Maniphest T96718

Regression: Photo "blowout" when rendering and denoising in 3.1.0
Closed, ResolvedBUG
Actions

Assigned To
Lukas Stockner (lukasstockner97)
Authored By
Fara Shimbo (FaraS)
Mar 22 2022, 8:45 PM
Tags
  • BF Blender
  • Render & Cycles
Subscribers
Alaska (Alaska)
Brecht Van Lommel (brecht)
Connor M. Denning (ConMan)
Fara Shimbo (FaraS)
Fernando Alcala (TriFractal)
Lukas Stockner (lukasstockner97)

Description

System Information
Operating system: Linux Mint 20.2
Graphics card: Nvidia 3070ti

Blender Version
Broken: 3.1.0 2022-03-08 18:16 c77597cd0e15 master
Worked: 3.0.0 2021-12-02 18:35 f1cca3055776 master

Caused by rB0781c22ceedc: Fix T95969, T91856: bake AO to vertex color artifacts after ray offset removal

When rendering the attached file in CPU compute, a number of fireflies appear, and upon denoising, these "blow out." I am using CPU compute because this file runs my graphics card out of memory. This file is rendered in Cycles.

DiscussingWithRuba.blend.tar.gz56 MBDownload

Exact steps for others to reproduce the error
Render the image in the attached file. It blows out reliably every time. Neither the fireflies nor the blowouts appear in 3.0.

I am including an image rendered in 3.0 (still runs the GPU out of memory), one "blow out" rendered in 3.1, and a tar.gz of the file.

Thanks!

Revisions and Commits

rB Blender
rB54f447ecde72 Fix T96718: Cycles invalid pixels when using bump normal for light emission

Related Objects

Mentioned In
P2839 3.1 R
Mentioned Here
rB0781c22ceedc: Fix T95969, T91856: bake AO to vertex color artifacts after ray offset removal
rBf1cca3055776: Blender 3.0 - version bump -> release
rBc77597cd0e15: Fix T95843: Invalid mesh runtime data after sculpt undo

Event Timeline

Fara Shimbo (FaraS) created this task.Mar 22 2022, 8:45 PM
Alaska (Alaska) changed the task status from Needs Triage to Confirmed.Mar 22 2022, 11:52 PM
Alaska (Alaska) triaged this task as High priority.
Alaska (Alaska) changed the subtype of this task from "Report" to "Bug".
Alaska (Alaska) added a project: Render & Cycles.
Alaska (Alaska) added subscribers: Brecht Van Lommel (brecht), Alaska (Alaska).

I can confirm this issue with the setup listed below.

The issue isn't "fireflies" in the typical sense. It is some rendering error which is resulting in extreme values. This is definitely an error.
Is there a specific object or something in the scene that's causing this? Maybe, but I haven't identified it yet.

Bisecting points to rB0781c22ceedc: Fix T95969, T91856: bake AO to vertex color artifacts after ray offset removal as the issue. CC @Brecht Van Lommel (brecht) can you take a look at it?

System Information
Operating system: Linux-5.16.0-4-amd64-x86_64-with-glibc2.33 64 Bits
Graphics card: NVIDIA GeForce RTX 3060/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 510.54
Blender version: 3.2.0 Alpha, branch: master, commit date: 2022-03-22 07:07, hash: rB47451af027e2

Alaska (Alaska) updated the task description.Mar 22 2022, 11:52 PM
Fara Shimbo (FaraS) added a comment.Mar 23 2022, 4:00 AM

I should add that the more light in the scene, the worse the problem (this render was done with the lights at about twice the power they were at in the file. Don't know if this will help track the bug, but here it is for what it's worth.

Connor M. Denning (ConMan) added a subscriber: Connor M. Denning (ConMan).Mar 23 2022, 6:49 AM
Brecht Van Lommel (brecht) closed this task as Resolved.Mar 23 2022, 2:16 PM
Brecht Van Lommel (brecht) claimed this task.

I can reproduce this in 3.1.0 but not 3.2.0 alpha, it appears to be solved there.
https://builder.blender.org/download/daily/

I'm also not sure rB0781c22ceedc6700a is related to this, I see no ambient occlusion in this file, but doesn't really matter either way.

