Noise seed for stereoscopic rendering #50024

New Issue

Gatis Kurzemnieks · 2016-11-14T16:36:49+01:00

Gatis Kurzemnieks commented

2016-11-14 16:36:49 +01:00

Not a bug, but design issue.
When rendering stereoscopic images, the same seed is used for both left and right eye renders. If there is a noise in frame it has the same pattern in both eyes. Your brain picks this up and because noise patterns are identical (there is no stereo offset) it visually looks like there is a noise sheet floating in the air. It's obvious only when viewing your rendered image in VR headset.

Currently to avoid this you need to have insanely high sample settings to avoid having noise at all.

It would be very welcome to enable randomizing seed for each eye, not only per-frame - I believe it would solve the issue.

Another great solution would be to add pre_render and post_render python callbacks that work per-eye. Currently they are called only once for both eyes.

Not a bug, but design issue. When rendering stereoscopic images, the same seed is used for both left and right eye renders. If there is a noise in frame it has the same pattern in both eyes. Your brain picks this up and because noise patterns are identical (there is no stereo offset) it visually looks like there is a noise sheet floating in the air. It's obvious only when viewing your rendered image in VR headset. Currently to avoid this you need to have insanely high sample settings to avoid having noise at all. It would be very welcome to enable randomizing seed for each eye, not only per-frame - I believe it would solve the issue. Another great solution would be to add pre_render and post_render python callbacks that work per-eye. Currently they are called only once for both eyes.

Gatis Kurzemnieks commented

2016-11-14 16:36:49 +01:00

Changed status to: 'Open'

Gatis Kurzemnieks commented

2016-11-14 16:36:49 +01:00

Added subscriber: @GatisKurzemnieks

Dalai Felinto commented

2016-11-16 17:10:58 +01:00

Added subscribers: @Sergey, @sebastian_k, @dfelinto

Dalai Felinto commented

2016-11-16 17:10:58 +01:00

@sebastian_k how are you handling this at the moment? @Sergey was this even a concern for the Caminandes VR project?

Aaron Carlisle commented

2016-11-16 17:13:08 +01:00

Added subscriber: @Blendify

Dalai Felinto was assigned by Sergey Sharybin

2016-11-18 16:04:52 +01:00

Sergey Sharybin commented

2016-11-18 16:04:52 +01:00

@dfelinto, we didn't really consider this as a big concern AFAIR. Something simple like this will fix it tho:

P428: Patch for #50024

diff --git a/intern/cycles/blender/blender_session.cpp b/intern/cycles/blender/blender_session.cpp
index c250a54..ed03ecf 100644
--- a/intern/cycles/blender/blender_session.cpp
+++ b/intern/cycles/blender/blender_session.cpp
@@ -498,7 +498,8 @@ void BlenderSession::render()
 		scene->film->tag_update(scene);
 		scene->integrator->tag_update(scene);
 
-		for(b_rr.views.begin(b_view_iter); b_view_iter != b_rr.views.end(); ++b_view_iter) {
+		int view_index = 0;
+		for(b_rr.views.begin(b_view_iter); b_view_iter != b_rr.views.end(); ++b_view_iter, ++view_index) {
 			b_rview_name = b_view_iter->name();
 
 			/* set the current view */
@@ -514,6 +515,12 @@ void BlenderSession::render()
 			                &python_thread_state,
 			                b_rlay_name.c_str());
 
+			/* Make sure all views have different noise patterns. */
+			if(view_index != 0) {
+				scene->integrator->seed += view_index;
+				scene->integrator->tag_update(scene);
+			}
+
 			/* Update number of samples per layer. */
 			int samples = sync->get_layer_samples();
 			bool bound_samples = sync->get_layer_bound_samples();

Assigning to you, feel free to commit the change :)

@dfelinto, we didn't really consider this as a big concern AFAIR. Something simple like this will fix it tho: [P428: Patch for #50024](https://archive.blender.org/developer/P428.txt) ``` diff --git a/intern/cycles/blender/blender_session.cpp b/intern/cycles/blender/blender_session.cpp index c250a54..ed03ecf 100644 --- a/intern/cycles/blender/blender_session.cpp +++ b/intern/cycles/blender/blender_session.cpp @@ -498,7 +498,8 @@ void BlenderSession::render() scene->film->tag_update(scene); scene->integrator->tag_update(scene); - for(b_rr.views.begin(b_view_iter); b_view_iter != b_rr.views.end(); ++b_view_iter) { + int view_index = 0; + for(b_rr.views.begin(b_view_iter); b_view_iter != b_rr.views.end(); ++b_view_iter, ++view_index) { b_rview_name = b_view_iter->name(); /* set the current view */ @@ -514,6 +515,12 @@ void BlenderSession::render() &python_thread_state, b_rlay_name.c_str()); + /* Make sure all views have different noise patterns. */ + if(view_index != 0) { + scene->integrator->seed += view_index; + scene->integrator->tag_update(scene); + } + /* Update number of samples per layer. */ int samples = sync->get_layer_samples(); bool bound_samples = sync->get_layer_bound_samples(); ``` Assigning to you, feel free to commit the change :)

Gatis Kurzemnieks commented

2016-11-18 16:12:21 +01:00

Awesome! Thank you Sergey!

Sebastian Koenig commented

2016-11-18 16:15:24 +01:00

In #50024#404772, @dfelinto wrote:
@sebastian_k how are you handling this at the moment?

Yeah, we did notice this noise veil kind of effect, but then again noise always looks crappy, so we just try to render as noisefree as possible. :)
And when doing so the "veil" is not noticable anymore. But different noise patterns per eye could be interesting. If I understand the code above correctly it seems like a great solution, since it is not totally random, but re-rendering parts of the image would still give the same result per eye so you can overlay with identical noise pattern, which is something we do quite often to fix stuff in final renders.

> In #50024#404772, @dfelinto wrote: > @sebastian_k how are you handling this at the moment? Yeah, we did notice this noise veil kind of effect, but then again noise always looks crappy, so we just try to render as noisefree as possible. :) And when doing so the "veil" is not noticable anymore. But different noise patterns per eye could be interesting. If I understand the code above correctly it seems like a great solution, since it is not totally random, but re-rendering parts of the image would still give the same result per eye so you can overlay with identical noise pattern, which is something we do quite often to fix stuff in final renders.

Dalai Felinto commented

2016-11-18 18:27:06 +01:00

Changed status from 'Open' to: 'Resolved'

Dalai Felinto closed this issue

2016-11-18 18:27:06 +01:00

Dalai Felinto commented

2016-11-18 18:27:06 +01:00

Fixed on 0f8e5f4fb4 Thanks for the report. And @sebastian_k great to count on you, thanks.

Fixed on 0f8e5f4fb4 Thanks for the report. And @sebastian_k great to count on you, thanks.

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

Noise seed for stereoscopic rendering #50024