Image -> loose text description -> SVG. Did you notice that ChatGPT writes a verbose description of the image first?