Ha ha, (w) oof.
While the training set for the model has not been released, my understanding is they started with the Objaverse dataset, converted GLBs to vector and face object strings (the same thing the model generates as output and we convert back to display), and then filtered out any objects with more than 500 faces. Whatever was left they used for training their model. Therefore any complex objects wouldn't make it in, only simple ones.
I'm still experimenting to discover what remains, because so long as you stick within the training data you'll usually get decent results, anything outside looks like it... kinda melted in the oven.