Multimodal · Generation

This module covers diffusion models and flow matching for image and video generation, plus the additional temporal-consistency and compute-cost problems introduced by video.

Chapters

1. Chapter 7: Diffusion Models, Flow Matching, and Image Generation
2. Chapter 8: Video Generation Models

Connections within the module

flowchart LR
    D[Diffusion / Flow Matching<br/>and Image Generation] --> V[Video Generation Models]

Video can reuse image-generation objectives and backbones, but also needs temporal representations, video compression, motion supervision, and history conditioning. Not every model uses a Transformer, and spatiotemporal patches do not imply a flow-matching objective.

Chapter 7 distinguishes training paths, prediction parameterizations, samplers, and conditional guidance. Chapter 8 adds subject identity, motion, and long-range temporal consistency. Straight conditional paths do not guarantee one-step generation, and smooth images do not guarantee correct motion. These counterexamples help clarify the methods' limits.

Return to Multimodal Topics.