LECTURES A GRATIS GLOBAL SERVICE
⌕ SEARCH GRATIS GLOBAL ↗
LECTURES
Transformers in Diffusion Models for Image Generation and Beyond
SOURCE: YOUTUBE · NO TRACKING UNTIL YOU PRESS PLAY · TROUBLE PLAYING? WATCH AT THE SOURCE ↗

Transformers in Diffusion Models for Image Generation and Beyond

75 MIN · EN · STATUS: [ STREAMING ]
RATE THIS
STANFORD

Sayak Paul of Hugging Face, part of Stanford's CS25 Transformers United course, lays out how transformer architectures have taken over diffusion models for image generation. He starts with the basics of how diffusion models are trained and the UNet-based architectures that once dominated the field, then explains why transformer-based designs replaced them. The talk works through the core architectural building blocks, the design choices available under different conditioning setups, and the varieties of attention mechanisms used in current state of the art open models. Paul, who works on Hugging Face's diffusers library and trains diffusion models daily, draws on that hands-on experience throughout, closing with open problems around making these models more efficient. The talk assumes familiarity with transformers and is aimed at viewers who already know the basics of generative modeling.

At a glance

Lecture facts

Runtime compared with the other 734 Mathematics lectures
Runtime1 h 15 m
Compared with MathematicsLonger than 63%
Source channelStanford Online (YouTube)