It does not draw.It develops.
Drag one control and watch a picture come out of pure random static, then drag it the other way and put the static back.
Drag the step.
Watch it come out.
In most image generators every run begins as a screen of random static. Drag the step control to take the noise away and let a picture form, then drag it the other way to put the noise back. Generating a picture runs the noising process in reverse, so both directions are worth watching.
You are looking at pure static, so every one of these starts out identical. Take the noise away and they part company.
The contact sheet
Frames taken at points across the same run, so you can see the noise coming off gradually rather than all at once. Choose one to jump there.
Forwards is how a picture gets made. Backwards is the noising the model learned to undo, adding noise to a real photograph a little at a time until nothing recognisable was left.
Each step takes off a little of what is left. With fewer steps every one has to do more, and the picture ends up softer and grainier. Past a point more steps change almost nothing and you are only spending time.
Detail reached stops moving well before the dial runs out. Noise left at the end keeps shrinking all the way, but in fractions of a percent that nothing in the picture shows.
The seed decides the starting static and the details that grow out of it. Move the dial away and come back to the same number, and the same picture returns.
The picture here is drawn by this page to show the shape of the process, and no image model is running in it. A real generator invents its detail as it goes, so its halfway frames look like plausible pictures of something else rather than a blurred version of the finished one. Provided by the Institute of AI for interest and learning.
Why static
becomes a picture.
It starts from static
There is no blank canvas and no album of pictures to cut and paste from. Every run begins with a field of random values, and the picture is built out of that field.
Training ran the other way
Real pictures had noise added to them a little at a time until nothing recognisable was left, and the model learned to undo one of those additions. Undoing them one after another is called denoising, which simply means taking a little of the noise away, and the whole method is called diffusion.
The words steer every step
The description is turned into numbers that sit alongside the picture, and they pull each guess towards something that matches the words. The description shapes the whole run rather than being applied at the finish.
The seed fixes the start
The starting static comes from a seed. Keep the words, the seed and the number of steps the same and the same picture comes back, which is how people return to a result they liked.
Membership is free. Accreditation is the standard.
Keep
exploring.
Get AI is a collection of games, experiences, and tools that make AI easier to understand from the Institute of AI.