DDPM Image Generation from Scratch | PyTorch + Streamlit Demo
This video demonstrates a complete Denoising Diffusion Probabilistic Model (DDPM) implemented from scratch using PyTorch for image generation.
The project includes a custom U-Net architecture with time-step embeddings, support for multiple noise schedules, and key training-stability techniques such as Exponential Moving Average (EMA) and gradient clipping. The full training pipeline was built end-to-end, including data loading, noise scheduling, sampling, and basic evaluation metrics.
A professional Streamlit web application was developed to showcase the model, featuring auto-configuration detection, real-time image generation, and a clean, user-friendly interface. The app allows interactive image generation for MNIST and CIFAR-10 datasets and is deployed on Streamlit Cloud with performance optimizations for smooth inference.
Key Highlights
• DDPM implemented from scratch in PyTorch
• U-Net with time embeddings
• Multiple noise schedules
• Training stability: EMA & gradient clipping
• End-to-end training and sampling pipeline
• Streamlit UI with real-time generation
• MNIST and CIFAR-10 image generation
• Cloud deployment on Streamlit
Live Demo
https://diffusion-model-generator.streamli...
Source Code (GitHub)
https://github.com/Samuel-Hailemariam-Seif...