Building Generative Image & Video models at Scale
summary
Shows that large-scale image and video generation trains diffusion models on latent autoencoders (EQ‑VAE / Stable Diffusion style) to cut tensor size by ~100x, uses JAX for sharded training, and applies guidance scale to trade diversity for prompt fidelity (doubling perceived quality while shrinking diversity). Demonstrates distillation into consistency models to reduce sampling steps, explains coarse‑to‑fine spectral view of diffusion, and compares deterministic versus stochastic samplers with concrete sampling‑pipeline tradeoffs.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 2:55Data Curation
- 4:02Representation
- 9:39Modeling: Diffusion Mechanism
- 20:01Network Architecture
- 22:25Training at Scale
- 23:33Sampling & Guidance
- 28:03Distillation
- 30:03Control Signals
related talks
30mHow to train a large diffusion model
22mEfficient Pretraining of Text-to-Image Models
01h 21mGenerative AI with Diffusion Models
10mDiffusion models
01h 17mDeep generative models for image...
17mThe State of Generative Media
30mRepetita Non Iuvant: Why Generative AI Models Cannot Feed Themselves
20mTraining a foundation image model and deploying it to millions of users