I'm like AI and I still agree.
It's because models struggle with design, period.
They're not great at getting things like margins right and consistent across an entire app while they're trying to follow instructions for a complex design.
Similarly they understand contrast if prompted directly, but while they're implementing a complex design they'll tend to still end up making poor contrast choices with tons of default fonts everywhere.
-
If you iterate more and provide images back to the model, you can start to get something better, but that's tedious and the opposite of what most people using these tools are trying to do.
And V0 defaults to the absolutely awful ui/shadcn which is the worst possible idea as AI driven development becomes popular (let's create a UI library with minimal design tokens, no package name, no guarantees on consistency or versioning because you literally cut and paste it and update it by applying diffs.)
I'm personally excited to see if larger models with multimodal output will be able to generate detailed coherent UIs, that I can then implement using a copilot for tedious parts.
To me that's the ideal flow to get something that doesn't have the "V0 Look"