DiffusionLab — DDPM From Scratch in PyTorch

A from-scratch implementation of a Denoising Diffusion Probabilistic Model (DDPM) in PyTorch.

This project explores the core ideas behind diffusion models by progressively building:

• MNIST data pipeline
• Forward image corruption
• Gaussian noise scheduling
• U-Net denoising architecture
• Sinusoidal timestep embeddings
• Noise prediction training objective
• Reverse diffusion sampling
• Automated tests
• Visualization of generated samples

Results

Forward Diffusion

Denoising Results

DDPM Generated Samples

Architecture

The model learns to reverse a gradual noising process.

Clean Image
↓
Add Gaussian Noise at timestep t
↓
Noisy Image x_t
↓
Time-Conditioned U-Net
↓
Predict Noise ε
↓
Reverse Diffusion
↓
Generated Image

Project Structure

DiffusionLab-DDPM-From-Scratch-in-PyTorch/
│
├── src/
│   ├── data.py
│   ├── noise.py
│   ├── visualize_noise.py
│   ├── unet.py
│   ├── train.py
│   ├── evaluate.py
│   ├── sample.py
│   ├── ddpm_scheduler.py
│   ├── train_ddpm.py
│   └── sample_ddpm.py
│
├── tests/
│   ├── test_noise.py
│   ├── test_unet.py
│   └── test_ddpm_scheduler.py
│
├── outputs/
├── notebooks/
├── requirements.txt
├── pytest.ini
└── README.md