An image model doesn't draw it removes noise.

Ask an AI for a picture and it does not start with a blank canvas. It starts with a full frame of random static — no subject, no edges, nothing — and every step after that only ever takes something away.

At each step the model predicts which specks in the frame are noise, and deletes them. Then it does it again, dozens of times over. The picture is never painted on; it is what is left standing when everything that wasn't the picture has been removed. This is why the process is called diffusion, and why it runs backwards from the way people imagine it: not adding detail, but subtracting randomness.

WHERE THE PROMPT ACTUALLY ACTS

The prompt does not draw anything either. Its job is to steer the judgement — to change what the model counts as noise on each pass. Describe a mountain at sunrise and you are not sketching a mountain; you are biasing, several dozen times in a row, which specks survive the next round of deletion. The words shape the filter, not the image.

WHAT THIS EXPLAINS

It explains why these models produce an image at all when given nothing but a sentence and a random seed: the seed IS the starting frame. Change it and the same prompt yields a different picture, because a different pile of static was standing there at step one. It also explains why the process is slow — each pass is a full prediction over the whole frame — and why partial results look like a photograph developing rather than a drawing being sketched.

ONE THING THIS VIDEO DOES NOT CLAIM

It does not explain HOW the model predicts which specks are noise. That is a trained network, and it is not something thirty seconds can honestly cover. The video treats the prediction as given and spends its time on what is done with it. And to be precise about the ending: the picture is not hidden inside the static waiting to be found — that is a nice image and it is false. What is true is the arithmetic. Nothing is added; things are only ever taken away.

HOW THIS VIDEO WAS MADE

Every frame is code — React and SVG rendered with Remotion. No stock footage, no AI-generated video, no sampled audio. The narration is synthesized text-to-speech: a generic voice generated from text, not a copy of anyone.

The counter on screen is real, not a prop. The field starts as 2200 individual specks and the video never creates another one — REMOVED counts the specks actually deleted at that moment in the timeline, and ADDED reads zero in every single frame, because the animation is built so that adding one is impossible. The closing line is something you can check against the screen rather than take on trust.

Deliberately absent: no model, product or company is named or drawn anywhere, and no prompt text ever appears on screen. The picture that emerges is a generic sun over two ridges — abstract on purpose, so it reads as "an image" rather than as any real output. There is no step count shown, because the number of steps is a setting someone chooses, not a fact about how the model works.

#ai #aiart #imagegeneration #diffusion #howitworks