Cycles Image reconstruction Filter, correct sample addition, correct gamma display #44391

New Issue

Mohamed Sakr · 2015-04-14T19:03:33+02:00

Mohamed Sakr commented

2015-04-14 19:03:33 +02:00

1- sample addition feels not correct, I'm used to PBRT2 where samples are added like this:

  float xyz[3];
  L.ToXYZ(xyz);
  pixel.Lxyz += xyz;

but in Cycles it addes RGB to RGB, which seems not correct. "need confirmation"

2- Cycles doesn't have any image reconstruction "I was shocked!!" , when testing a simple rendering with Gaussian Filter of size 2. and size 5., it just altered how samples are created, not final image quality.

here is a good source for an image reconstruction function!
https://github.com/mmp/pbrt-v2/blob/master/src/film/image.cpp#L77-L137

3- gamma display is incorrect, test this in Cycles stand alone, --output image.jpg (will result in false gamma, and it is exactly as the displayed image in OpenGL buffer in Blender or in Cycles) , and --output image.exr (will result in correct gamma)

1- sample addition feels not correct, I'm used to PBRT2 where samples are added like this: ``` float xyz[3]; L.ToXYZ(xyz); pixel.Lxyz += xyz; ``` but in Cycles it addes RGB to RGB, which seems not correct. "need confirmation" 2- Cycles doesn't have any image reconstruction "I was shocked!!" , when testing a simple rendering with Gaussian Filter of size 2. and size 5., it just altered how samples are created, not final image quality. here is a good source for an image reconstruction function! https://github.com/mmp/pbrt-v2/blob/master/src/film/image.cpp#L77-L137 3- gamma display is incorrect, test this in Cycles stand alone, --output image.jpg (will result in false gamma, and it is exactly as the displayed image in OpenGL buffer in Blender or in Cycles) , and --output image.exr (will result in correct gamma)

Mohamed Sakr commented

2015-04-14 19:03:33 +02:00

Changed status to: 'Open'

Mohamed Sakr commented

2015-04-14 19:03:33 +02:00

Added subscriber: @MohamedSakr-2

Aaron Carlisle commented

2015-04-14 21:48:48 +02:00

Added subscriber: @Blendify

Julian Eisel commented

2015-04-14 21:56:59 +02:00

Added subscriber: @JulianEisel

Julian Eisel commented

2015-04-14 21:56:59 +02:00

Ehh, sounds to me this is a set of todos, but no real bugs? Changing to todo for now

Mohamed Sakr commented

2015-04-15 02:13:37 +02:00

well I consider (1) as a bug because it adds wrong values "if I'm correct" , they should be added in XYZ color space (not RGB)
about (2) it is misleading for the user, user expects an antialiasing filter, here it doesn't do any reconstruction
about (3) again it is for the "correct" color viewing during render, so user doesn't add light to dark areas "while they are not black if they got gamma correction"

well I consider (1) as a bug because it adds wrong values "if I'm correct" , they should be added in XYZ color space (not RGB) about (2) it is misleading for the user, user expects an antialiasing filter, here it doesn't do any reconstruction about (3) again it is for the "correct" color viewing during render, so user doesn't add light to dark areas "while they are not black if they got gamma correction"

Sergey Sharybin commented

2015-04-15 18:36:41 +02:00

Changed status from 'Open' to: 'Archived'

Sergey Sharybin closed this issue

2015-04-15 18:36:41 +02:00

Sergey Sharybin self-assigned this 2015-04-15 18:36:41 +02:00

Sergey Sharybin commented

2015-04-15 18:36:41 +02:00

Will reply on each of the points:

PBRT is working in XYZ space internally because of various reasons (well, it actually does spectrum which is rather straightforward to convert to XYZ) and for the user usable result it's then gets converted to linear RGB space. Cycles on the other hand is biased and uses linear RGB space internally so there's no need to do any sort of conversion when adding samples/passes.
AA filter and image reconstruction are quite separate things, they both happening in integrator but they're happy to exists separately. Now, image reconstruction is rather a hack which allows to get rid of fireflies by canceling some samples. This is a cheat which is fine for static images but which could cause some artifacts in animations. Also, PBRT is just a renderer and in Blender we've got compositor, which means it makes sense to do it as compo node (which for good results might require deep compositing, but it's something we need to support eventually anyway).
It seems the issue is in standalone only? Than it's better to mail me or to bf-cycles ML exact steps of reproducing the issue. This tracker is only about bugs reproduceable in Blender itself.

So thanks for the report, but such reports does not really belong to here.

