Latent Diffusion
A diffusion model that learns to denoise in a compressed "latent" space rather than raw pixels - much faster and cheaper.
Think of it like
Instead of sculpting a full-size marble statue from scratch, you carve a small clay model, then enlarge it. Latent diffusion works in the small space.
Example
Stable Diffusion and DALL-E 2 use latent diffusion. You describe "a red bicycle in the rain" - the model denoises a compressed latent vector, then decodes it to a high-res image in seconds.
How it actually works
Standard diffusion adds noise to pixels; latent diffusion adds noise to latent codes (output of a VAE encoder). The autoencoder compresses images, drastically reducing compute (4-8x speedup). Denoising happens in latent space, then the decoder reconstructs pixels. Enables cheap, fast text-to-image. Trade-off: lossy compression can blur fine details, but empirically results are excellent. Usually paired with a text encoder (CLIP) and sometimes a classifier-free guidance mechanism.
For product teams
Democratizes image generation - runs on consumer hardware, enables real-time interactive editing, powers creative applications.
For engineers
Architecture: VAE (encoder/decoder) + diffusion UNet in latent space + optional text encoder. Training: ~500K A100-hours on image-caption pairs.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome