AI Hub
All papers
Architecture

High-Resolution Image Synthesis with Latent Diffusion Models

LMU Munich·December 20, 2021

Robin Rombach, Andreas Blattmann, Dominik Lorenz

View on arXiv

TL;DR

Runs diffusion in a compressed latent space instead of pixel space, slashing the compute needed for high-resolution image generation.

Why it matters

The method behind Stable Diffusion. Moving diffusion into latent space is what made open, consumer-GPU image generation possible.

Related models

Related terms