Noise to Picture

Noise to Picture

Diffusion models feel like magic: type a prompt, an image appears. This playground shows the actual mechanism instead. Drag one step slider from 0 to 100 and watch a fixed 8×8 picture emerge from a field of pure random noise — one denoising pass at a time, never painted onto a blank canvas. Every pixel is a plain, deterministic blend: target × clarity + noise × (1 − clarity), where clarity rises with the step. You also get the residual noise level and the signal-to-noise ratio climbing as the picture resolves. Runs entirely in your browser (0 uploads, works offline).

How to use it

1. Drag the denoising step slider from 0 (pure noise) toward 100 (clean picture), or tap a preset — Pure noise, Halfway, or Clean picture.
2. Watch the 8×8 canvas resolve and the clarity, noise level, and signal-to-noise readouts update live.
3. Scan the filmstrip below to see the same seed at steps 0 · 25 · 50 · 75 · 100 side by side — the coalescing is the whole point.
4. Hit Reseed noise to start from a different random field; the picture it resolves into never changes, only the noise it climbs out of.
5. Open the X-ray panel to see the literal per-pixel formula, then copy or share your scrub.

What this clears up (the fundamentals)

• Generation is denoising, not painting — the model starts from random noise and removes it step by step. Nothing is drawn stroke by stroke onto a blank canvas.
• The step slider is the noise schedule — each position sets how much noise is left. Early steps are mostly noise; later steps are mostly signal. That progression is the "diffusion process."
• The seed fixes the starting noise — same seed, same starting field, forever. Change the seed and you climb out of a different noise field toward the same target. That is why identical prompts with the same seed are reproducible.
• Signal-to-noise is a number, not a vibe — as the step rises, step / (101 − step) grows from 0 toward large values. "The image got clearer" becomes a measurable ratio.

Where it's used

A multimodal enrichment demo and a highly shareable lead magnet — the fastest way to feel why "AI made a picture" is really iterative denoising. It's a miatz build-lab concept playable: play it here, then learn to build the noise schedule, the seeded field, and the step-by-step reveal yourself inside AIE-130.

FAQ

Does this call a real image model?

No. The picture here is a fixed 8×8 teaching target so you can see the mechanism clearly. A production diffusion model runs the same idea — blend a learned prediction toward less noise each step — over millions of pixels with a neural network doing the denoising.

What is the "clarity" number?

Clarity is step / 100, rounded to two places: how much of each pixel comes from the target versus the noise. At clarity 0 the canvas is pure noise; at clarity 1 it's the clean picture. The residual noise level is simply 1 − clarity.

Why does reseeding change the noise but not the picture?

Because the target is fixed and the seed only sets the starting noise field. Denoising always heads toward the same target, so a new seed just gives it a different noise to climb out of — exactly like re-rolling the seed on a real diffusion model with the same prompt.

What is the signal-to-noise ratio here?

It's step / (101 − step), rounded to two places — 0 at pure noise, 100 at the last clean step. It rises as the denoising step climbs, giving you a single honest number for "how much signal has surfaced."

Is anything uploaded?

No. The noise field, the blend, and every readout are computed entirely in your browser — nothing is transmitted, stored, or logged. Turn off your Wi-Fi and it still works.

Limits

A teaching model, not a real sampler: it uses a linear noise-to-target blend and a fixed 8×8 target, so you can see the schedule instead of a black box. Real diffusion uses a learned denoiser, a non-linear noise schedule, and guidance — the shape of the reveal (noise → signal, step by step, seed-reproducible) is exactly right; the pixel math is deliberately simplified.

Related

Part of the miatz playgrounds. Explore the rest from the playground catalog.

Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.