ImgPilot

A canvas on the left, a painting on the right. Outline a jar, scribble two green stems into it, and the right-hand pane fills in with a watercolour of flowers in a glass vase — not after you press a button, but while the pencil is still moving.

I created it and built it at Lepton AI. The interesting constraint was latency, not quality. Image models were already good enough; what they were not was fast enough to sit inside a feedback loop. A latent consistency model reaches an image in four denoising steps rather than fifty, and four steps on a single A10 keeps up with a hand moving on a canvas. That changes what the prompt box is for — you stop writing a paragraph and hoping, and start steering something you can already see.

The canvas is Excalidraw. Every change exports it to a bitmap that goes to the model as the image-to-image input at strength 0.7 — high enough to repaint everything, low enough that your composition survives. The prompt is split across two fields, what you are drawing and what it should be painted like, because the second half is the part people cannot invent on the spot: it is a menu of five media and nine movements, so "oil painting, impressionism" is two clicks instead of a sentence.

The back end is a Lepton photon — the model, its dependencies and its resource shape in one Python file — which is what let the whole thing be a demo of the platform as much as a toy.

The hosted service was shut down in June 2024 and imgpilot.com is offline. The repository is still up, front end and photon both, and runs against your own GPU.