He mentions they used quite different methods to get to the end-result. I'd love to know how much changing the prompt impacts things: like telling claude to only draw and blend, similar to sol, etc.
There's likely not training data for this and Claude seemed to spend a lot of time blending the background.
