aijadugar/diffusers

DDPM — Fashion-MNIST

A simple DDPM implementation built from scratch using PyTorch.

The model learns to predict the noise added to a real Fashion-MNIST image.

Pipeline

Training:

x₀ → add noise → xₜ → U-Net → predicted noise
                          ↓
                     MSE Loss

Sampling:

Gaussian Noise
      ↓
    U-Net
      ↓
  Denoising
      ↓
    U-Net
      ↓
     ...
      ↓
 Generated Image

Model

The U-Net takes:

(xₜ, t)

and predicts:

εθ(xₜ, t)

The U-Net contains:

• Time embedding
• Input convolution
• 2 downsampling blocks
• 2 middle ResBlocks
• 2 upsampling blocks
• Skip connections
• GroupNorm
• SiLU activation
• Final output convolution

Input:

[B, 1, 28, 28]

Output:

[B, 1, 28, 28]

The output represents predicted Gaussian noise.

Forward Diffusion

We use:

xₜ = √ᾱₜ x₀ + √(1 - ᾱₜ) ε

where:

ε ~ N(0, I)

Training

For every batch:

1. Take a real image x₀.
2. Select a random timestep t.
3. Generate Gaussian noise ε.
4. Create xₜ using the forward diffusion process.
5. Give xₜ and t to the U-Net.
6. Predict the noise.
7. Calculate MSE between predicted and actual noise.
8. Update the U-Net.

Reverse Diffusion

During generation, we start from:

xT ~ N(0, I)

and move backwards:

xT → xT-1 → ... → x1 → x0

At every timestep, the U-Net predicts the noise and the DDPM scheduler performs one reverse step.

Checkpoint

Model weights are saved using:

.safetensors

Example:

checkpoints/ddpm_fashion_mnist.safetensors

The model architecture must be recreated before loading the weights.

Dataset

Dataset:

Fashion-MNIST

Image size:

28 × 28

Channels:

1

Goal

The goal of this project is to understand DDPM from first principles:

Forward Diffusion
      ↓
   U-Net
      ↓
Noise Prediction
      ↓
Reverse Diffusion
      ↓
Image Generation