bot66/MNISTDiffusion

Built with depthwise convolutions, residual shortcuts, and simplified timestep embeddings, the model achieves full DDPM functionality at minimal footprint through architectural simplification rather than quantization. Trains end-to-end on MNIST using PyTorch with configurable hyperparameters, producing both inference checkpoints (9.1MB with EMA) suitable for educational purposes or resource-constrained environments. References the canonical DDPM paper and implements the core denoising diffusion framework as a learning resource for understanding generative modeling fundamentals.

143 stars. No commits in the last 6 months.