My code modification is like this

I refer to the dragon_specular.ini format and create the checkpoints.ini file myself

When running nlt_test.py I'm getting the following error:

Running postproc.py to generate the json gives the following error:

Am I missing some files to run postproc.py?

Below is the folder structure I'm using:

With the following files within ./checkpoints/test_30/

And here is the content of the checkpoints.ini

Hi, Since I want to relight the human image for a given HDRI map, can I test it on my own image data and how? Thanks

1. I want to run the nlt_test.py code. I don't know how to specify the CKPT path，
   Can I just specify the path
   
2. I don't quite understand how CKPT-30 and CKPT-10 are used. Could you give me a demo
   
3. If the realization of custom characters in the new scene of the image synthesis, how to achieve, I hope you can give some advice
   I hope to receive your prompt reply. Thank you.

Hello, I am a beginner in neural rendering, I hope you can give me some code examples, about light and shadow re-illumination，like this example

Hi,
This is a very gread project, Thanks for open source. I have a question: how to use this project achieve human image-based relighting? Looking forward to reply, Thanks.

This is a very well-designed experiment, so please do not take my words to be overly critical. It is just that what you are proposing in your paper is not entirely true. You say:

"With this learned LT, one can relight the scene photorealistically with a directional light or an HDRI map, synthesize novel views with view-dependent effects, or do both simultaneously, all in a unified framework using a set of sparse, previously seen observations. "

This last statement would indicate that you could take a small number of photographs from different perspectives (such as those presented in the instant-ngp model) and then relight a static observation / scene (i.e. single photograph). Reading more deeply into your paper, this proposal is contradicted. In fact, not only does this project require OLAT data as an input (specifically, hundreds of OLAT images from over 60 RGB cameras placed around the performer) --- which could hardly be qualified as "sparse" -- but it appears you also require a proxy mesh (distributed in a Blender file). I do not see that any rendering is possible without a mesh --- and a reliable, high-resolution mesh cannot be generated using a set of sparse, previously seen observations, as you state.

To make a long story short, all of this could already be done many years ago using photogrammetry and GPU-based rendering engines (such as Cycles, Redshift, and Octane). Perhaps you can explain to me why we would want to take the time to invest thousands of dollars in an OLAT-capture cage when in just a fraction of the time with a single DSLR camera (and a bit of patience), the same quality output could be achieved? This isn't theoretical, I have done it.

In other words, this project would only be useful if it were able to relight scenes from a small set of images. Clearly, from the questions being posted here, the average reader has misunderstood what the original paper was proposing. Perhaps if the language were simplified or at least brought closer to reality, we wouldn't waste our time looking at the code thinking that it could be used to "relight a scene from a set of sparse views", OLAT-derived or not. It cannot.

TL;DR -- instead of downloading 19GB of "rendered data" for a dragon, just take a 2Mb OBJ of the same dragon, import it into Blender, put some point-lights / HDRI around it, and render it in Cycles. No need to mess with "neural light transport", which would require 1) having access to a dragon figurine and 2) taking hundreds of photos of that dragon with actual lights (in a studio or OLAT cage), crossing your fingers, and hoping for the same results as the OBJ in Cycles.

It is likely I am missing the point of this paper. It seems silly to go to so much trouble to do something that has another more simplistic approach that has been around for over a decade now.

Hi,

I was trying to run NLT on my own dataset, and I found that while trying to run I ran into an error to run gen_file_stats.py, but that file doesn't seem to exist. How can I generate Data Status w/o it or where does the file exist?

Thanks!

Edit: I see that postproc.py actually does this, but it might be worth updating the error msg.

please you can add a demo colab to test your own images