Multiple input about 3dfacereconstruction-lap HOT 1 CLOSED

Hi，thanks for your interest to our work. We will release the training code later. Right now for multi-input condition, you can concatenate images along channel dimension and feed them to netA and netD, and set rand_inner_batch as the number of images. The target image should be fed into refine_netA and refine_netD.

For the issue on Asian faces, it is due to the bias of training dataset. The released model is fine-tuned on CelebAMask-HQ dataset in which most identities are non-Asian. You may fine-tune the model on custom dataset. For the training losses and implementation, you may refer to unsup3D.

from 3dfacereconstruction-lap.