Cycles bump map is not super sampled but filtered #43495

New Issue

Jason Ellis · 2015-01-31T03:26:58+01:00

Jason Ellis commented

2015-01-31 03:26:58 +01:00

System Information
OS X Yosemite

Blender Version
Latest

Short description of error

Exact steps for others to reproduce the error
Render high res bump tex close-up and then downsample.
Render high red bump from a distance.
Second result is blurry and misses features.

My guess is that it's per pixel pre-filtered, however, it would be better to sample the bump map at full res.

**System Information** OS X Yosemite **Blender Version** Latest **Short description of error** **Exact steps for others to reproduce the error** Render high res bump tex close-up and then downsample. Render high red bump from a distance. Second result is blurry and misses features. My guess is that it's per pixel pre-filtered, however, it would be better to sample the bump map at full res.

Jason Ellis commented

2015-01-31 03:26:58 +01:00

Changed status to: 'Open'

Jason Ellis commented

2015-01-31 03:26:58 +01:00

Added subscriber: @jasonellis

Aaron Carlisle commented

2015-01-31 03:43:03 +01:00

Added subscriber: @Blendify

Aaron Carlisle commented

2015-01-31 03:43:03 +01:00

Please post a .blend and some pictures and a youtube video will help

Jason Ellis commented

2015-01-31 23:10:15 +01:00

Here is an example.

BumpSamplingBug.zip

Here is an example. [BumpSamplingBug.zip](https://archive.blender.org/developer/F140178/BumpSamplingBug.zip)

Jason Ellis commented

2015-02-01 00:22:17 +01:00

A workaround is, of course, to render at, say, 10x the target size, reduce the sample count by 10x and then downsample the result by 10x. (This is also what I have done for comparison in the example I have provided.)

Sergey Sharybin commented

2015-02-10 09:15:51 +01:00

Changed status from 'Open' to: 'Archived'

Sergey Sharybin closed this issue

2015-02-10 09:15:51 +01:00

Sergey Sharybin self-assigned this 2015-02-10 09:15:51 +01:00

Sergey Sharybin commented

2015-02-10 09:15:51 +01:00

This is because mipmaps are not implemented in Cycles. Well, they are implemented but only when using OpenShadingLanguage backend. For until mipmaps are implemented for SVM such cases would always be noisy and wrong.

Thanks for the report, but it is a known thing TODO.

This is because mipmaps are not implemented in Cycles. Well, they are implemented but only when using OpenShadingLanguage backend. For until mipmaps are implemented for SVM such cases would always be noisy and wrong. Thanks for the report, but it is a known thing TODO.

Jason Ellis commented

2015-02-10 13:54:27 +01:00

Just out of curiosity, couldn't you continue to use the OSL backend but make mipmap texture lookups at level N-log_2(sample_count) instead of at N? That should work fine if I'm not mistaken. (Ideally, you should create mipmaps of bump and normal maps with a Gaussian lobe model since a simple average is not a good statistical model for directions - [x]. Otherwise, it might be worthwhile to add a (default?) option to completely deactivate mipmapping for bump and normal maps i.e. to take samples at the full resolution.).

By the way, this problem is not only restricted to bump maps since mipmapping with a sample sizes of the entire fragments (this is what I assume is happening here) generally has a lower image quality compared to supersampling with mipmapped subsamples.

Could you point me to the source code in which the texture lookup is done? Perhaps, I can provide a patch.

http://www.cs.columbia.edu/cg/normalmap/normalmap.pdf

Just out of curiosity, couldn't you continue to use the OSL backend but make mipmap texture lookups at level N-log_2(sample_count) instead of at N? That should work fine if I'm not mistaken. (Ideally, you should create mipmaps of bump and normal maps with a Gaussian lobe model since a simple average is not a good statistical model for directions - [x]. Otherwise, it might be worthwhile to add a (default?) option to completely deactivate mipmapping for bump and normal maps i.e. to take samples at the full resolution.). By the way, this problem is not only restricted to bump maps since mipmapping with a sample sizes of the entire fragments (this is what I assume is happening here) generally has a lower image quality compared to supersampling with mipmapped subsamples. Could you point me to the source code in which the texture lookup is done? Perhaps, I can provide a patch. - [x] http://www.cs.columbia.edu/cg/normalmap/normalmap.pdf

Sergey Sharybin commented

2015-02-10 14:53:27 +01:00

If you want to look into sampling, then for OSL the closest entry point is https:*developer.blender.org/diffusion/B/browse/master/intern/cycles/kernel/osl/osl_services.cpp;298d8681a0ac28ccc96e593a02ad34a5cef624c6$842 and for SVM think this is a good start https:*developer.blender.org/diffusion/B/browse/master/intern/cycles/kernel/svm/svm_image.h;298d8681a0ac28ccc96e593a02ad34a5cef624c6$53

Looking into mipmaps support for SVM is also welcome, but that's kinda can of worms because of limited VRAM on GPUs.. But still be welcome to look into :)

If you want to look into sampling, then for OSL the closest entry point is https:*developer.blender.org/diffusion/B/browse/master/intern/cycles/kernel/osl/osl_services.cpp;298d8681a0ac28ccc96e593a02ad34a5cef624c6$842 and for SVM think this is a good start https:*developer.blender.org/diffusion/B/browse/master/intern/cycles/kernel/svm/svm_image.h;298d8681a0ac28ccc96e593a02ad34a5cef624c6$53 Looking into mipmaps support for SVM is also welcome, but that's kinda can of worms because of limited VRAM on GPUs.. But still be welcome to look into :)

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 bump map is not super sampled but filtered #43495