I tried using Gemini/Claude/GPT-4/Deepseek. The results fell far short of expectation. What can be done to improve the results, or, are there any generative models more suitable for the task?
You're looking for a vector art GPT and there isn't one yet. The reason for this is that no training data has been assembled that could be used to build a model of vector art, at least AFAIK.