Denoising Diffusion Models from Scratch

Part I · Chapter 00 · 5 sections · 85 min

Essential mathematical background for understanding diffusion models

1. 0.1Probability Fundamentals15m
2. 0.2Gaussian Distributions Deep Dive20m
3. 0.3Information Theory Essentials15m
4. 0.4Variational Inference Primer20m
5. 0.5Markov Chains and Stochastic Processes15m

Part I · Chapter 01 · 4 sections · 49 min

Context and motivation for diffusion models

1. 1.1The Generative Modeling Problem12m
2. 1.2Landscape of Generative Models15m
3. 1.3The Diffusion Model Intuition12m
4. 1.4Historical Context and Key Papers10m

Part II · Chapter 02 · 5 sections · 80 min

Understanding how to systematically destroy data with noise

1. 2.1Defining the Forward Process15m
2. 2.2The Noise Schedule18m
3. 2.3Closed-Form Sampling at Any Timestep20m
4. 2.4Properties of the Forward Process12m
5. 2.5Implementing the Forward Process15m

Part II · Chapter 03 · 5 sections · 100 min

Learning to denoise - the heart of diffusion models

1. 3.1The Reverse Process Goal12m
2. 3.2Deriving the True Reverse Distribution25m
3. 3.3Parameterizing the Reverse Process18m
4. 3.4The Training Objective Derivation25m
5. 3.5Score Matching Perspective20m

Part II · Chapter 04 · 4 sections · 60 min

Deep dive into training objectives and their properties

1. 4.1The Simplified Loss Explained15m
2. 4.2Loss Weighting Strategies18m
3. 4.3Connection to Denoising Autoencoders12m
4. 4.4Numerical Analysis of the Loss15m

Part III · Chapter 05 · 5 sections · 88 min

The neural network backbone of diffusion models

1. 5.1Why U-Net?12m
2. 5.2U-Net Building Blocks18m
3. 5.3Time Conditioning15m
4. 5.4Attention in Diffusion U-Net18m
5. 5.5Complete U-Net Implementation25m

Part III · Chapter 06 · 5 sections · 80 min

Assembling all components into a working DDPM

1. 6.1The Complete DDPM Class20m
2. 6.2Training Loop Implementation18m
3. 6.3Sampling Algorithm15m
4. 6.4Practical Training Tips15m
5. 6.5Debugging and Visualization12m

Part III · Chapter 07 · 5 sections · 70 min

Faster and better generation algorithms

1. 7.1Problems with Ancestral Sampling10m
2. 7.2DDIM: Deterministic Sampling20m
3. 7.3DDIM Implementation15m
4. 7.4Advanced Samplers Overview15m
5. 7.5Choosing a Sampler10m

Part IV · Chapter 08 · 5 sections · 80 min

Controlling what the model generates

1. 8.1Unconditional vs Conditional Models12m
2. 8.2Class-Conditional Diffusion15m
3. 8.3Classifier Guidance18m
4. 8.4Classifier-Free Guidance20m
5. 8.5Implementing Classifier-Free Guidance15m

Part IV · Chapter 09 · 4 sections · 63 min

Understanding the path to modern text-to-image models

1. 9.1Text Conditioning Overview15m
2. 9.2Cross-Attention Mechanism18m
3. 9.3CLIP and Contrastive Learning15m
4. 9.4From Pixel Space to Latent Space15m

Part V · Chapter 10 · 3 sections · 37 min

Setting up data for training a diffusion model

1. 10.1Dataset Selection10m
2. 10.2Data Preprocessing15m
3. 10.3Building the Training Pipeline12m

Part V · Chapter 11 · 4 sections · 62 min

Complete training setup and execution

1. 11.1Model Configuration12m
2. 11.2Complete Training Script20m
3. 11.3Training Monitoring15m
4. 11.4Common Issues and Solutions15m

Part V · Chapter 12 · 4 sections · 57 min

Using the trained model and measuring quality

1. 12.1Generating Images12m
2. 12.2Evaluation Metrics18m
3. 12.3Qualitative Analysis12m
4. 12.4Interactive Demo15m

Part VI · Chapter 13 · 4 sections · 68 min

Understanding the architecture behind Stable Diffusion

1. 13.1The Latent Diffusion Idea15m
2. 13.2The VAE Component18m
3. 13.3Diffusion in Latent Space15m
4. 13.4Stable Diffusion Architecture Overview20m

Part VI · Chapter 14 · 4 sections · 54 min

ControlNet, IP-Adapter, and beyond

1. 14.1ControlNet Concept15m
2. 14.2Image-to-Image Generation15m
3. 14.3IP-Adapter and Image Prompts12m
4. 14.4Multi-Modal Conditioning12m

Part VI · Chapter 15 · 3 sections · 53 min

The unified mathematical framework

1. 15.1Score-Based Generative Models18m
2. 15.2Diffusion as SDEs20m
3. 15.3Unified Framework15m

Part VII · Chapter 16 · 3 sections · 39 min

Making models practical

1. 16.1Model Acceleration15m
2. 16.2Quantization and Efficiency12m
3. 16.3Serving Diffusion Models12m

Part VII · Chapter 17 · 4 sections · 44 min

Current research directions

1. 17.1Video Generation12m
2. 17.23D Generation12m
3. 17.3Beyond Images10m
4. 17.4Open Problems and Research Directions10m