🌫️ From the noise - picture
Neural network artists do NOT start with a blank slate, but with pure noise - like ripples on an old TV. Choose what to “guess” and move the slider: with each step the model removes some of the noise until a picture emerges from the chaos.
What we draw:
How it works: diffusion
The model was trained the other way around: they took millions of real pictures and gradually NOISE them to complete ripples, remembering every step. So she learned to predict what a “slightly less noisy” version of any noise looks like.
When you ask for a picture, the model starts with pure random noise and applies this skill many times: removing a drop of noise, then another, and another - dozens of steps. Your prompt (what to draw) at every step tells you WHICH direction to remove the noise so that exactly what is needed appears.
Therefore, more steps mean a cleaner result, and an accurate prompt means a more accurate picture. The same principle of “many small refinements instead of one jump” works with text: a clear query leads the model to the desired result step by step. You can practice in a free lesson.
Start for free