Cycles- MIS gives better result with resolution superior to input HDRi's resolution #49952

New Issue

mathieu menuet · 2016-11-06T16:32:53+01:00

mathieu menuet commented

2016-11-06 16:32:53 +01:00

System Information
All

Blender Version
Broken: 2.78a

Short description of error
Using a resolution of MIS superior to the input image gives better results. The quality should improve up to image resolution and then stay constant.

Exact steps for others to reproduce the error
Open attached .blend. Included is an HDRi from Greg Zaal under CC-BY with a resolution of 512x256 : bug mis hdr.blend
Render animation:

frame 1 is at 512 for MIS (should be the best result already)
frame 2 is at 1024 = 4x more pixels, interpolated
Frame 3 at 2048 = 16x more pixels, interpolated

Frame 4 at 4096 = 64x interpolated pixels, but still it's even better than frame 3

**System Information** All **Blender Version** Broken: 2.78a **Short description of error** Using a resolution of MIS superior to the input image gives better results. The quality should improve up to image resolution and then stay constant. **Exact steps for others to reproduce the error** Open attached .blend. Included is an HDRi from Greg Zaal under CC-BY with a resolution of 512x256 : [bug mis hdr.blend](https://archive.blender.org/developer/F393517/bug_mis_hdr.blend) Render animation: - frame 1 is at 512 for MIS (should be the best result already) ![hdri0001.jpg](https://archive.blender.org/developer/F393506/hdri0001.jpg) - frame 2 is at 1024 = 4x more pixels, interpolated ![hdri0002.jpg](https://archive.blender.org/developer/F393511/hdri0002.jpg) - Frame 3 at 2048 = 16x more pixels, interpolated ![hdri0003.jpg](https://archive.blender.org/developer/F393510/hdri0003.jpg) - Frame 4 at 4096 = 64x interpolated pixels, but still it's even better than frame 3 ![hdri0004.jpg](https://archive.blender.org/developer/F393505/hdri0004.jpg)

mathieu menuet commented

2016-11-06 16:32:53 +01:00

Changed status to: 'Open'

mathieu menuet commented

2016-11-06 16:32:53 +01:00

Added subscriber: @bliblubli

mathieu menuet commented

2016-11-06 16:33:18 +01:00

Added subscribers: @brecht, @LukasStockner, @Sergey

Lukas Stockner commented

2016-11-06 16:50:37 +01:00

Ah, yes, I wanted to look into that.
But, I think you forgot to attach the .blend?

Ah, yes, I wanted to look into that. But, I think you forgot to attach the .blend?

mathieu menuet commented

2016-11-06 16:54:47 +01:00

True :D
bug mis hdr.blend

True :D [bug mis hdr.blend](https://archive.blender.org/developer/F393516/bug_mis_hdr.blend)

Lukas Stockner self-assigned this 2016-11-06 18:48:20 +01:00

Lukas Stockner commented

2016-11-06 18:48:20 +01:00

Yes, I can reproduce the issue, and it should not behave like that.
Interestingly, my first assumption was that it's due to the texture interpolation - but, when switching to closest, the difference is much higher, which makes me suspect an off-by-one bug somewhere in the sampling.

Yes, I can reproduce the issue, and it should not behave like that. Interestingly, my first assumption was that it's due to the texture interpolation - but, when switching to closest, the difference is much higher, which makes me suspect an off-by-one bug somewhere in the sampling.

blender-admin commented

2016-11-06 19:12:45 +01:00

This issue was referenced by blender/cycles@512a999023

This issue was referenced by blender/cycles@512a999023619926786eced4d1c0728456e41e57

blender-admin commented

2016-11-06 19:12:45 +01:00

This issue was referenced by f89fbf580e

This issue was referenced by f89fbf580eae6202cef9da08756fd415ca34a8f3

Lukas Stockner commented

2016-11-06 20:38:16 +01:00

Changed status from 'Open' to: 'Resolved'

Lukas Stockner closed this issue

2016-11-06 20:38:16 +01:00

Lukas Stockner commented

2016-11-06 20:42:32 +01:00

Okay, should work now.

To give an example how bad this bug could be:

Old code, 256spp, Closest Interpolation, Resolution 512:

New code, 8spp, Closest Interpolation, Resolution 512:

Regarding linear interpolation: That needs twice the image resolution right now to look good - it could be improved, but that would probably make Closest interpolation worse. Still, I guess that linear interpolation is the usual case, so I'll try hacking together a patch for piecewise-linear sampling to test the actual difference.

Okay, should work now. To give an example how bad this bug could be: Old code, 256spp, Closest Interpolation, Resolution 512: ![old_mis512_closest_256spp.png](https://archive.blender.org/developer/F393563/old_mis512_closest_256spp.png) New code, **8spp**, Closest Interpolation, Resolution 512: ![new_mis512_closest_8spp.png](https://archive.blender.org/developer/F393565/new_mis512_closest_8spp.png) Regarding linear interpolation: That needs twice the image resolution right now to look good - it could be improved, but that would probably make Closest interpolation worse. Still, I guess that linear interpolation is the usual case, so I'll try hacking together a patch for piecewise-linear sampling to test the actual difference.

Brecht Van Lommel commented

2016-11-06 20:48:55 +01:00

Are you saying linear interpolation is worse than before? That's definitely the usual case, closest is not important for HDRI backgrounds.

Lukas Stockner commented

2016-11-06 21:03:56 +01:00

No, that was maybe a bit unclear - I was saying that linear interpolation at 1:1 resolution is still terrible (maybe a tiny bit better, but hard to say), while 2:1 resolution or better now is clean as well, which it wasn't before.

mathieu menuet commented

2016-11-06 21:39:36 +01:00

Thanks for the quick fix Lukas. The image is much cleaner indeed :) And requiring twice the resolution only is much better.

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- MIS gives better result with resolution superior to input HDRi's resolution #49952