MrTornado24 / Next3D

[CVPR 2023 Highlight] Next3D: Generative Neural Texture Rasterization for 3D-Aware Head Avatars
https://mrtornado24.github.io/Next3D/
472 stars 29 forks source link

Issue about portrait inversion. #16

Open stickice opened 1 year ago

stickice commented 1 year ago

Hi, Thanks for your great work! I notice that the code of One-shot portrait reenactment and stylization is not released, so I try to invert image using PTI. However, I found that the latent code produced by PTI is [1,18,512]. The latent code(ws) in gen_samples_next3d.py is [1,28,512].

https://github.com/MrTornado24/Next3D/blob/bdab84fc12cb61bd1002a3efb2c0a707a1139d97/gen_samples_next3d.py#L198

And I notice that the same latent code in EG3D is [1,14,512]. Could you explain the changes or improvements in Next3D? Thanks a lot!

wamiq-reyaz commented 1 year ago

The code is decomposed into two parts - one for the generation of the texture for the FLAME mesh and the other for the original EG3D components - the triplane generator, the superresolution network and the mouth/teeth generator.

https://github.com/MrTornado24/Next3D/blob/bdab84fc12cb61bd1002a3efb2c0a707a1139d97/training_avatar_texture/triplane_next3d.py#L123