A diffusion model that fits in a single file

What it does minDiffusion is a bare-bones educational implementation of Denoising Diffusion Probabilistic Models built entirely in PyTorch. It trains a simple U-Net with time embeddings to reverse noise into images, supporting MNIST, CIFAR-10, and CelebA datasets. The entire core algorithm lives in superminddpm.py, a single file that clocks in at under 200 lines.

The interesting bit The project treats brevity as a feature, not a bug. By refusing to bury the sampling loop or noise schedule behind abstraction layers, it forces you to confront exactly how diffusion ticks—something that is often obscured in heavier frameworks.

Key highlights

• Core DDPM implementation in under 200 lines of self-contained Python
• Trains on standard benchmarks: MNIST, CIFAR-10, and CelebA
• Generates recognizable CIFAR-10 samples in roughly two hours on a single RTX 3090
• No exotic dependencies—just PyTorch and torchvision
• Refactored training script (train_cifar10.py) included for longer runs

Caveats

• Advanced techniques like DDIM and classifier guidance are listed as TODOs, not yet implemented
• The U-Net and time embeddings are intentionally simple; output quality is educational-grade, not production-grade
• README showcases CIFAR-10 and MNIST results, but does not display CelebA samples

Verdict This is for students or practitioners who want to see the diffusion loop without navigating a monolithic framework. Look elsewhere if you need pretrained weights, advanced samplers, or state-of-the-art fidelity.