What is Image-to-Image? — one analogy

1The analogy

Bring an old photo to a photo studio and ask to have it reshot with the same composition. The studio doesn't hand you a copy of the photo you brought. It sets the photo aside as a reference and shoots an entirely new one, working out the background, lighting, and pose fresh. What comes back resembles the photo you brought, but it's clearly a different shot.

How closely it resembles the original is up to the request. Ask for it to stay nearly the same, and the layout and colors both follow the original closely; ask for a big change, and only a faint trace of the layout survives while everything else comes out new.

Image-to-image is this same request. Hand over a starting picture and get a new picture built from it in return.

2In detail

Starting from a picture instead of blur

Given only a sentence, a generator starts from a screen full of blur and clears it away step by step. Image-to-image only swaps the starting point. Instead of pure blur, it starts from the handed-over picture with a certain amount of blur mixed in.

Since the starting point is the handed-over picture half-buried in blur, the big shapes and the light and dark areas of that picture survive as the blur gets cleared. At the same time, the details get rebuilt fresh according to the sentence and the model's own habits. It's not painting over the original — it's redrawing a picture that resembles the original, from scratch.

The dial for how much to change

Tools call it different things, but there's always a value that sets how much changes. This value decides how much blur gets mixed in at the start. Mix in a little and the original stays sharp; mix in a lot and only a faint trace of the original survives.

Set the value very low and the result comes out nearly identical to the original, with no sign it was touched. Push it all the way up and handing over a starting picture stops meaning anything. Somewhere in the middle is the sweet spot for borrowing just the original's layout and getting everything else fresh. Starting low and nudging it up bit by bit is the fastest way to find where that comfortable range sits.

The starting picture and the sentence split the job

What the starting picture locks down is placement. What sits where on the screen, which areas are bright or dark, which way the eye is drawn — all of that comes from the starting picture. What the sentence locks down is content. What that shape actually is, what color and material it's made of — the sentence decides that.

So when the two disagree, the result comes out strained. Ask for a portrait-oriented scene starting from a landscape-oriented picture and the result looks forced. Leave a large shape sitting in the starting picture unmentioned in the sentence and it gets filled in with something unrelated. Naming what's already in the starting picture, even briefly, in the sentence makes the result noticeably steadier.

Where it's handy

The most welcome use is rescuing a rough sketch. Block in a few patches of color to rough out a layout, describe the content in a sentence, and a polished picture that keeps that layout comes back. It's faster and more precise than trying to describe a mental layout in a long sentence.

It also gets used a lot to get a different texture out of an existing photo. Feed in a photo of an alley taken by hand and ask for a painterly texture, and a picture comes back that keeps the alley's layout while only the material and brushwork change. There's no need to describe what's where on the screen, so the sentence gets much shorter too.

It's good for pulling several pictures out of one favorite too, with only the texture changing. Feed the same starting picture in with a slightly different sentence each time, and a set comes back with the layout locked in place while only the mood shifts. Shifting the season or the time of day while keeping a photo's layout untouched works the same way.

3More precisely

Image-to-image converts the original into the compressed form the model works with, mixes in a set amount of blur, and starts the clearing partway through instead of from the very beginning. A change value that keeps most of the steps means the walk covers only the later portion of the full process, so a lower value also means fewer steps and a faster result. Push the value to its maximum and it's the same as not using the starting picture at all.

The analogy breaks down in one place. A photo studio looks at the photo brought in as a reference, but image-to-image doesn't keep the original off to the side as a reference — it uses it directly as the starting screen. So any blur or shake already present in the original gets mixed into the start and carries through into the result. And where a photo studio leaves everything untouched except what was asked for, this method rebuilds the entire screen, so spots meant to stay untouched can shift too. Painting in only the area meant to be fixed is the right move when just a part needs to stay put.

4Try it yourself

Sketch to Photo ailearn.space Draw a few simple lines and they become the starting point for a photo-like picture

5Common misconceptions

It's easy to think it's a feature for editing the original photo, but actually it uses the original as a starting point and builds a new picture from scratch.
It's easy to think it only changes the style, but actually the content on screen changes too, depending on the value.
It's easy to think the sentence can be written carelessly once a starting picture is given, but actually the starting picture only sets placement — the content is set by the sentence.

7One-line summary

In shortImage-to-image starts from a starting picture with blur mixed in and builds a new picture from there, with a change dial that decides how much of the original survives.

Did this analogy help?

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