Diffusion: Generate by Un-noising

Start from noise. A network predicts how to remove a little noise at each step. After many steps you get an image (or a latent that a decoder turns into pixels).

Training flips the story: take a real sample, add noise on a schedule, and teach the net to reverse that damage. At inference the model walks the same path backward from pure noise toward structure.

Why this prototype shape matters

A working diffusion demo can show “noise → structure” as the story of generative vision. The artifact itself is the lecture: you watch order appear.

Core loop

flowchart TD
  Noise["Pure noise<br/>or noisy latent"] --> Pred["Network predicts<br/>noise or clean signal"]
  Pred --> Step["Remove a little noise"]
  Step --> More{"More steps?"}
  More -->|yes| Pred
  More -->|no| Image["Image or decoded latent"]

Training vs inference

Phase	Input	Job
Train	Real sample + known noise level	Predict the noise (or the clean target)
Infer	Random noise + step schedule	Iteratively denoise toward a sample

Contrast

• Diffusion makes images (or other signals) by sampling.
• Computer vision reads images into measurements about the world.

Both often share encoders and latents; the direction of the arrow differs.

Run

File	What it does
toy_denoise_1d.py	1D signal: train a tiny noise predictor, then denoise from pure noise (stdlib only)
noise_to_structure.html	2D canvas: scrub through denoise steps in the browser

# Terminal — watch MSE fall as structure returns
python diffusion/toy_denoise_1d.py
python diffusion/toy_denoise_1d.py --steps 40 --save diffusion/out_denoise.ppm

# Browser — open the HTML file locally (no server needed)
# Linux: xdg-open diffusion/noise_to_structure.html
# macOS: open diffusion/noise_to_structure.html

WHAT YOU JUST SAW

Training adds known noise to a clean signal and teaches a network to predict that noise. You had the ground truth — the model learns the forward corruption.

Inference starts from random noise and walks backward. No clean sample is given; the network must guess what to remove at each step. That train/infer asymmetry is the heart of diffusion: learn destruction forward, perform creation backward.