Improve Principled Hair Chiang importance sampling for tilt #114780

New Issue

Brecht Van Lommel · 2023-11-13T11:00:05+01:00

Brecht Van Lommel commented

2023-11-13 11:00:05 +01:00

The Cycles implementation is based on PBRT.

There was a fix for the importance sampling with tilt, that we should port over:
https://github.com/mmp/pbrt-v3/pull/256

The Cycles implementation is based on PBRT. There was a fix for the importance sampling with tilt, that we should port over: https://github.com/mmp/pbrt-v3/pull/256

Brecht Van Lommel added the

labels 2023-11-13 11:00:05 +01:00

Brecht Van Lommel added this to the Render & Cycles project 2023-11-13 11:00:07 +01:00

Amine Bensalem commented

2023-11-15 13:00:31 +01:00

I'd like to give it a shot :)

Brecht Van Lommel commented

2023-11-15 15:39:46 +01:00

Great, thanks.

👍 1 🚀 1

Amine Bensalem referenced a pull request that will close this issue

2023-11-21 16:11:52 +01:00

WIP: Fix #114780: Cycles: Importance sampling correction #115241

Kanishk-Tiwari commented

2023-12-25 22:15:45 +01:00

Not resolved yet?

Amine Bensalem commented

2023-12-26 13:46:37 +01:00

@Kanishk-Tiwari Still on it , I'm waiting for my new PC , the old one fried

Deyuan Liu commented

2024-01-04 07:39:54 +01:00

Hi, @brecht, I am a student, and I also want to make some contribution to cycles in my spare time, there might be some useful algorithms in pbrt-v4 and not implemented in cycles, can I try to implement these algorithms in cycles? I had read some chapters of pbrt-v4 and some cycles codes (e.g. old version principled-bsdf and some light implementation).

❤️ 1 🚀 1

Brecht Van Lommel commented

2024-01-04 12:58:43 +01:00

@liudeyuan the list of things we want to improve in Cycles is in:
https://projects.blender.org/blender/blender/projects/13

Generally we try to go from problem -> solution. If you find a particular algorithm in pbrt-v4 that solves a problem, on the list or not, then you can explain it and we can check if it's a good fit for Cycles.

@liudeyuan the list of things we want to improve in Cycles is in: https://projects.blender.org/blender/blender/projects/13 Generally we try to go from problem -> solution. If you find a particular algorithm in pbrt-v4 that solves a problem, on the list or not, then you can explain it and we can check if it's a good fit for Cycles.

👍 1 ❤️ 1

Holly Liu commented

2024-02-12 19:57:53 +01:00

Hi @brecht, is this issue still relevant and if so, can I go ahead and test @HamilcarR 's code with a different kind of fur as mentioned in #115241 ? Likely something that is similar to sheep fur as sheep is already tested.

Brecht Van Lommel commented

2024-02-12 20:15:39 +01:00

@holly-l it's still relevant, though I'm not sure if @HamilcarR plans to continue work on this.

What I think the main thing to do is test this on an single close up hair strand, and verify that results are correct. It's really hard to tell on a complex hair setup what is right or wrong.

If it's not clear how to verify if it's correct, it's best to pick another Good First Issue. As I can't really explain what to look for without a background in BSDFs and sampling algorithms.

Helpful for testing this mostly:

Build with WITH_CYCLES_DEBUG=ON
In the preferences, enable "Developer Extras", then in "Experimental" enable Cycles Debug options
The in the render properties, there is a Debug panel with options to compare Direct Sampling, Next Event Estimation and MIS.

@holly-l it's still relevant, though I'm not sure if @HamilcarR plans to continue work on this. What I think the main thing to do is test this on an single close up hair strand, and verify that results are correct. It's really hard to tell on a complex hair setup what is right or wrong. If it's not clear how to verify if it's correct, it's best to pick another Good First Issue. As I can't really explain what to look for without a background in BSDFs and sampling algorithms. Helpful for testing this mostly: * Build with `WITH_CYCLES_DEBUG=ON` * In the preferences, enable "Developer Extras", then in "Experimental" enable Cycles Debug options * The in the render properties, there is a Debug panel with options to compare Direct Sampling, Next Event Estimation and MIS.

👍 1

Amine Bensalem commented

2024-02-12 21:31:57 +01:00

@holly-l it's still relevant, though I'm not sure if @HamilcarR plans to continue work on this.

What I think the main thing to do is test this on an single close up hair strand, and verify that results are correct. It's really hard to tell on a complex hair setup what is right or wrong.

If it's not clear how to verify if it's correct, it's best to pick another Good First Issue. As I can't really explain what to look for without a background in BSDFs and sampling algorithms.

Helpful for testing this mostly:

Build with WITH_CYCLES_DEBUG=ON

In the preferences, enable "Developer Extras", then in "Experimental" enable Cycles Debug options

The in the render properties, there is a Debug panel with options to compare Direct Sampling, Next Event Estimation and MIS.

Hello , Sorry I didn't update this issue earlier .
I had IRL stuff to attend to these last 2 months , came back 3 days ago and I had to finish another issue as I was working on it before this one.
I'm currently fully on this now.
Like you said it's a bit hard on a complex hair model , and the pull request's examples may not be reflective of real changes because I forgot about image compression when comparing the images.
For now , it's really hard to see any meaningful differences between the bugged version and the new one . I think the math is correct in the last commit if I'm not mistaken, but like you said , on complex hair models it's difficult to see differences .
The problem should be more noticeable on a cuticles with a 4° angle ... Right now I'm just making more tests with simpler hair fibers.

