STARPOWER · WVY-IMAGE-COMPOSITION-101

How to Generate Intentionally

This isn't an image model you use. It's one you learn from. A 7.8M-parameter model built from scratch, trained on 6 transparent PNGs and 8 colors, that draws scenes it was never shown.

Try it

Pick the pieces and the model draws the scene right here in your browser, no server. Each result tells you whether the model ever saw that exact scene.

Top animal
Bottom animal
Background
Prompt (you can type it too)
Generated by the model
Correct answer (stacked by hand)
Loading the model (15 MB)…

Never seen · flipped stack
Hidden in every color. It only saw the reverse stack, so it has to flip it from word order alone. 32 scenes.
Never seen · new color
It knows this arrangement, just never on this background. 87 scenes.
Seen in training
It studied this exact image, so this checks memory, not invention. 361 scenes.
Outside its world
The same animal twice was never in the dataset at all. Worth trying, to see where it breaks.

How it works

01Few clean ingredients6 animal PNGs + 8 colors. That's the whole world it knows.
02Stack them like layersBackground → bottom animal → top animal, with an exact caption for each.
03Hold some back480 possible scenes. It only studies 361, and the rest are the test.
04Words + positions → pixelsEvery word gets a tag for where it sits in the sentence, so "cat … dog" ≠ "dog … cat". Then 2 layers of neurons turn that into a 48×48 image.

Look closely at never-seen results and you'll spot faint ghosts: blurry outlines of other animals in the background. That's the model blending in "the average of everything it ever saw in that spot," the seam between memorizing and generalizing.