FilterHN

I understand prose and code and raster images, but how does SVG work? The SVG structure and visual result seem so disconnected, I would have thought current LLMs would have a harder time on arbitrary descriptions of imagery.

▲

XCSme

2 days ago

[-]

Svgs are just shapes, they (models) mostly only use sqaures, circles, etc. Rarely paths.

Plus most models are multi-modal, so they have a visual understanding too.

▲

happytoexplain

2 days ago

[-]

Right, but rendering the visual understanding is based on neighboring pixels normally, which is much more directly related to the visual than plaintext like SVG tags.

▲

XCSme

2 days ago

[-]

I kind of disagree, because describing an image is more like writing an svg than rendering pixels.

If I ask you to describe how would you draw a cat head, you could do it in text like: "a big circle as the head, 2 small circles as the eyes, 2 triangles for ears, 3 lines on each side of the mouth as moustaches, etc..."

▲

happytoexplain

2 days ago

[-]

Hmm, sure. I'm still surprised - it also has to say where they are in coordinate space. It feels like the way genAI works ("what comes next") is not amenable to this use case (demonstrably I'm wrong, of course).

▲

XCSme

2 days ago

[-]

I remember when I was learning WebGL/OpenGL and had to draw some test shapes to test my shader, I would manually think what the vertex positions should be to draw a triangle, a pyramid, etc. I think for AIs it's quite easy, because most are quite decent at math, and have been trained on many geometry problems, and probably also a lot of OpenGL code too, and 3D assets.

▲

tmaly

2 days ago

[-]

the DeepSeek V4 Flash is impressive for the cost

▲

XCSme

2 days ago

[-]

I use V4 Flash as my main model, it really is exceptionally capable for the price.

It can follow instructions quite well, if you have a plan, it usually executes it to completion and you get code that works.

My only concern with it, is that it's mostly served by Chinese companies on OpenRouter.