The way this works is it generates python code via gpt4 which generates the requested geometry.
From the python script I parse all the relevant parameters and generate a just in time UI. Here is a small demo. Note that its not very reliable for complex parts.
I think for more complex tasks trying to 1-shot the solution is too difficult. It would probably be good to mimic what humans do and introduce a feedback loop.
I think it should be possible to make an agent that first comes up with steps to model a part and then executes each step while analyzing the result of each step using gpt4-vision and possibly fixing any mistakes.
Its also surprisingly good with chairs (or maybe not so surprisingly given that the prompt contains an example of an L-bracket)
