type python3
Only need to do this once:
python3 scripts/preload_models.py
On a Linux container with Docker for Apple silicon
You can't access the Macbook M1/M2 GPU cores from the Docker containers so performance is reduced but for development purposes it's fine.
Prerequisites
Install Docker
On the Docker Desktop app, go to Preferences, Resources, Advanced. Adjust the CPUs and Memory to the largest amount available to avoid this Issue. You may need to increase Swap and Disk image size too.
Go to Hugging Face, and click "Access repository" to Download sd-v1-4.ckpt (~4 GB) to ~/Downloads.
You'll need to create an account but it's quick and free.
Create a Docker volume for the downloaded model file
docker volume create my-vol
Populate the volume using a lightweight Linux container. You just need to create the container with the mountpoint; no need to run it.
docker create --name dummy --mount source=my-vol,target=/data alpine
cd ~/Downloads # or wherever you saved sd-v1-4.ckpt
docker cp sd-v1-4.ckpt dummy:/data
Launch and set up a container
Start a container for Stable Diffusion
docker run -it \
--platform linux/arm64 \
--name stable-diffusion \
--hostname stable-diffusion \
--mount source=my-vol,target=/data \
debian
# or arm64v8/debian
You're now on the container. Set it up:
apt update && apt upgrade -y && apt install -y \
git \
pip3 \
python3 \
wget
GITHUB_STABLE_DIFFUSION="-b docker-apple-silicon https://github.com/santisbon/stable-diffusion.git"
# you won't need to close and reopen your terminal after this because we'll source our .<shell>rc file
cd /data && wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-aarch64.sh -O anaconda.sh \
&& chmod +x anaconda.sh && bash anaconda.sh -b -u -p /anaconda && /anaconda/bin/conda init bash && source ~/.bashrc
# update conda
conda update -y -n base -c defaults conda
cd / && git clone $GITHUB_STABLE_DIFFUSION && cd stable-diffusion
# When path exists, pip3 will (w)ipe.
# restrict the Conda environment to only use ARM packages. M1/M2 is ARM-based. You could also conda install nomkl.
PIP_EXISTS_ACTION=w
CONDA_SUBDIR=osx-arm64
# Create the environment
# conda env create -f environment.yaml && conda activate ldm
conda create -y --name ldm && conda activate ldm
pip3 install -r requirements-linux-arm64.txt
python3 scripts/preload_models.py
mkdir -p models/ldm/stable-diffusion-v1 \
&& chown root:root /data/sd-v1-4.ckpt \
&& ln -sf /data/sd-v1-4.ckpt models/ldm/stable-diffusion-v1/model.ckpt
[Optional] Face Restoration and Upscaling
cd .. # by default expected in a sibling directory
git clone https://github.com/TencentARC/GFPGAN.git
cd GFPGAN
pip3 install basicsr # used for training and inference
pip3 install facexlib # face detection and face restoration helper
pip3 install -r requirements.txt
python3 setup.py develop
pip3 install realesrgan # to enhance the background (non-face) regions and do upscaling
# pre-trained model needed for face restoration
wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth -P experiments/pretrained_models
cd ..
cd stable-diffusion
python3 scripts/preload_models.py # if not, it will download model files from the Internet the first time you run dream.py with GFPGAN and Real-ESRGAN turned on.
Usage
Startup
With the Conda environment activated (conda activate ldm), run the interactive interface that combines the functionality of the original scripts txt2img and img2img:
Use the more accurate but VRAM-intensive full precision math because half-precision requires autocast and won't work.
By default the images are saved in outputs/img-samples/.
If you're on a docker container set the output dir to the Docker volume.
# If on Macbook
python3 scripts/dream.py --full_precision
# If on Linux container
python3 scripts/dream.py --full_precision -o /data
You'll get the script's prompt. You can see available options or quit.
dream> -h
dream> q
Text to Image
For quick (and rough) results test with 5 steps (default 50), 1 sample image.
Increase steps to 100 or more for good (but slower) results.
The prompt can be in quotes or not.
dream> The hulk fighting with sheldon cooper -s5 -n1
dream> "woman closeup highly detailed" -s 150
# Reuse previous seed and apply face restoration (if you installed GFPGAN)
dream> "woman closeup highly detailed" --steps 150 --seed -1 -G 0.8
TODO: example for upscaling. The -U option currently doesn't work on Mac.
If you're on a container and set the output to the Docker volume (or moved it there with mv outputs/img-samples/ /data/) you can copy it easily wherever you want.
# On your host Macbook (you can use the name of any container that mounted the volume)
docker cp dummy:/data/ ~/Pictures
Image to Image
You can also do text-guided image-to-image translation. For example, turning a sketch into a detailed drawing.
Strength is a value between 0.0 and 1.0, that controls the amount of noise that is added to the input image. Values that approach 1.0 allow for lots of variations but will also produce images that are not semantically consistent with the input. 0.0 preserves image exactly, 1.0 replaces it completely.
Make sure your input image size dimensions are multiples of 64 e.g. 512x512. Otherwise you'll get Error: product of dimension sizes > 2**31'. If you still get the error try a different size like 512x256.
If you're on a docker container, copy your input image into the Docker volume
docker cp ~/Pictures/sketch-mountains-input.jpg dummy:/data/
Try it out generating an image (or 4).
# If you're on your Macbook
dream> "A fantasy landscape, trending on artstation" -I ~/Pictures/sketch-mountains-input.jpg --strength 0.8 --steps 100 -n4
# If you're on a Linux container on your Macbook
dream> "A fantasy landscape, trending on artstation" -I /data/sketch-mountains-input.jpg --strength 0.8 --steps 100 -n1
Web Interface
You can use the script with a graphical web interface
python3 scripts/dream.py --full_precision --web
and point your browser to http://127.0.0.1:9090
Notes
Some text you can add at the end of the prompt to make it very pretty:
cinematic photo, highly detailed, cinematic lighting, ultra-detailed, ultrarealistic, photorealism, Octane Rendering, cyberpunk lights, Hyper Detail, 8K, HD, Unreal Engine, V-Ray, full hd, cyberpunk, abstract, 3d octane render + 4k UHD + immense detail + dramatic lighting + well lit + black, purple, blue, pink, cerulean, teal, metallic colours, + fine details, ultra photoreal, photographic, concept art, cinematic composition, rule of thirds, mysterious, eerie, photorealism, breathtaking detailed concept art painting art deco pattern, by hsiao, ron cheng, john james audubon, bizarre compositions, exquisite detail, extremely moody lighting, painted by greg rutkowski makoto shinkai takashi takeuchi studio ghibli, akihiko yoshida
The original scripts should work as well.
python3 scripts/orig_scripts/txt2img.py --help
python3 scripts/orig_scripts/txt2img.py --ddim_steps 100 --n_iter 1 --n_samples 1 --plms --prompt "new born baby kitten. Hyper Detail, Octane Rendering, Unreal Engine, V-Ray"
python3 scripts/orig_scripts/txt2img.py --ddim_steps 5 --n_iter 1 --n_samples 1 --plms --prompt "ocean" # or --klms