LDM
A paper published in 2022 by Robin Rombach et al. proposing that the diffusion process be performed in latent space rather than pixel space, significantly reducing computational cost. It forms the theoretical foundation of the Stable Diffusion model.
Timeline (last 90 days)
VAE typically has a compression ratio of 8×8 (spatial dimensions reduced by 64×), originating from the Latent Diffusion Model (LDM) framework of Stable Diffusion
In 2021, Robin Rombach and others from Ludwig Maximilian University of Munich proposed the Latent Diffusion Model (LDM), which uses a pre-trained VAE to compress images into a low-dimensional latent space (typically 1/8 of the original resolution) before performing diffusion denoising.
All Facts (2)
VAE typically has a compression ratio of 8×8 (spatial dimensions reduced by 64×), originating from the Latent Diffusion Model (LDM) framework of Stable Diffusion
50%UnverifiedIn 2021, Robin Rombach and others from Ludwig Maximilian University of Munich proposed the Latent Diffusion Model (LDM), which uses a pre-trained VAE to compress images into a low-dimensional latent space (typically 1/8 of the original resolution) before performing diffusion denoising.
50%