FrozenBurning / Text2Light

[SIGGRAPH Asia 2022] Text2Light: Zero-Shot Text-Driven HDR Panorama Generation
https://frozenburning.github.io/projects/text2light/
Other
598 stars 47 forks source link

lighting estimation #3

Closed Len-Li closed 2 years ago

Len-Li commented 2 years ago

Hi, thanks for sharing your great work!

I am curious that is it possible to do lighting estimation using GAN inversion as in Stylelight?

FrozenBurning commented 2 years ago

It is always possible to do the inversion. One naive way is that you can use the encoders of our dual codebook to derive latent representations from input images/patches. And then use that as a condition to sample the full panoramas. Another possible way would be using gradient-based optimization. You may need careful implementation of back-propagation when dealing with the quantization in codebooks.

Len-Li commented 2 years ago

It is always possible to do the inversion. One naive way is that you can use the encoders of our dual codebook to derive latent representations from input images/patches. And then use that as a condition to sample the full panoramas. Another possible way would be using gradient-based optimization. You may need careful implementation of back-propagation when dealing with the quantization in codebooks.

Thanks for your reply. I write a simple script to optimize a latent code, it seems that latent code does not update. Do you have any advice to deal with the quantization in codebooks?

FrozenBurning commented 2 years ago

You may refer to some optimization-based techniques which deal with back-propagation through quantized codebooks.