Deeper

Lab · c22.s04.L1

Noise and denoise

A denoising step does not paint; it replaces the noisy picture with the model's estimate of the finished picture plus a smaller amount of noise, so repeating the swap sharpens the picture.

Question: How does removing noise create an image?

Noise and denoiseTwo square panels. The left, what the model sees, starts as pure static and only near the last step resolves into a pale round face on a dark ground. The right, the model's guess, starts as a soft pale disc and sharpens sooner, showing two dark eyes and a dark mouth line. A bar shows the noise falling from 100% to 0%, an error readout falling from 0.50 to 0.00, and Play and Step move through the steps.Noise and denoiseHow does removing noise create an image?what the model seesthe model's guessnoise lefterror of the guessfrom the finished picturedenoiser blurOne step does not paint.It replaces the noisy picture with the model's guess at the finished picture plus a smaller dose of noise.Repeat the swap: the guess sharpens and what is left is a little more of the picture.Space plays or pauses; Left and Right step one recorded step; Home and End go to the first and last step.Step 0 of 20schedule step t = 999100%0.504.47 px
Synthetic Synthetic: the 28 x 28 face is drawn by data/scripts/c22.s04.L1.py, the starting noise is one fixed random draw (seed 7), and the denoiser is a stand-in that returns the true picture blurred by an amount that falls from 4.5 pixels to zero as the noise falls. No trained network is involved; a real diffusion model learns this estimate from data.
Lab controls
600 ms per step

Left and Down slow by 200 ms, Right and Up speed up by 200 ms; Home sets 200 ms, End 1600 ms

Frame 1 of 16: Step 0 of 20100% of the noise is left; the denoiser blur is 4.47 px and the error is 0.50. The left panel is pure static: the starting noise. The guess is already a soft pale disc. The model supplies the shape; the noise supplies the variation.

Keyboard: Space plays or pauses, Left and Right step, Home goes to the first frame, End goes to the last frame.

Text description of this lab

Two square panels. The left, what the model sees, starts as pure static and only near the last step resolves into a pale round face on a dark ground. The right, the model's guess, starts as a soft pale disc and sharpens sooner, showing two dark eyes and a dark mouth line. A bar shows the noise falling from 100% to 0%, an error readout falling from 0.50 to 0.00, and Play and Step move through the steps.

Return to the section that hosts this lab

Caption status: synthetic. Every lab declares one of published, re-simulated, schematic, synthetic or interpolated.