GPU memory usage and incorrect rendering?

Hi, @derksend , nice work and presentation. However, I've got some questions. Hope you don't mind.

1) It seems to me that shadow neural radiance fields somehow are GPU memory hungry. I have to set the image downsampling factor to 8 to feed the Data Fusion Contest images into a single GPU with 12G memory. The training image resolution seems to be only 128x128, shown below in the ’train.png'. Any idea of training shadow neural radiance fields on high-res (2048x2048) images on limited-memory GPU?

2) The 16th row in the attached ’train.png’ seems to have strange patterns; only the diagonal of the images are correctly rendered. Any idea why this is the case? I’ve observed such phenomena on all the four sites I tried: JAX_165, JAX_166, JAX_167, JAX_168.

Looking forward to your reply!

Thank you,
Kai