Day 9 · Un-Smash the Picture

✨ Lab · Noise → Picture

Press it. Watch a picture appear.

On the left is pure static — random fuzz, no picture at all. Pick what you're aiming for, hit Denoise, and watch the machine remove a little noise each step until the picture emerges on the right. This is the money moment. 🪄

start: pure static 🌫️

→

step 0 / 20

🤝 Honest note: a REAL model spent weeks learning to guess the clean picture hiding in the noise. Our toy cheats — it peeks at the target picture to know what to remove. But the step-by-step motion you're watching is exactly what real diffusion does. The dance is real; only the "brain" is faked.

🎚️ Play · The Steps Dial

How many steps? Speed vs. quality.

Every real AI art tool has this exact dial: number of sampling steps. Few steps = fast but rough. Many steps = slow but crisp. Set the dial, press Paint, and compare the results yourself.

fast & rough20slow & crisp

stepscrispness: —

press Paint

Try both ends: paint at 30 steps (crisp!), then drag down to 4 and paint again (rough!). More steps means the model takes smaller, gentler denoise nudges — so it lands more precisely on a clean picture. 🎯

🐞 Break-It · Too Few Steps

Only 2 steps? Get a blob.

Slam the steps down to the minimum and paint. The model doesn't get enough tries to clean up the static, so you get a rough, blotchy mess. Press it a few times.

2 steps · messy!

Real-world tie-in: that blobby look is exactly why super-fast image generators can look rough — they took fewer denoising steps to save time. Speed always costs a little quality. ⚡

🏷️ The pro wordsRunning the melt backwards is reverse diffusion, and each cleanup move is denoising. The number of moves is your sampling steps, and the recipe that decides how much to remove each time is the scheduler (real ones are named DDPM and DDIM). The model's actual job is to predict the noise to subtract — and the network that does it is usually a U-Net. You just ran the whole loop by hand.

🎯 Boss · Budget Painter

Hit the target within budget

You're a pro artist on a deadline. Each round gives you a crispness target and a step budget (fewer steps = faster = cheaper). Pick just enough steps to reach the target without blowing your budget. Win 3 rounds for the badge!

Loading round…

your steps5

Won: 0 / 3

your painting

🚀 Beyond the Basics

From "any picture" to "the picture I asked for"

Right now our toy peeks at a target to know what to build. A real model doesn't peek — it learned. But that raises a question: if it just makes some clean picture, how do you get the one you want?

✨

Denoise (today)

Static → a picture, one predicted-noise step at a time. The core engine of every AI image tool.

💬

Guidance (Day 10)

Whisper "a red dragon" at every step and the denoiser steers toward it. Words become pictures.

🎚️

Your controls

Steps, seed, and prompt strength — the exact knobs you tuned here are the real ones in Stable Diffusion & friends.

🏷️ Why no giant model here?Real reverse diffusion needs a trained U-Net and a beefy GPU — too heavy for a 4GB laptop. So our toy peeks at the answer instead of predicting it. Everything else — the schedule, the step-by-step denoise motion, the steps/quality trade-off — is the real thing, honest and tiny. 💜

🎉 Day 9 Complete!

You ran reverse diffusion — the actual engine behind AI art. Start from static, remove a little noise each step, and a picture appears. You also found the real speed-vs-quality trade-off every artist tunes: more sampling steps = crisper, slower.

Day 10: Words into Pictures →

Tomorrow: we stop peeking at a target. Instead you'll type words and steer the denoiser toward them, every single step. Type "dragon," get a dragon. That's text-to-image. 💬✨