Will reply on each of the points: 1. PBRT is working in XYZ space internally because of various reasons (well, it actually does spectrum which is rather straightforward to convert to XYZ) and for the user usable result it's then gets converted to linear RGB space. Cycles on the other hand is biased and uses linear RGB space internally so there's no need to do any sort of conversion when adding samples/passes. 2. AA filter and image reconstruction are quite separate things, they both happening in integrator but they're happy to exists separately. Now, image reconstruction is rather a hack which allows to get rid of fireflies by canceling some samples. This is a cheat which is fine for static images but which could cause some artifacts in animations. Also, PBRT is just a renderer and in Blender we've got compositor, which means it makes sense to do it as compo node (which for good results might require deep compositing, but it's something we need to support eventually anyway). 3. It seems the issue is in standalone only? Than it's better to mail me or to bf-cycles ML exact steps of reproducing the issue. This tracker is only about bugs reproduceable in Blender itself. So thanks for the report, but such reports does not really belong to here.

Mohamed Sakr commented

2015-04-15 19:22:57 +02:00

Hey Sergey,

1- you are correct, that's why I needed a confirmation :) as I wasn't sure, but I guess implementing a Spectrum class and giving the user an option to choose between Spectrum and RGB is very good "this is needed before extending Cycles with photons, Caustics, etc.."

2- I disagree at this point, imagine an image which needs 1000 spp to converge in animation, it may need only 500 spp to converge in animation with image reconstruction.

3- the issue is in both standalone and in Blender, I confirmed it with a friend "and by tests" , saving image to OpenEXR saves it with correct gamma "2.2" or to be exact "as written in Cycles"

1.055f * powf(c, 1.0f / 2.4f) - 0.055f;

so if .exr -> bright image with correct feedback (only in save, not in preview buffer)
else -> dark image (similar to preview).

Hey Sergey, 1- you are correct, that's why I needed a confirmation :) as I wasn't sure, but I guess implementing a Spectrum class and giving the user an option to choose between Spectrum and RGB is very good "this is needed before extending Cycles with photons, Caustics, etc.." 2- I disagree at this point, imagine an image which needs 1000 spp to converge in animation, it may need only 500 spp to converge in animation with image reconstruction. 3- the issue is in both standalone and in Blender, I confirmed it with a friend "and by tests" , saving image to OpenEXR saves it with correct gamma "2.2" or to be exact "as written in Cycles" 1.055f * powf(c, 1.0f / 2.4f) - 0.055f; so if .exr -> bright image with correct feedback (only in save, not in preview buffer) else -> dark image (similar to preview).

Sergey Sharybin commented

2015-04-20 14:56:39 +02:00

I didn't tell that image reconstruction does not help, i was only telling that in blender it's not necesserily belong to render engine. Plus, it's not a magic bullet either, it wouldn't help you much with effects like DOF and motion blur.

Gamma 2.2 is not correct for exr files, they're expected to be in linear scene space. What are the steps of reproducing the exr file which is not in linear space from blender?

I didn't tell that image reconstruction does not help, i was only telling that in blender it's not necesserily belong to render engine. Plus, it's not a magic bullet either, it wouldn't help you much with effects like DOF and motion blur. Gamma 2.2 is not correct for exr files, they're expected to be in linear scene space. What are the steps of reproducing the exr file which is not in linear space from blender?

Mohamed Sakr commented

2015-04-20 17:04:50 +02:00

I can't reproduce it from blender, the saved image is exactly = to the previewed image "in any format" , both appear with linear gamma which gives the user wrong impression "at least when he tries to render in progressive mode to test lights and shaders"

in standalone, OpenEXR gamma get changed "corrected", but .jpg for example gives the same impression as preview.

I can't reproduce it from blender, the saved image is exactly = to the previewed image "in any format" , both appear with linear gamma which gives the user wrong impression "at least when he tries to render in progressive mode to test lights and shaders" in standalone, OpenEXR gamma get changed "corrected", but .jpg for example gives the same impression as preview.

Sergey Sharybin commented

2015-04-28 00:45:23 +02:00

First you say that it happens in blender then you say you can't reproduce it in blender. That's confusing.

So if there's something wrong in blender itself, which exact steps you're doing, what you're expecting and what results you're having?

First you say that it happens in blender then you say you can't reproduce it in blender. That's confusing. So if there's something wrong in blender itself, which exact steps you're doing, what you're expecting and what results you're having?

Mohamed Sakr commented

2015-04-30 21:27:21 +02:00

sorry for this, my first test was inside blender "no images saved" , and images were the same, thought it would make the same problem after saving "like in Cycles Standalone" , but the save mechanism is different in Blender and Cycles Standalone.

so the problem is ONLY in the Standalone.

sorry for this, my first test was inside blender "no images saved" , and images were the same, thought it would make the same problem after saving "like in Cycles Standalone" , but the save mechanism is different in Blender and Cycles Standalone. so the problem is ONLY in the Standalone.

Sign in to join this conversation.

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Cycles Image reconstruction Filter, correct sample addition, correct gamma display #44391