A latent space is the compressed internal representation a model works in.
Similar things end up near each other, even though nobody arranged them that way.
Directions inside that space often carry meaning you can actually use.
Generating in a small latent space is far cheaper than generating full-size pixels.
That is exactly what latent diffusion does, and why image generation became practical.
A latent space is the compressed internal representation a model works in.
Latent means hidden, and these are the features nobody designed and the model found.
An image of a million pixels might be represented by a few thousand numbers.
Similar things land near each other, without anyone arranging that.
Structure appears because the model needed it to do its job.
Directions in the space often carry meaning you can use.
Move along one direction and a face gains a smile, and along another it ages.
Those directions are found afterwards by looking, not built in.
Interpolating between two points produces plausible things in between.
That is the test of a well-formed space, since gaps would decode to noise.
Working in a small latent space is far cheaper than working in full-size pixels.
That is exactly what latent diffusion does, and it is why image generation became practical.
An image is compressed by a VAE, denoised in latent space, then decoded once at the end.
The same idea appears in language, where an embedding is a latent representation of text.
Latent spaces are not interpretable dimension by dimension, so any meaning found is emergent.
One question on this card. Every option explains itself once you pick, right or wrong.
18 concepts in this module, then a graded test with every answer explained.