search

dunbar12138 / pix2pix3D

pix2pix3D: Generating 3D Objects from 2D User Inputs
http://www.cs.cmu.edu/~pix2pix3D/
MIT License
1.64k stars 131 forks source link

Pretrained weights from EG3D #15

Open zhouwy19 opened 1 year ago

zhouwy19 commented 1 year ago

Thank you for your great work.

The paper says that 'we can significantly reduce the training time to 4 hours if we initialize parts of our model with pretrained weights from EG3D'. As I known, EG3D just has weights for FFHQ dataset. But pix2pix3D uses CelebaHQ-Mask dataset. So I want to know what weights do you use?

Thank you.

dunbar12138 commented 1 year ago

Hi, thanks for your interest.

We use the FFHQ checkpoint of EG3D as it is also a dataset of human faces. We also report metrics without pre-trained weights for a fair comparison with other baselines.

zhouwy19 commented 1 year ago

Thank you for your quick reply and your great work. I would like to ask about when you will open source code.

zhouwy19 commented 1 year ago

At the same time, I also want to consult some designs of pix2pix3D.

For gt {Ic, Is}, you have adopted reconstruction loss of images. That means that for a given mask Is and a set of random z, all z will generate Ic for the given Is. Will this lead to a deterioration in the diversity of the model.

According to my understanding, without reconstruction loss for images, the network can also complete 'pix2pix' by monitoring the reconstruction loss of Is and the generated mask, and different random z can generate different results.

Thank you.