Diffusion Models, Playable

EP 01

The forward process — drown a picture in noise

Training starts with destruction. Take a clean image, add a little Gaussian noise, then a little more, until nothing is left but static. Drag the slider — you are the forward process. Every intermediate frame becomes a training example: “given this mess, what did the cleaner version look like?”

Noise step t t = 0 (clean)

Left: your image at step t. Right: what the model must learn — the noise that was added.

The trick: destroying is easy and perfectly known. So the model never has to “learn to paint” — it only learns to answer “what noise was just added?”, a much easier question, millions of times.

EP 02

The reverse process — order out of static

Generation runs the film backwards. Start from pure random points and repeatedly remove a little predicted noise. Watch 1,200 particles that begin as formless static get nudged, step by step, into a shape. Every step is small; the miracle is the accumulation.

step 0 / 60

Conceptual demo: the “denoiser” here knows the target distribution directly; a real model learns it from data. The choreography — noise → small steps → structure — is exactly the same.

Key intuition: no single step creates the image. Each step only makes the cloud slightly less improbable. Sixty tiny corrections later, improbability has nowhere left to hide.

EP 03

The prompt is a steering wheel

Same starting noise, different destinations. A text prompt doesn't select a stored picture — it tilts every denoising step toward regions that match the description. Pick a “prompt” below and generate from the identical noise seed. The static is the same; the pull is different.

seed #7 · pick a prompt

Why “same seed, same vibe”: artists reuse seeds because the seed fixes the noise, and the noise fixes the composition's skeleton. The prompt then decides what that skeleton becomes. You just did the same thing.