The problem is likely not the SVG aspect. It's the fact that LLMs don't have any real cross-training between visual concepts and the syntax that produces them. We have this problem with CSS, too, which LLMs are notoriously bad at.
It's actually easier to do bitmaps which is why the visual models create those. You can describe any painting with a few words about content and style but that doesn't give you the SVG syntax needed to represent those things as shapes.
I had Gemini help me make a sparklines charting component that uses SVG. The SVG itself is relatively simple and is parameterized. So it's really only making an SVG component that gets data, rather than designing icons. I tried to get them to do that without any success. I stick to open source icon libraries and there happens to be a meta list on the front of HN right now
The problem is likely not the SVG aspect. It's the fact that LLMs don't have any real cross-training between visual concepts and the syntax that produces them. We have this problem with CSS, too, which LLMs are notoriously bad at.
It's actually easier to do bitmaps which is why the visual models create those. You can describe any painting with a few words about content and style but that doesn't give you the SVG syntax needed to represent those things as shapes.
I had Gemini help me make a sparklines charting component that uses SVG. The SVG itself is relatively simple and is parameterized. So it's really only making an SVG component that gets data, rather than designing icons. I tried to get them to do that without any success. I stick to open source icon libraries and there happens to be a meta list on the front of HN right now
https://news.ycombinator.com/item?id=46665411