Generative Models, Diffusion Models, Computer Vision, U-Net Implementation

💻 GitHub Repository

DDPM-AFHQ: End-to-End Diffusion Model Implementation

This project implements a full Denoising Diffusion Probabilistic Model (DDPM) using the AFHQ dataset (cats subset). The work follows the modern diffusion modeling pipeline, covering the forward process, reverse denoising process, cosine noise schedule, U-Net denoiser, training with L1 noise prediction loss, and visualization of both forward and backward diffusion. All training experiments—including FID logging, progressive sampling quality, and full 10k-step training—were completed on a A100 GPU.

Diffusion Forward & Reverse Processes

The forward process gradually adds Gaussian noise:

\(x_t = \sqrt{\bar{\alpha}_t} x_0 + \sqrt{1-\bar{\alpha}_t}\epsilon, \qquad \epsilon \sim \mathcal{N}(0,I)\)

The reverse process uses the U-Net to predict $\epsilon_\theta(x_t,t)$ and reconstruct $x_0$ through the posterior mean:

\(\tilde{\mu}_t = \frac{\sqrt{\alpha_t}(1-\bar{\alpha}_{t-1})}{1-\bar{\alpha}_t} x_t + \frac{\sqrt{\bar{\alpha}_{t-1}}(1-\alpha_t)}{1-\bar{\alpha}_t} \hat{x}_0\)

Forward Diffusion (PDF Figure)

Figure 1: Forward diffusion process (noise increases gradually)

Reverse Diffusion (PDF Figure)

Figure 2: Reverse diffusion process (model reconstructs from noise)

U-Net Noise Predictor

A lightweight U-Net was implemented with:

• Downsampling/upsampling blocks
• Skip connections
• GroupNorm + SiLU
• Time embedding injected into each residual block
• Middle-layer attention at low resolution

This network predicts the added noise $\epsilon_\theta(x_t,t)$, enabling reconstruction of $x_0$.

U-Net Diagram (PDF Figure)

Figure 3: U-Net architecture

Training Pipeline & Configuration

Objective:
\(\mathcal{L} = \| \epsilon - \epsilon_\theta(x_t,t) \|_1\)

Setup:

• Dataset: AFHQ (cats)
• Timesteps: T = 50
• Optimizer: AdamW (1e-3)
• Batch size: 32
• Cosine noise schedule (Nichol & Dhariwal 2021)
• Hardware: Google Colab A100

Training Loss Curve (PDF Figure)

Figure 4: Training loss curve

FID Tracking (PDF Figure)

Figure 5: FID progression over training steps

My Role & Contributions

As the sole developer, I:

• Implemented the full forward diffusion process
• Implemented reverse process sampling (p_sample, p_sample_loop, sample)
• Built the U-Net noise prediction model
• Engineered full training pipeline and L1 noise prediction loss
• Logged FID and produced visualizations (forward, backward, final samples)
• Completed full 10k-step training and evaluation
• Extracted and documented all results shown above

This project refined my understanding of variational inference, generative modeling, and the finer dynamics of diffusion processes.