Here's a secret that sounds fake but isn't: modern AI art doesn't paint on a big picture. It squishes everything down to a tiny canvas, does all its thinking there, then blows it back up. That squished world is called the latent — and walking through it makes pictures morph. 🌈
Pick a picture. The machine encodes it into a tiny N×N grid (the middle box — that's the whole picture, stored in just a few color blocks!), then a decoder blows it back to full size on the right. Drag the N slider and watch the trade-off: fewer blocks = smaller & blurrier, more blocks = detailed but heavier.
grid size (N×N blocks)
The AI doesn't store every pixel — it stores the tiny grid and imagines the details back. That's why it's fast enough to run at all. 💾
Here's the magic trick. Squish two pictures into latents, then take a step-by-step walk from one to the other. Every in-between step decodes into a real picture — so A smoothly morphs into B. Slide t, or hit ▶ Auto-morph and watch.
I froze a morph at a secret spot between 🅰️ and 🅱️ (shown as thumbnails). Study the mix and set your slider to where you think it stopped — 0 = pure A, 1.00 = pure B. Get within 0.12 and it counts. Crack it 3 times to win the badge!
your guess (0 = A, 1 = B)
Squishing to a tiny canvas isn't just a speed hack — that compressed idea-space is where almost every cool AI trick happens:
Diffusion on a 64×64 latent instead of a 512×512 image is ~64× less work. This is the whole reason AI art runs on a laptop.
Walk between two latents (like you just did) and you get smooth morphs — the trick behind "AI transition" videos.
Nudge a latent in the right direction and you can add a smile, change hair, or swap a season — without repainting a single pixel by hand.
You learned the trick that makes AI art possible on a normal computer: it works on a tiny compressed canvas — the latent — then a decoder blows it back up. And walking between two latents gives you smooth morphs. That little squished space is where all the magic lives.