Stable Diffusion Implementation

From-scratch implementation of a Stable Diffusion model, divided into two main parts: the model_package for the model implementation and packaging, and the serving_api for deployment of the model via a FastAPI interface.

Table of Contents

1. Project Structure
2. Setup
3. Usage
4. Configuration
5. License

Project Structure

stable-diffusion/
│
├── model_package/
│   ├── sd_model/
│   │   ├── config/
│   │   ├── data/
│   │   ├── model/
│   │   │   ├── clip.py
│   │   │   ├── sampler.py
│   │   │   ├── unet_diffusion.py
│   │   │   └── vae.py
│   │   ├── util/
│   │   ├── finetuning.py
│   │   └── predict.py
│   ├── tests/
│   ├── requirements/
│   ├── setup.py
│   └── tox.ini
│
└── serving_api/
    ├── app/
    │   ├── api.py
    │   ├── config.py
    │   ├── main.py
    │   ├── schemas/
    │   └── tests/
    ├── Procfile
    ├── requirements.txt
    └── tox.ini

Setup

Usage

Model Package

Finetuning

Prediction

Serving API

Model can be deployed via CI/CD pipeline implemented in Jenkinsfile (currently its Railway deployment). If one wants to test API locally:

1. Navigate to the serving_api directory:

cd ../serving_api

2. Start the FastAPI server:

uvicorn app.main:app --host 0.0.0.0 --port 8001

3. The API will be accessible at http://localhost:8001.

API Endpoints

• POST /api/v1/generate: Generate an image from a text prompt.

Example payload:

{
    "inputs": [
        {
            "config": {
                "mode": "text_to_image",
                "prompt": "A beautiful painting of a sunset over a mountain range",
                "uncond_prompt": "",
                "strength": 0.75,
                "do_cfg": True,
                "cfg_scale": 7.5,
                "num_inference_steps": 50,
                "seed": 42,
                "device": "cuda",
                "idle_device": "cpu"
            }
        }
    ]
}

Configuration

Configuration files for the model are located in the model_package/sd_model/config directory. Adjust the base_config.yaml file as needed for your setup.

License

This project is licensed under the MIT License - see the LICENSE file for details.