search

Kunhao-Liu / StyleRF

[CVPR 2023] StyleRF: Zero-shot 3D Style Transfer of Neural Radiance Fields
https://kunhao-liu.github.io/StyleRF/
144 stars 11 forks source link

Question for room dataset #25

Closed SYSUykLin closed 8 months ago

SYSUykLin commented 8 months ago

We run your code on llff room dataset. And we use the images_8(downsample 8 times). In the second stage, train_feature.py can only achieve 15 PSNR. Does it normal? In the first stage, train.py can achieve 35 PSNR.

Kunhao-Liu commented 8 months ago

Hi, it is not normal. You can try training with images_4.

SYSUykLin commented 8 months ago

Thanks for your reply. Another question, do u try replica dataset? How large RAM does it need? I use your model as baseline to compare

SYSUykLin commented 8 months ago

I try blender dataset and my RAM is not enough(64G RAM).

Kunhao-Liu commented 8 months ago

If the RAM is not enough, you can use only half of the images or even fewer in the second stage. The performance won't be affected too much.