10 KiB
Table of Contents
Step 1 - Get the Model
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.
Step 2 - Installation
Option A - 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. Increase the CPUs and Memory to avoid this Issue. You may need to increase Swap and Disk image size too.
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 --platform linux/arm64 --name dummy --mount source=my-vol,target=/data alpine # or arm64v8/alpine
# Copy the model file to the Docker volume. We'll need it at run time.
cd ~/Downloads # or wherever you saved sd-v1-4.ckpt
docker cp sd-v1-4.ckpt dummy:/data
Setup
# Set the fork you want to use.
GITHUB_STABLE_DIFFUSION="https://github.com/santisbon/stable-diffusion.git"
cd ~
git clone $GITHUB_STABLE_DIFFUSION
cd stable-diffusion/docker-build
chmod +x entrypoint.sh
# download the Miniconda installer. We'll need it at build time.
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-aarch64.sh -O anaconda.sh && chmod +x anaconda.sh
Build the Docker image. Give it any tag -t that you want.
Tip: Make sure your shell session has the env variable set (above) with echo $GITHUB_STABLE_DIFFUSION.
docker build -t santisbon/stable-diffusion \
--build-arg gsd=$GITHUB_STABLE_DIFFUSION \
--build-arg sdreq="requirements-linux-arm64.txt" \
.
Run a container using your built image e.g.
docker run -it \
--rm \
--platform linux/arm64 \
--name stable-diffusion \
--hostname stable-diffusion \
--mount source=my-vol,target=/data \
--expose 9090 \
--publish 9090:9090 \
santisbon/stable-diffusion
Tip: Make sure you've created the Docker volume (above)
Option B - 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
Setup
# Set the fork you want to use.
GITHUB_STABLE_DIFFUSION="https://github.com/santisbon/stable-diffusion.git"
git clone $GITHUB_STABLE_DIFFUSION
cd stable-diffusion
mkdir -p models/ldm/stable-diffusion-v1/
PATH_TO_CKPT="$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
Face Restoration and Upscaling
# by default expected in a sibling directory to stable-diffusion
cd .. && git clone https://github.com/TencentARC/GFPGAN.git && cd GFPGAN
# basicsr: used for training and inference. facexlib: face detection / face restoration helper.
pip3 install basicsr facexlib \
&& 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 ../stable-diffusion
Only need to do this once. If we don't preload models it will download model files from the Internet when you run dream.py. Used by the core functionality and by GFPGAN/Real-ESRGAN.
python3 scripts/preload_models.py
Step 3 - Usage (time to have fun)
Startup
If you're on a Linux container the dream script is automatically started and the output dir set to the Docker volume you created earlier.
If you're directly on macOS follow these startup instructions.
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.
# If on Macbook
python3 scripts/dream.py --full_precision
# By default the images are saved in outputs/img-samples/.
You'll get the script's prompt. You can see available options or quit.
dream> -h
dream> q
Text to Image
For quick (but bad) image results test with 5 steps (default 50) and 1 sample image. This will let you know that everything is set up correctly.
Then 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 0.75
# TODO: example for upscaling.
You'll need to experiment to see if face restoration is making it better or worse for your specific prompt. The -U option for upscaling has an Issue.
If you're on a container the output is set to the Docker volume. You can copy it wherever you want.
You can download it from the Docker Desktop app, Volumes, my-vol, data.
Or you can copy it from your Mac terminal. Keep in mind docker cp can't expand *.png so you'll need to specify the image file name:
# On your host Macbook (you can use the name of any container that mounted the volume)
docker cp dummy:/data/000001.928403745.png /Users/<your-user>/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 /Users/<your-user>/Pictures/sketch-mountains-input.jpg dummy:/data/
Try it out generating an image (or 4).
The dream script needs absolute paths to find the image so don't use ~.
# If you're on your Macbook
dream> "A fantasy landscape, trending on artstation" -I /Users/<your-user>/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 0.75 --steps 100 -n1
Web Interface
You can use the dream script with a graphical web interface. Start the web server with:
python3 scripts/dream.py --full_precision --web
If it's running on your Mac point your Mac web browser to http://127.0.0.1:9090
Press Control-C at the command line to stop the web server.
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, 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