Do two networks have to be used to model color and density separately

[BCL123456-BAL]

I tried fitting the color and density with an MLP like the original nerf. There is no problem with training and testing, but when doing color editing, there is a problem.

Test images during training：

Result after color editing：

[yuehaowang]

Thanks for your question.

In our implementation, we model density and appearance in two separate fields.
Nevertheless, I believe fitting density and appearance with one MLP is also okay since our method does not require any parameter fine-tuning. I guess the issue you met may stem from misuse of the density and appearance channels.

[BCL123456-BAL]

Thank you for your reply, I have found the problem, which lies in the processing of light.