reproduce problem

[VLadImirluren]

I use your house geometry and apperance on 4 * 4090, but still meet problem.
How can I solve..
Thanks

[VLadImirluren]

there are too much noise on rgb and geometry..

[VLadImirluren]

and the rgb not right...such as the door on the left hand side. I try for many times, but problem still exist

[VLadImirluren]

and the rgb not right...such as the door on the left hand side. I try for many times, but problem still exist

the rgb doesn't match the geometry....

[ARuiChen]

1、set "if_use_bump = False" in config file if your appearance has too much noise. This means closing the prediction of Kn in the paper. Actually, this may also be due to the background is fixed black. You can add "black background" in the prompt to alleviate this problem.
2、Strategy 0 is more prone to texture and geometric misalignment, and using strategy 1 or 2 can alleviate this problem. At the same time, it is necessary to adjust the range of FOV and set the minimum value higher. The problem of misalignment is mainly caused by the small FOV of certain perspectives. But please note that if the FOV range is set improperly, the details may not be rich enough.
3、 Train longer.

[VLadImirluren]

1、set "if_use_bump == False" if your appearance has too much noise. Actually, this is because the background is fixed black. You can add "black background" in the prompt to alleviate this problem 2、Strategy 0 is more prone to texture and geometric misalignment, and using strategy 1 or 2 can alleviate this problem. At the same time, it is necessary to adjust the range of FOV and set the minimum value higher. The problem of misalignment is mainly caused by the small FOV of certain perspectives. But please note that if the FOV range is set improperly, the details may not be rich enough. 3、 Train longer.

thanks for you advice.
I am curious of the reason why I can't reproduce correctly by using your config.
What's the main point you think?
thanks

[ARuiChen]

Because the number of viewpoints used in one iteration is different (the number of GPUs is different). The sampling algorithm I proposed in the supplementary materials contributes to global consistency in appearance and geometry modeling, but it depends on a large batch size.

[VLadImirluren]

Because the number of viewpoints used in one iteration is different (the number of GPUs is different). The sampling algorithm I proposed in the attached materials contributes to global consistency in appearance and geometry, but it depends on a large batch size.

Do you think I can solve this problem by using double batchsize on 4gpu？

[ARuiChen]

I think this is possible, but I haven't tried it before, so I need you to give it a try and let me know if it works.

[VLadImirluren]

I think this is possible, but I haven't tried it before, so I need you to give it a try and let me know if it works.

okay, I will

[VLadImirluren]

I think this is possible, but I haven't tried it before, so I need you to give it a try and let me know if it works.

yes, it works. thank you

[ARuiChen]

Good to hear that. Feel free to ask again if you have any questions.