I baked xkcd-artist: a 1.72 MB Int8 ONNX browser model that plans 12-DOF stick-figure comics & charts in 0.55ms (zero garbled speech bubbles)
Image diffusion models draw decent stick figures, yet speech-bubble text still melts into unreadable soup. Over the weekend I trained and quantized xkcd-artist to separate neural stage planning from vector ink lettering.