Variation
Slightly different results pulled from the same request
- Variation means keeping the request the same and getting back several results that differ slightly.
- The results differ not because the request changed, but because the starting blur gets drawn fresh each time.
- Once one favorite turns up, that favorite can become the starting point for pulling several more, with a dial that sets how closely they resemble it.
- Rather than fighting to nail one perfect result, generating several and picking one is usually faster, and even the misses show where the request needs fixing.
- It serves a different purpose from interpolation, which bridges two results, and consistency, which keeps the same subject steady across many results.
Contents
1The analogy
Pour water into an ice tray and freeze it, and each compartment gives back an ice cube. They came out of the same tray, so the size and shape are all close to identical. Pick each one up, though, and where the air bubbles sit and how the surface cracked differ a little from cube to cube.
Typing one request and getting several pictures back works the same way. The tray is the request, and the small differences in each compartment are the fresh chance that slips in every time it freezes. Not a single word of the request has to change for each compartment to come out different.
If one of the ice cubes turns out to be a favorite, its shape can be used to cast a new tray and freeze another batch. This time, the results come out closer to the favorite that was picked. How close they come is set by turning a dial.
2In detail
Why the same request gives a different picture
The process of making a picture starts from a hazy, messy screen of blur. This starting blur gets drawn fresh every single time it's made. The number used to draw it is called the seed, and a different seed means a completely different starting blur.
The request just sets which direction to head in. A different starting point means arriving somewhere different even while walking in the same direction. That's why a request for a cat gets honored, while the fur pattern and the pose it's sitting in come out different every time.
Fix the seed by hand and the same picture can be called back up again. Variation runs the other way — it deliberately changes the seed every time to pull several sibling pictures.
Why pull several instead of one
Turning a picture in your head into a few lines of text with total accuracy is hard. There's a lot that's awkward to put into words, like the direction of the light or the composition. Getting several pictures back at once and picking whichever one comes closest means the eye can choose what words couldn't quite capture.
Even the ones that miss aren't necessarily worth throwing out. If all four pictures come back as a night scene, that means something in the request read as dark. Several pictures drifting the same way isn't chance — that's a sign something in the request needs fixing.
That's why image tools often hand back two or four pictures with a single click. The whole screen is built around the assumption that picking is part of the process.
Pulling more from a favorite
Sometimes a favorite picture turns up, but one small thing about it — the shape of a hand, say — is a letdown. In that case, that picture gets fed back in as the starting point, gets deliberately jostled by a small amount, and gets redrawn. How much it gets jostled is the same dial that sets how closely the results resemble the favorite.
Jostle it only a little and the layout and colors stay put while just the fine details shift. Jostle it a lot and only the mood survives while a different picture comes out. Push the dial all the way and the result has nothing to do with the original picture anymore, so nudging it up bit by bit is the better approach.
How this differs from editing the request
Editing the request changes the destination itself. Swap the background to a beach or specify a color, and the result lands somewhere entirely different. Variation keeps the destination the same and only changes the path taken to get there.
So there's an order to follow. When what's wanted points in a completely different direction, edit the request; when the direction is right and only the details are a letdown, pull several variations instead. Mixing the two and constantly fiddling with the request makes it impossible to tell what actually caused an improvement.
Keeping interpolation and consistency separate
Interpolation takes two results and fills in the space between them. What comes out is arranged in a line, with an order from one end to the other. Variation lines results up side by side, like siblings, with no order to them.
Consistency aims for close to the opposite. It's about making sure the same person or the same object comes out identical across several pictures. Push variation harder and consistency breaks down. That's why both get used together when making a picture story.
3More precisely
Variation isn't the name of one single feature — it's an umbrella term covering a handful of methods. Redrawing with only the seed changed, feeding a finished result back in as input and redrawing it with a controlled amount of jostling, and nudging the neighborhood around where a result sits in the model's internal space and drawing from there — all of these get called variation. Different tools use different names for the feature and different names for the dial, so it takes a look at the screen to figure out which method a given tool actually means.
The analogy breaks down in places. An ice tray has a fixed number of compartments, but the results that can be pulled are effectively endless. And where the difference between ice cubes is pure chance, the difference between pictures isn't random in an unconstrained way — it only wobbles within whatever range the request allows. Write the request in enough detail and the range of variation narrows on its own. If turning the dial hard still only produces similar pictures, the request has probably already pinned down most of the answer.
4Try it yourself
5Common misconceptions
It's easy to think a different result every time means the AI is unstable, but actually the starting blur is deliberately redrawn each time, and fixing the seed number brings the same picture back.
It's easy to think running variation again and again keeps improving the result, but actually it isn't improving — it's just spreading sideways, so there's a point where picking one and stopping makes sense.
It's easy to think variation can keep the same character consistent across several shots, but actually pushing it harder changes the face more, so consistency needs a separate method of its own.
7One-line summary
In shortVariation keeps the request the same and only changes the starting blur to pull back several sibling results, leaving the choice of a favorite for last.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02