Papers · Latent Diffusion / Stable Diffusion
AI PAPER

High-Resolution Image Synthesis with Latent Diffusion Models

Moved the diffusion process to a compressed latent space, making high-quality image generation 10-100x more efficient and enabling Stable Diffusion's open-source revolution.

Authors
Robin Rombach, Andreas Blattmann, Dominik Lorenz +2
Institution
Ludwig Maximilian University of Munich
Published
April NaN, 2022
Citations
12,000
Category
Multimodal
Impact
notable
PAPER EXPLAINED

Plain-English Summary

Standard diffusion models operate directly on full-resolution images, which is computationally expensive. Latent diffusion first compresses images into a much smaller latent space using an autoencoder, then performs the diffusion process in this compressed representation. This is like drawing a sketch first and then expanding it to a full painting — the creative decisions happen in the efficient sketch space.

This efficiency gain (10-100x) made high-quality image generation accessible on consumer GPUs. When Stability AI trained and released the model as Stable Diffusion with open weights, it triggered an explosion of community creativity and tool development.

Key Innovation

The two-stage approach (autoencoder compression followed by latent diffusion) separates perceptual compression from generative modeling. The autoencoder reduces a 512x512 image to a 64x64 latent representation with minimal perceptual loss. The diffusion model then operates in this compressed space, requiring far less computation per step.

Cross-attention conditioning allows the model to incorporate text embeddings (from CLIP) directly into the denoising process, enabling precise text-to-image generation without separate conditioning stages.

Impact on the Field

Stable Diffusion’s open release (August 2022) democratized image generation. Within months, millions of people were generating images locally. A massive ecosystem of tools, interfaces (Automatic1111, ComfyUI), and model extensions (LoRA adapters, ControlNet, IP-Adapter) emerged. The open-source nature enabled rapid innovation that proprietary systems could not match.

The paper also established latent diffusion as the dominant approach for efficient generative modeling, influencing all subsequent diffusion-based systems.

Models That Built on This

Stable Diffusion 1.5, 2.0, SDXL, and SD3 all build on latent diffusion. Stable Video Diffusion extends to video generation. Flux (from Black Forest Labs, by the same researchers) continues the approach. DALL-E 3, Midjourney, and Imagen all adopt latent diffusion principles. The community has created hundreds of thousands of custom models and LoRA adapters. The latent diffusion approach now underpins the entire commercial image generation industry.