search

nv-tlabs / GET3D

Other
4.17k stars 374 forks source link

Pretrained Model and Source Code #1

Closed XuankeShi closed 1 year ago

XuankeShi commented 1 year ago

Waiting.......

MMUmedaGIT commented 1 year ago

anxiously!

SteveJunGao commented 1 year ago

We're in the final stage of code review, we will release by the end of this week (likely Friday).

Unfortunately, we won't be able to release the pretrained weights, but we will provide all the details for training the model.

Lhun commented 1 year ago

@SteveJunGao did I read this right? It was only trained for a week to get your pretrained weights which had the results in the video? 🤯 I'm really excited to see people take the code and run with it, I imagine datasets like ObjectNet3D and many others could be used here, and then for the textures and images, Stable Diffusion? I think for scene generation this is really exciting, Ikea and HomeDepot and Quixel have huge datasets...

MMUmedaGIT commented 1 year ago

Can it be trained on just 8 GB of VRAM?

SteveJunGao commented 1 year ago

It only needs 1 week to train the model to get the results in the videos. But we were training on 8 V100 GPUs, and each GPU needs 16GB memory (or 32 GB memory depending on the model you choose). If training on 8 A100 GPUs, it can be ~2.5x faster.

It might be possible to train with 8GB GPU memory if reducing the network size, rendering lower image resolution, etc. We haven't tested this yet, but would be great to see the results!

Please stay tuned! We're releasing it soon, and we can talk more about it with the code!

XuankeShi commented 1 year ago

We're in the final stage of code review, we will release by the end of this week (likely Friday).

Unfortunately, we won't be able to release the pretrained weights, but we will provide all the details for training the model.

I think if you can release the pre-trained model to a certain extent, it will be of great benefit to the publicity and promotion of your work. It can provide great convenience for relevant researchers by providing a pre-trained model. Essentially, we both hope to promote the advancement of computer vision 3D understanding. At the same time, I also wish that your work can be noticed by more people.

SteveJunGao commented 1 year ago

Unfortunately, we are not able to release the pretrained models, due to licensing restrictions of the data.

SteveJunGao commented 1 year ago

Source code is uploaded!

zhanghongyong123456 commented 1 year ago

Source code is uploaded!

Thank you

SteveJunGao commented 1 year ago

As a small update, We have released the pretrained modelshere