Generative AI with Diffusion Models
summary
Walks through denoising diffusion probabilistic models with the concrete claim that diffusion now outperforms GANs on image synthesis, explains the forward (thousands of iterative noising steps) and reverse denoising math, and shows implementations from nn.labml.ai and the Hugging Face diffusers library plus references to Stable Diffusion. Notes practical costs: models can reach ~7 GB and generation is multi-iteration and slow, and a diffusion-based segmentation attempt works in quality but proves too large/latency-heavy for production.
AI-written summary · about
chapters (8)
- 0:00Welcome to diffusion models
- 10:14Generative AI history for images
- 23:43Diffusion model building blocks
- 35:58Diffusion basics
- 44:53Forward diffusion process
- 56:09Reverse diffusion and denoising
- 1:04:39GANs vs diffusion models
- 1:13:20Implementing diffusion models
related talks
44mGenerative Adversarial Networks (GANs) and Stable Diffusion
10mDiffusion models
41mBuilding Generative Image & Video models at Scale
53mGenerative AI: Overview and use cases
01h 04mGeneralization in Diffusion Models Arises from Geometry-Adaptive Harmonic Rrepresentations
01h 09mLearning Generative Models from Corrupted Data
49mGenerative Diffusion Models: Development, Training and Releasing Kandinsky 2.1
30mHow to train a large diffusion model