> @holly-l it's still relevant, though I'm not sure if @HamilcarR plans to continue work on this. > > What I think the main thing to do is test this on an single close up hair strand, and verify that results are correct. It's really hard to tell on a complex hair setup what is right or wrong. > > If it's not clear how to verify if it's correct, it's best to pick another Good First Issue. As I can't really explain what to look for without a background in BSDFs and sampling algorithms. > > Helpful for testing this mostly: > * Build with `WITH_CYCLES_DEBUG=ON` > * In the preferences, enable "Developer Extras", then in "Experimental" enable Cycles Debug options > * The in the render properties, there is a Debug panel with options to compare Direct Sampling, Next Event Estimation and MIS. > > Hello , Sorry I didn't update this issue earlier . I had IRL stuff to attend to these last 2 months , came back 3 days ago and I had to finish another issue as I was working on it before this one. I'm currently fully on this now. Like you said it's a bit hard on a complex hair model , and the pull request's examples may not be reflective of real changes because I forgot about image compression when comparing the images. For now , it's really hard to see any meaningful differences between the bugged version and the new one . I think the math is correct in the last commit if I'm not mistaken, but like you said , on complex hair models it's difficult to see differences . The problem should be more noticeable on a cuticles with a 4° angle ... Right now I'm just making more tests with simpler hair fibers.

Holly Liu commented

2024-02-13 00:17:10 +01:00

You can explain what to look for. I do have a background in computer graphics from my alma mater.

Amine, are you testing the single strand sheep hair that was mentioned or others as well? I think we should try not to overlap the work being done.

You can explain what to look for. I do have a background in computer graphics from my alma mater. Amine, are you testing the single strand sheep hair that was mentioned or others as well? I think we should try not to overlap the work being done.

Amine Bensalem commented

2024-02-13 03:14:39 +01:00

You can explain what to look for. I do have a background in computer graphics from my alma mater.

Amine, are you testing the single strand sheep hair that was mentioned or others as well? I think we should try not to overlap the work being done.

Yes , I'm working with a bunch of hair types , but you can check the one you want , no worries.

I think it all boils down to a difference in specular reflection in the hair. the longitudinal scattering gives you the specular reflection of light along the hair fiber.
since this is what being modified , we should at least expect a difference in the size of the highlight , at a high cuticle angle , between the old code , and the new one.
For now , my tests have been the same between the two at low angle, the only difference between the two images is just the sampling's noise.
Also , I will edit my pull request about this , but the images diffs I posted are not actual differences , I made the stupid mistake of saving the images in jpeg , and what you see is probably due to compression.
right now like I said , there's no real differences except white noise (at low angle)

> You can explain what to look for. I do have a background in computer graphics from my alma mater. > > Amine, are you testing the single strand sheep hair that was mentioned or others as well? I think we should try not to overlap the work being done. > Yes , I'm working with a bunch of hair types , but you can check the one you want , no worries. I think it all boils down to a difference in specular reflection in the hair. the longitudinal scattering gives you the specular reflection of light along the hair fiber. since this is what being modified , we should at least expect a difference in the size of the highlight , at a high cuticle angle , between the old code , and the new one. For now , my tests have been the same between the two at low angle, the only difference between the two images is just the sampling's noise. Also , I will edit my pull request about this , but the images diffs I posted are not actual differences , I made the stupid mistake of saving the images in jpeg , and what you see is probably due to compression. right now like I said , there's no real differences except white noise (at low angle)

Amine Bensalem commented

2024-02-14 06:42:27 +01:00

so , I'm finally starting to get a bit of results from the new version :

old render	new render	specular	difference

the old code seems to exhibit areas of hair that have a specular effect that is "darker" in the direction of light

so , I'm finally starting to get a bit of results from the new version : |old render | new render| specular| difference | |-----------------|-------------------|----------------|--------------| |![chiang_old](/attachments/688d0d9b-1dad-49b3-939a-abb0a1b64383)| ![chiang_new](/attachments/61c8d3bd-5d55-4bdd-a24a-9efd7bebe851)| ![image](/attachments/a000cfe5-626e-4731-b648-6e6de8a4f5f0) | ![image](/attachments/70aedbb8-2508-4762-bb44-ae27ab0d48f2) | the old code seems to exhibit areas of hair that have a specular effect that is "darker" in the direction of light

diff_300s.bmp

2.7 MiB

chiang_new_300s.bmp

2.7 MiB

chiang_old_300s.bmp

2.7 MiB

hair_nodes-female_hair_styles.blend

51 MiB

image.png

1.2 MiB

image.png

1.2 MiB

blond_curly_chiang_300s.mkv

17 MiB

image.png

1.2 MiB

image.png

2.3 MiB

Brecht Van Lommel commented

2024-02-14 21:00:45 +01:00

It looks visually pleasing.Though a single hair render with simple light setup would still help understand if it's correct.

Amine Bensalem commented

2024-02-15 02:18:49 +01:00

little update : Those are single fibers samples , there's definitely a better specular highlight

old	new	diff

little update : Those are single fibers samples , there's definitely a better specular highlight |old|new|diff| |----|-------|-------| |![chiang_old_300s.bmp](/attachments/616bcd44-02d4-4103-9d9f-3f514a619787)|![chiang_new_300s.bmp](/attachments/5065ea7a-007b-41b0-af8d-82dbc6ea3518)|![chiang_diff.bmp](/attachments/e2916a04-9660-4706-bde9-3a5555337237)|

hair_nodes-female_hair_styles.blend

51 MiB

output.mkv

34 MiB

chiang_new_300s.bmp

42 MiB

chiang_diff.bmp

42 MiB

chiang_old_300s.bmp

42 MiB

Sign in to join this conversation.