Fara Shimbo (FaraS) added a comment.Mar 24 2022, 12:17 AM

I just downloaded 3.2 alpha, and while the resulting problem is slightly different, it's still there.

Alaska (Alaska) reopened this task as Confirmed.Mar 24 2022, 1:07 AM

I can continue to reproduce the issue in builds from the build bot and my own build of latest master. I will re-open the report in the mean time.

If knowing what CPU is used during rendering helps in some way, I am using the Ryzen 9 5950X.

My own build of master: 3.2.0 Alpha, branch: master, commit date: 2022-03-23 22:25, hash: rB51380b9346a5
Build bot: 3.2.0 Alpha, branch: master, commit date: 2022-03-22 22:06, hash: rB5c27391b0f8c

Alaska (Alaska) removed Brecht Van Lommel (brecht) as the assignee of this task.Mar 24 2022, 1:08 AM
Fernando Alcala (TriFractal) added a subscriber: Fernando Alcala (TriFractal).Mar 24 2022, 7:48 AM

I can reproduce on Ubuntu 21.10 with a Threadripper 3000.
Cannot reproduce on Windows 10 with an i7 4790k or Threadripper.

On Ubuntu, it isn't just the denoise that breaks, there are a bunch of additional fireflies that aren't on the Windows version. The fireflies are different every time too.
Windows i7/TR noisy


Windows i7/TR denoise

Ubuntu TR noisy render 1

Ubuntu TR noisy render 2

Ubuntu TR denoised render 1

Ubuntu TR denoised render 2

Fara Shimbo (FaraS) added a comment.Mar 24 2022, 2:00 PM

I am using an AMD Ryzen 5 3600X 6-Core Processor

Philipp Oeser (lichtwerk) renamed this task from Photo "blowout" when rendering and denoising in 3.1.0 to Regression: Photo "blowout" when rendering and denoising in 3.1.0.Mar 25 2022, 11:13 AM
Pratik Borhade (PratikPB2123) mentioned this in P2839 3.1 R.Mar 27 2022, 6:46 AM
Lukas Stockner (lukasstockner97) claimed this task.May 1 2022, 1:23 AM
Lukas Stockner (lukasstockner97) added a subscriber: Lukas Stockner (lukasstockner97).

Can reproduce, the problem appears to be caused by the shader on rock.005.

Something is going wrong with SVM there, it's reading an uninitialized location from the stack, which returns garbage, which in turn gets used as an emission value, resulting in the bright pixels. That also explains the differences between compilers/platforms.

Connor M. Denning (ConMan) added a comment.EditedMay 1 2022, 8:06 PM
In T96718#1349609, @Lukas Stockner (lukasstockner97) wrote:

Can reproduce, the problem appears to be caused by the shader on rock.005.

Something is going wrong with SVM there, it's reading an uninitialized location from the stack, which returns garbage, which in turn gets used as an emission value, resulting in the bright pixels. That also explains the differences between compilers/platforms.

Hello I'm having similar issue with GPU and CPU rendering outputting very different results on there tiles in hybrid rendering cuda and opencl. the density seems to be off between them.
this file should show it.

Broken volumes.blend606 KBDownload

Alaska (Alaska) added a comment.May 2 2022, 2:10 AM

@Connor M. Denning (ConMan) , reports about different bugs should be made in different bug reports. Please report your issue elsewhere by selecting from the top of Blender Help → Report a bug and filling out all the relevant information.

Connor M. Denning (ConMan) added a comment.May 3 2022, 4:57 PM
In T96718#1349867, @Alaska (Alaska) wrote:

@Connor M. Denning (ConMan) , reports about different bugs should be made in different bug reports. Please report your issue elsewhere by selecting from the top of Blender Help → Report a bug and filling out all the relevant information.

Are you sure my issue isn't the same issue cause I'm thinking the issue I'm having might be this issue. I'm happy to make a bug report if it's not.

Brecht Van Lommel (brecht) added a comment.May 3 2022, 5:07 PM

It's not the same issue, please make a separate bug report.

Brecht Van Lommel (brecht) closed this task as Resolved by committing rB54f447ecde72: Fix T96718: Cycles invalid pixels when using bump normal for light emission.May 4 2022, 8:05 PM
Brecht Van Lommel (brecht) added a commit: rB54f447ecde72: Fix T96718: Cycles invalid pixels when using bump normal for light emission.