Can somebody explain how the workflow works here exactly? Is the LLM trained on SVG? If so: could it hallucinate SVG properties or so? Or is it a regular image

Can somebody explain how the workflow works here exactly? Is the LLM trained on SVG? If so: could it hallucinate SVG properties or so? Or is it a regular image generating AI that vectorizes raster images afterwards with traditional tooling? I'm a noob in that field but I'm curious about potential risks. ;)