From a grid of numbers to a picture.
A small convolutional decoder turns a 48×48 grid of codes into a 384-pixel drawing. Walking between the codes of real photographs, it draws shapes no photograph contains.
- 11subjects
- 88photographs
- 384pixels
Loading the network…
Loading…
Open full screen
Two networks. One idea.
Each network learns a compact space of real things. Moving through that space — and drawing every step — is what you see.
A whole object in 24,576 numbers.
A triplane autoencoder packs each real 3D model into 3 × 32 × 32 × 8 numbers. The decoder turns any such code into a coloured surface, evaluated on a dense grid with WebGPU or WebGL and sphere-traced every frame.
- 588objects
- 98classes
- 1.31Mweights
No server. No upload.
Both networks are a few megabytes, load once and run entirely in your browser, on your own GPU. Nothing you do leaves your device.
- 0servers
- 2neural nets
- ~2 MBto start