Unofficial Fastapi implementation Stable-Diffusion API

UNOFFICIAL, Stable-Diffusion api using FastAPI

Samples

Text2Image-01	Text2Image-02
Image2Image-01	Image2Image-02
Inpaint-01	Inpaint-02

Features

• [x] long-prompt-weighting support
• [x] text2image
• [x] image2image
• [x] inpaints
• [x] negative-prompt
• [x] celery async task (check celery_task branch)
• [x] original ckpt format support
• [ ] object storage support
• [ ] stable-diffusion 2.0 support
• [ ] token size checker
• [ ] JAX/Flax pipeline

Requirements

API

fastapi[all]==0.80.0
fastapi-restful==0.4.3
fastapi-health==0.4.0
service-streamer==0.1.2
pydantic==1.9.2
diffusers==0.3.0
transformers==4.19.2
scipy
ftfy

Frontend

streamlit==1.12.2
requests==2.27.1 
requests-toolbelt==0.9.1 
pydantic==1.8.2
streamlit-drawable-canvas==0.9.2

API

/text2image

create image from input prompt

inputs:

- prompt(str): text prompt
- num_images(int): number of images
- guidance_scale(float): guidance scale for stable-diffusion
- height(int): image height
- width(int): image width
- seed(int): generator seed

outputs:

- prompt(str): input text prompt
- task_id(str): uuid4 hex string
- image_urls(str): generated images url

/image2image

create image from input image

inputs:

- prompt(str): text prompt
- init_image(imagefile): init image for i2i task
- num_images(int): number of images
- guidance_scale(float): guidance scale for stable-diffusion
- seed(int): generator seed

outputs:

- prompt(str): input text prompt
- task_id(str): uuid4 hex string
- image_urls(str): generated images url

Environment variable

# env setting is in 
>> ./core/settings/settings.py

RUN from code (API)

1. install python Requirements

pip install -r requirements.txt

2. downlaod and caching huggingface model

python huggingface_model_download.py
# check stable-diffusion model in huggingface cache dir 
[[ -d ~/.cache/huggingface/diffusers/models--CompVis--stable-diffusion-v1-4 ]] && echo "exist"
>> exist

3. update settings.py in ./core/settings/settings.py

# example
class ModelSetting(BaseSettings):
    MODEL_ID: str = "CompVis/stable-diffusion-v1-4" # huggingface repo id
    ENABLE_ATTENTION_SLICING: bool = True
...
class Settings(
    ...
):
    HUGGINGFACE_TOKEN: str = "YOUR HUGGINGFACE ACCESS TOKEN"
    IMAGESERVER_URL: str  = "http://localhost:3000/images"
    SAVE_DIR: str = 'static'
    ...

4. RUN API from code

bash docker/api/start.sh

RUN from code (frontend)

1. install python Requirements

pip install \
    streamlit==1.12.2 \
    requests==2.27.1 \
    requests-toolbelt==0.9.1 \
    pydantic==1.8.2 \
    streamlit-drawable-canvas==0.9.2

2. RUN streamlit frontend

streamlit run inpaint.py

RUN using Docker (docker-compose)

1. Image Build

docker-compose build

3. update docker-compose.yaml file in repo root

version: "3.7"
services:
  api:
    ...
    volumes:
      # mount huggingface model cache dir path to container root user home dir
      - /model:/model  # if you load pretraind model 
      - ...
    environment:
      ...
      MODEL_ID: "CompVis/stable-diffusion-v1-4" 
      HUGGINGFACE_TOKEN: {YOUR HUGGINGFACE ACCESS TOKEN}
      ...

    deploy:
      ...
  frontend:
    ...

4. Container RUN

docker-compose up -d 
# or API only
docker-compsoe up -d  api
# or frontend only
docker-compsoe up -d  frontend

References

• CompVis/stable-diffusion
• huggingface, stable-diffusion
• teamhide/fastapi-boilerplate