Question about render quality and multi-scale about torch-ngp HOT 5 CLOSED

Yes, I'm still trying to figure out what causes the low quality. However, the multi-scale density grid is quite tricky in implementation, and I'm not 100% sure it causes the blurring, so the progress may be slow. (I assume you are talking about the multi-scale density grid in ray marching, not the hash encoder?)
Besides, if you are looking for better quality (instead of speed), you can try without --cuda_ray and increase --num_steps for now.

from torch-ngp.

The real reason is instant-ngp use all train/val/test set to train, which is set to default in GUI mode now. You can try it now!

from torch-ngp.

@ashawkey Thanks for your awesome work! I have tried official instant-ngp with only training data, still achieving high performance, >34 PSNR with 1 min on LEGO. I wonder if any clue has been figured out for the quality issue.

from torch-ngp.

@JaminFong You are right, the performance in PSNR is still worse compared to instant-ngp (as I mentioned in readme's update logs). Honestly, I haven't figured out why for now, but I'm still trying.
If you have any idea, welcome to report in issue/pull request.

from torch-ngp.

@ashawkey Thanks for the reply! I really admire your implementation work. I am also delving into ngp and happy for further discussion if any idea got.

from torch-ngp.