Page MenuHome

EEVEE 2.93 Incorrect glossy values
Confirmed, NormalPublic

Description

System Information
Operating system: win 10
Graphics card: rx 6800

Blender Version
Broken: 2.93
Worked: 2.92

Short description of error

Recent eevee ao refactor looks like it produces inferior results. geometry is clearly less defined and in new implementation you can even notice glossy shading underneath AO something that is not present in old AO

2.92

2.93

Event Timeline

Leonardo (leonardos) updated the task description. (Show Details)
Leonardo (leonardos) updated the task description. (Show Details)
Leonardo (leonardos) added a comment.EditedTue, Feb 23, 5:12 AM

Here is a more practical example it looks like there is an issue with more then just AO:

This is 100% exact same scene, only re baked with same settings and everything the same in 2.93 because irradiance from 2.92 doesnt work in 2.93

2.92

2.93

It appears to be a continuation of T85720 and it appears that rBcd1a08398437bb6be4ea82bb553ef3e8dcaee7ab commit didnt solve the issue completely therefore T85720 is not resolved properly either.

Leonardo (leonardos) added a comment.EditedTue, Feb 23, 5:45 AM

GlossDir 2.92


GlossDir 2.93

GlossCol 2.92

GlossCol 2.93

Leonardo (leonardos) renamed this task from EEVEE AO Refactor produces inferior results to EEVEE 2.93 Incorrect glossy values.Tue, Feb 23, 5:47 AM
Evan Wilson (EAW) changed the task status from Needs Triage to Confirmed.EditedTue, Feb 23, 9:54 AM

I am not certain what the result of the AO refactoring is intended to look like. I can confirm however that there is an issue with the glossy passes in EEVEE.

The first issue is that they don’t appear in the viewport when selected from the menu. The shadow pass doesn’t either.

The second issue is how the glossy passes appear in a render. Testing different sample counts, it became clear that the glossy color pass is being divided by the number of samples, while the glossy direct is getting multiplied. The only way to get the values to match the levels from 2.92 and earlier is to render with a single sample.

This is caused by 3 things:

  • The AO is more correct as it gives less dark values on some regions.
  • The number of samples was considerably reduced leading to less features for really high radii. This is planned to be fixed.
  • The thickness heuristic does not yet have a parameter to tweak it. This might reduce shadowing. This is planned to be fixed.

This is caused by 3 things:

  • The AO is more correct as it gives less dark values on some regions.
  • The number of samples was considerably reduced leading to less features for really high radii. This is planned to be fixed.
  • The thickness heuristic does not yet have a parameter to tweak it. This might reduce shadowing. This is planned to be fixed.

Ah, great. Is AO then going to be finished for 2.93? What about glossy pass issues? EEVEE is unusable at the moment :/ was really looking forward to test out new AO

@Leonardo (leonardos) Please report the Glossy color pass bug in another report. Do not discuss many issues on one ticket, it makes it hard to follow and manage.

Yes AO is going to be finished for 2.93.

@Clément Foucault (fclem) and @Leonardo (leonardos) I’ll create the glossy pass bug report, since I should have split it off before.

I created T86036 for the Viewport AOV issue, and T86037 for the Rendered AOV issue.