Why do I get a different picture from the same description?

Because the process starts from random noise and that randomness is different each time. It is not inconsistency in the model. Where a product lets you set the starting point, repeating an exact result becomes possible, which is the setting worth finding if you are trying to produce a series.

Why can it not change just one part of a picture?

Because the description influences every step of the refinement rather than being applied to a region, so a small change to the wording moves the whole image. Some products offer ways of masking an area and regenerating only that, which is a different feature layered on top and worth asking about specifically.

Is this the same kind of model as the ones that write?

No, and they are built and trained quite differently. Text models predict what comes next, one piece at a time, in a fixed direction. These refine a whole picture at once, repeatedly. Products increasingly present both behind one interface, which makes the underlying difference easy to miss.

Glossary

Diffusion model

A diffusion model starts from visual noise and refines it step by step towards a picture matching your description, which is how most image generators work and why they behave as they do.

In plain terms

Rather than drawing, it un-blurs. The system begins with a screen of random speckle and repeatedly removes a little of the randomness, each time nudging what remains towards the description you gave. After enough passes a picture has emerged. That sounds like an odd way to make an image and it explains most of what these tools do that seems arbitrary.

Why it matters

You can use image generators perfectly well without knowing any of this, and it is the one page here you could reasonably skip. What it buys, if you read it, is that several irritating behaviours stop being mysterious. Why the same description twice gives you different pictures. Why asking for more steps helps up to a point and then does not. And why fine detail in a corner of a large image is the part most likely to be wrong.

How it works

Training runs the process backwards. The model is shown pictures with increasing amounts of randomness added and learns to undo each step, so by the end it knows how to move from noise towards something that looks like a real image. Generating is that learnt undoing, run from pure noise.

Your description steers each step rather than being read once at the start. That is why wording changes propagate through the whole picture instead of adjusting one part of it, and why these tools are poor at leave everything and change only this, a request that would be trivial for an editing program.

The starting noise is random, which is the direct cause of the same description producing different pictures. Where a product lets you fix that starting point, repeating a result becomes possible, and that setting is the single most useful one for anybody trying to produce a consistent set.

More refinement steps improve the result up to a point and then stop mattering, while costing time throughout. Every product picks a default somewhere sensible, and the returns fall off sharply enough that turning it up is rarely the answer to a picture that came back wrong.

One picture, from noise to result

Five moments in a process that runs many more times than this, and three familiar behaviours drop out of it. The starting speckle is random, so two identical requests genuinely cannot produce identical pictures unless something fixes that start, which is why a repeat of yesterday's result is a setting rather than a rewording. The description steers every step rather than being consulted once, so changing a word early in a sentence can rearrange a picture entirely. And detail resolves last, in the passes where the broad composition has already settled, which is exactly why fine text and the far corners of a large image are where errors survive. None of that is essential to using these tools, and all of it turns arbitrary-looking behaviour into something predictable.

Seen in the wild

Ask for the identical picture twice and notice you get two different ones, which is the random starting point rather than any inconsistency in the model.
ChatGPT
Change one word in a description and see how much of the whole picture moves, which is the description steering every step rather than being read once.
Google Gemini

Common misconceptions

People assume

It assembles pieces of pictures it has seen.

In fact

It does not hold a library to assemble from. What it learnt was how to move from randomness towards something image-like given a description, and each result is produced afresh from noise. That is a real distinction, and it is separate from the ongoing arguments about what was in the training material.

People assume

More steps will fix a bad result.

In fact

Additional refinement improves things up to a point and then does nothing but cost time. A picture that came back wrong is usually answered by changing the description, or by fixing the starting point and iterating from there, rather than by asking for more of a process that had already finished converging.

Telling them apart

Diffusion model vs Image generation

Diffusion model

The mechanism: starting from noise and refining towards a description, which explains the behaviour you notice.

Image generation

The capability: producing pictures from descriptions, with all the practical questions about rights, consistency and control.

Read the second before buying anything. Read this one when a tool is doing something that looks arbitrary and you would like to know why.

Questions

Why do I get a different picture from the same description?: Because the process starts from random noise and that randomness is different each time. It is not inconsistency in the model. Where a product lets you set the starting point, repeating an exact result becomes possible, which is the setting worth finding if you are trying to produce a series.
Why can it not change just one part of a picture?: Because the description influences every step of the refinement rather than being applied to a region, so a small change to the wording moves the whole image. Some products offer ways of masking an area and regenerating only that, which is a different feature layered on top and worth asking about specifically.
Is this the same kind of model as the ones that write?: No, and they are built and trained quite differently. Text models predict what comes next, one piece at a time, in a fixed direction. These refine a whole picture at once, repeatedly. Products increasingly present both behind one interface, which makes the underlying difference easy to miss.

Key takeaways

It starts from random noise and repeatedly removes randomness, steering towards your description.
The random start is why the same description gives different pictures each time.
The description influences every step, which is why small wording changes move the whole image.
Extra refinement steps help up to a point and then only cost time.
It is a different kind of model from the ones that write, despite sharing an interface.

Tools that use this

ChatGPT
The same description twice, showing the random starting point.
Google Gemini
One changed word moving the whole picture.

Last checked July 2026

All glossary terms