magic-research / getavatar Goto Github PK

Hello, Thank you for your greate work.

I have a question on the training code.
I can see that "Gmain" and "Gnorm" phase are separated.
Isn't normal map can be generated during "Gmain" phase?

Is there a reason you have separated normal and texture generation?

According to the provided training instructions, the training effect of the normal map is much worse than that of the texture. May I ask what is the specific reason or are there any ways to improve it? Thanks very much！
!

Hello, thank you for sharing your work. I have a few questions on your model. My first question is that does your model output obj file?
And my second question is that does your model generate texture ?

Thanks.

Hi,such a great work! but i get some problems
May I ask how to implement queries on textures and normals, and where are these implemented in the code?

Can the model train on 3090 Ti? And it only have 24G memory.

Hello while experimenting I am facing weird behavior and would like to ask for tips.

When training, the normal vector of the generated mesh gets flipped.

So like in 20th epoch, the normal vector is facing outwards, however around 40th epoch the normal vector starts becoming "noisy". And in 40th epoch the normal vector get completely flipped so facing inwards.

Have you guys encountered this kind of phenomenon ? Could I get a tip on what may be causing this issue please?

May I ask why I am running the code of render rgb images of THUMAN2.0 with blender the following error occurred：
Blender 3.2.2 (hash bcfdb14560e7 built 2022-08-02 23:38:01)
Read blend: /home/youzehong/jt/blender-3.2.2-linux-x64/GETAvatar/preprocess/test.blend
GPU 0 enabled
non-existent directory
Error: Cannot read 'white_bg.png': No such file or directory
RuntimeError: Error: Cannot read 'white_bg.png': No such file or directory

If I want to train the model on other non 3D scanning format datasets such as deepfashion, do I need to obtain SMPL labels in the following format:
label info:
0:16 world2camera_matrix shape (16,)
16:19 global_orient shape (3,)
19:88 body_pose shape (69,)
88:98 betas shape (10,)
"""
If so, could you introduce me a work that can obtain SMPL parameters in this format? Thank you very much!

I found that the 3D scans in the official store of Render People is very expensive (About 29 USD one scan). May I ask if this work has cost a lot of money on this dataset?

Thanks for such an awesome work! How could I export the obj files of the human meshes? I notice there is this function: extract_mesh_with_marching_cube. Can I use this in the script gen_multiview for mesh extraction?