🧠 A Photo in a 12KB Neural Net

Don't store the pixels — store a function. A tiny network learns f(x, y) → color until the function is your photo, then the whole image fits in ~12KB of weights. Watch it fight JPEG at the same file size.

original

the 12KB network

render

JPEG, same budget

untrained

Hit ▶ compress and watch a blur become your photo as the network trains (a minute or two is plenty). Then try the 4× render — the net is a continuous function, so unlike pixels it can be sampled at any resolution. Export the weights: that JSON file is the photo.