What is Latent Space? — one analogy

1The analogy

A drawer divided into compartments doesn't hold things just anywhere. Long things go in the left row, short things in the right row, round things in the back row, angular things up front — every compartment has a character. Open the drawer, and you can guess roughly what's where without inspecting every item one by one.

In this drawer, position itself is information. Items in neighboring compartments share similar properties; an item at the opposite end of the drawer is completely different.

Set the dividers closer together, and the compartments split into finer slices. Add a new compartment between two existing ones, and whatever goes there sits about halfway between the two neighboring items. Latent Space is a drawer just like this.

2In detail

Position is the property

A model doesn't work with a photo or a sentence directly. It first shrinks it down to a short bundle of numbers. That single bundle of numbers corresponds to one spot in the drawer, and Latent Space is the collection of every spot a compressed value could possibly sit in.

A single spot carries several numbers attached to it. Unlike a drawer with just two directions, side to side and front to back, this space has anywhere from dozens to hundreds of directions. It's hard for a person to picture in their head, but measuring the distance between two spots and moving along a direction still works exactly the same way.

What matters is that these spots aren't scattered randomly. While the compressing practice repeats over and over, the spots get organized so that similar data ends up close together. They organize this way because it's an advantage when it comes to rebuilding.

Close means similar, far means different

Photos of the same kind end up in neighboring spots. Not because their pixels look similar, but because the properties the model decided matter are similar. Even with a totally different background color, two photos of the same subject can end up sitting close together.

This property gets used for search. Compress whatever you're looking for the same way to find its spot, then pull out whatever sits nearby. Things with similar properties can turn up even when no name or word overlaps at all.

Direction takes on meaning too

Once the spots settle into order, directions pick up rules too. Nudge a little in one direction and the result gets brighter; move in a different direction and it looks older. This is exactly why changing just one single property becomes possible.

But these directions weren't assigned by a person. They're found afterward, by a person moving around and experimenting once training is done, so they aren't cleanly separated from each other. Trying to change just the brightness and having the background shift along with it is common.

Landing between two spots gives an in-between result

Fix the spots of two pieces of data, walk between them at even steps, and generate a result at each one, and you get a sequence that gradually shifts from one to the other. This is different from overlapping two pictures and blurring them together — the result at a middle spot is a complete, valid picture in its own right.

For this walk to work well, the space between spots needs to be packed tightly. If every piece of data gets only a single point with empty gaps in between, a spot in the middle produces something unrecognizable. This is exactly why an approach that remembers data as a range instead of a point exists.

Outside the region is empty

Latent Space being vast doesn't mean every spot in it is usable. There's a region where the training data actually settled, and everything outside it is a spot that was never practiced on, not once. Pull a value from out there, and the result comes out mangled or nonsensical.

In drawer terms, it's like setting something down on the bare floor outside the compartments. A spot exists there, but nobody ever decided what it's supposed to mean. This is exactly why generating something new means pulling a value from near the region where the training data clustered, rather than plugging in just anything.

3More precisely

Latent Space is a multidimensional space where the vectors a model builds by compressing its input come to sit. Each axis isn't a category a person defined — it's a value decided during training, and no single axis maps cleanly onto one human-recognizable property. Looking at the relationship between two spots means measuring the straight-line distance between them or the angle of the direction connecting them.

The analogy breaks down in places. A drawer's compartments can be touched, and a person builds the dividers, but a spot in Latent Space is a continuous value with no marked-off gradations, and no person knows what standard divided it up. A drawer has two directions, but this space actually runs to hundreds of directions, and some corners of it resist the sense of distance we're used to.

One more thing: Latent Space gets built separately for every model. Feed the same photo into a different model, and it lands in a completely different spot, even if the two models were trained on very similar data. That's exactly why a value obtained from one model can't just be carried over and used in another without some kind of translation between the two spaces.

4Try it yourself

5Common misconceptions

It's easy to think each axis of Latent Space is responsible for one item like color or size, but actually several properties are tangled together on a single axis, so changing just one isn't simple.
It's easy to think a nearby spot means similar appearance, but actually closeness reflects properties the model considered important, so the actual look can differ quite a bit.
It's easy to think you can pull a plausible result from anywhere in the space, but actually stray outside the region the training data occupied, and the result comes out unrecognizable.

7One-line summary

In shortLatent Space is the place where compressed values end up sitting, and because position and direction carry meaning there, it's what lets you find similar things or generate something in between.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02