Fix: difference of transmissive closures in EEVEE blended and dithered mode #125233

Weizhen Huang · 2024-07-22T16:20:35+02:00

Weizhen Huang commented

2024-07-22 16:20:35 +02:00

This fixes two bugs:

for transmissive closures, blended mode square the color, and
dithered square the color, weight and the pdf. Both were wrong.
The correct way is to square the color and weight, but divide the
pdf (closure_color_weight / total_weight) only once.
3fcd80db68 normalized the color after picking the closure, which
doesn't work for transmissive closures, where we apply the color and
weight twice. Therefore, delay the division until when evaluating the
picked closure. This fixes #125219.

This fixes two bugs: 1. for transmissive closures, blended mode square the color, and dithered square the color, weight and the pdf. Both were wrong. The correct way is to square the color and weight, but divide the pdf (`closure_color_weight / total_weight`) only once. 2. 3fcd80db68 normalized the color after picking the closure, which doesn't work for transmissive closures, where we apply the color and weight twice. Therefore, delay the division until when evaluating the picked closure. This fixes #125219.

Weizhen Huang added 1 commit 2024-07-22 16:20:43 +02:00

Fix #125219 : Translucent BSDF in Blended mode has wrong weight 8116232630

The color was normalized after picking the closure, which doesn't work
for translucent and refraction closures, where the color is applied
twice.
Therefore, delay the division until when evaluating the picked closure.

Weizhen Huang added this to the Viewport & EEVEE project 2024-07-22 16:20:47 +02:00

Weizhen Huang commented

2024-07-22 16:22:26 +02:00

@fclem can you check the compilation on Linux, or is blender-bot build supposed to capture compilation failure?

Weizhen Huang added 2 commits 2024-07-22 20:08:13 +02:00

Cleanup: reduce copy paste d1efc5f927

Fix color in deferred f575f996c4

Weizhen Huang changed title from ~~Fix #125219: Translucent BSDF in Blended mode has wrong weight~~ to WIP: Fix #125219: Translucent BSDF in Blended mode has wrong weight

2024-07-22 20:08:16 +02:00

Weizhen Huang commented

2024-07-22 20:26:18 +02:00

Turns out that blended and dithered already looked different before my change, but it was only obvious at grazing angles or when the closure weight is small:

4.2	Dithered	Blended	Cycles
Glass with 0.1 weight

The choice to square color and weight (which includes the closure weight and Fresnel), because this is what Cycles does (albeit a better Fresnel).

Turns out that blended and dithered already looked different before my change, but it was only obvious at grazing angles or when the closure weight is small: | 4.2| Dithered | Blended | Cycles | -------- | -------- | -------- |--| | Glass with 0.1 weight | ![dithered.png](/attachments/ebc16959-1c2a-4ec1-8a3a-f6e990bbdc16)|![blended.png](/attachments/6c9e08d9-5bce-4e74-be2c-36baee9b17c1)|![cycles.png](/attachments/1042acec-ed9b-4418-85be-cef5122bcc79)| The choice to square color and weight (which includes the closure weight and Fresnel), because this is what Cycles does (albeit a better Fresnel).

dithered.png

37 KiB

glass.blend

1.0 MiB

cycles.png

50 KiB

blended.png

34 KiB

Weizhen Huang requested review from Clément Foucault 2024-07-22 20:26:32 +02:00

Weizhen Huang added 1 commit 2024-07-22 21:12:33 +02:00

Pre-multiply color by weight 583e4559b9

Weizhen Huang changed title from ~~WIP: Fix #125219: Translucent BSDF in Blended mode has wrong weight~~ to Fix: difference of transmissive closures in EEVEE blended and dithered mode

2024-07-22 21:21:03 +02:00

Weizhen Huang commented

2024-07-22 22:16:47 +02:00

EEVEE should either have the option to test everything with both dithered and blended mode, or have some dedicated test as the attached one:

EEVEE 4.2	EEVEE PR	Cycles

EEVEE should either have the option to test everything with both dithered and blended mode, or have some dedicated test as the attached one: | EEVEE 4.2 | EEVEE PR | Cycles| | -------- | -------- | -------- | | ![Screenshot 2024-07-22 at 22.15.35.png](/attachments/c56e1ee6-8369-426f-8eff-019786e13ec1)|![Screenshot 2024-07-22 at 22.15.45.png](/attachments/669b6b51-4757-49d8-b216-5ccf7812f997)|![Screenshot 2024-07-22 at 22.15.55.png](/attachments/fde0767d-e021-4a62-99b2-d61fe2f5976e)|

Screenshot 2024-07-22 at 22.15.45.png

995 KiB

Screenshot 2024-07-22 at 22.15.35.png

1.1 MiB

transmissive_mix_blended_dithered.blend

746 KiB

Screenshot 2024-07-22 at 22.15.55.png

1.0 MiB

Weizhen Huang referenced this pull request

2024-07-23 12:21:40 +02:00

Fix #125219: Transmissive BSDF different in Blended and Dithered mode #125299

Weizhen Huang commented

2024-07-23 15:20:04 +02:00

Ok so this fix is also biased, because we are computing the sum of squares, which is not the same as the square of sums

EEVEE PR	Cycles	EEVEE 4.2	EEVEE main

Seems that PR still performs better in most cases. The most obvious difference is when mixing glass with glass (lower left)

Ok so this fix is also biased, because we are computing the sum of squares, which is not the same as the square of sums |EEVEE PR| Cycles | EEVEE 4.2 |EEVEE main| | -------- | -------- | -------- |--| | ![pr.png](/attachments/767a70dd-e9d7-4093-b8be-581f1db2809e) | ![cycles.png](/attachments/66ea962b-866d-497d-9666-800a8e058161) | ![4.2.png](/attachments/116cd78c-8bb9-486e-9996-afea6e0df387) |![main.png](/attachments/b44e7f0e-0a25-4102-b411-e8fab4b58b07)| Seems that PR still performs better in most cases. The most obvious difference is when mixing glass with glass (lower left)

mix_shader.blend

989 KiB

pr.png

110 KiB

4.2.png

113 KiB

cycles.png

112 KiB

main.png

113 KiB

❤️ 1

Weizhen Huang referenced this issue from a commit

2024-07-23 19:40:35 +02:00

Fix #125219: Transmissive BSDF different in Blended and Dithered mode

Weizhen Huang commented

2024-07-24 12:27:44 +02:00

I think the proper way would be to pick the closure twice using different random numbers, and multiply the resolved result. This way we are estimating the square of the sums.
Potential problem is that we might need an additional bin for that. I will try this in another PR.

I think the proper way would be to pick the closure twice using different random numbers, and multiply the resolved result. This way we are estimating the square of the sums. Potential problem is that we might need an additional bin for that. I will try this in another PR.

Weizhen Huang closed this pull request

2024-07-24 12:27:47 +02:00

Pull request closed

Please reopen this pull request to perform a merge.

Sign in to join this conversation.

No reviewers

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Fix: difference of transmissive closures in EEVEE blended and dithered mode #125233

Pull request closed