i had this idea for an AI workflow:<p>1. describe an svg drawing in great detail<p>1.5. ask the llm to label its drawing because it's fun to watch<p>2. generate the svg<p>3. img-to-img with sdxl<p>it's a silly demo but kinda fun, source: <a href="https://www.val.town/v/substrate/scribble" rel="nofollow">https://www.val.town/v/substrate/scribble</a><p>you can coax Llama3 into generating better SVGs using multiple steps, actually – in a previous version the SVG had gradients, stacked layers, dot grid textures... even animations<p>but i had to do a roundabout thing to rasterize the SVG in the val sandbox so i removed all that<p><a href="https://x.com/vprtwn/status/1811710646729277740" rel="nofollow">https://x.com/vprtwn/status/1811710646729277740</a>