search

wyysf-98 / CraftsMan

CraftsMan: High-fidelity Mesh Generation with 3D Native Diffusion and Interactive Geometry Refiner
https://craftsman3d.github.io/
430 stars 22 forks source link

Problem about aligned VAE #23

Closed zkaiWu closed 2 months ago

zkaiWu commented 3 months ago

Thanks for your work. And I have a question, what are the differences between michelangelo aligned vae and michelangelo vae

wyysf-98 commented 2 months ago

Hi, the aligned-vae is to compatibility with the original Michelangelo code: https://github.com/NeuralCarver/Michelangelo, and they use contrastive learning to align the encoded shape latent space align with the CLIP. We also train another shape vae without this constrain, namely w/o aligned vae

zkaiWu commented 2 months ago

Are there any performance gaps between them?

wyysf-98 commented 2 months ago

theoretically, I think VAE without alignment may have higher bounds for reconstruction performance as it has less constrain. But the released code may have the similar performance after my evaluation

zkaiWu commented 2 months ago

Thanks a lot!!