Blend modes are broken #75844

New Issue

Richard Antalik · 2020-04-17T20:00:33+02:00

Richard Antalik commented

2020-04-17 20:00:33 +02:00

System Information
Operating system: Windows-10-10.0.18362-SP0 64 Bits
Graphics card: Radeon RX550/550 Series ATI Technologies Inc. 4.5.13587 Core Profile Context 20.4.1 26.20.15029.20013

Blender Version
Broken: version: 2.90 (sub 0), branch: arcpatch-D7314 (modified), commit date: 2020-04-17 15:25, hash: 0e2de2b212
Worked: Never

Short description of error
Output of various blend modes is different depending on whether byte or float image is blended
Right now I am not sure which result is correct (or if any)

In example file Blend is set to Linear burn with brush strength of 0.5 (I assume this will give brush pixels alpha of 0.5)

I have found this bug while working on D7314, so I can work on fixing blend modes as well, or at least I would like to be involved in how this will be fixed. Of course, if this is actually bug and not only me being dumb.

Exact steps for others to reproduce the error
#75844.blend

Open file
Paint both textures with pre-set brush

Result is different

As a consequence, some blend modes also work unexpected in the Video Sequence Editor, see #76376, #74440, #79333 for details.
Also note that this was once changed in fdee84fd56, so we are now in a "fix-for-fix" scenario...

**System Information** Operating system: Windows-10-10.0.18362-SP0 64 Bits Graphics card: Radeon RX550/550 Series ATI Technologies Inc. 4.5.13587 Core Profile Context 20.4.1 26.20.15029.20013 **Blender Version** Broken: version: 2.90 (sub 0), branch: arcpatch-[D7314](https://archive.blender.org/developer/D7314) (modified), commit date: 2020-04-17 15:25, hash: `0e2de2b212` Worked: Never **Short description of error** Output of various blend modes is different depending on whether byte or float image is blended Right now I am not sure which result is correct (or if any) In example file Blend is set to Linear burn with brush strength of 0.5 (I assume this will give brush pixels alpha of 0.5) I have found this bug while working on [D7314](https://archive.blender.org/developer/D7314), so I can work on fixing blend modes as well, or at least I would like to be involved in how this will be fixed. Of course, if this is actually bug and not only me being dumb. **Exact steps for others to reproduce the error** [#75844.blend](https://archive.blender.org/developer/F8479363/T75844.blend) - Open file - Paint both textures with pre-set brush Result is different As a consequence, some blend modes also work unexpected in the Video Sequence Editor, see #76376, #74440, #79333 for details. Also note that this was once changed in fdee84fd56, so we are now in a "fix-for-fix" scenario...

Richard Antalik commented

2020-04-17 20:00:33 +02:00

Added subscriber: @iss

blender-admin commented

2020-04-17 20:00:33 +02:00

#88407 was marked as duplicate of this issue

blender-admin commented

2020-04-17 20:00:33 +02:00

#86367 was marked as duplicate of this issue

blender-admin commented

2020-04-17 20:00:33 +02:00

#85996 was marked as duplicate of this issue

blender-admin commented

2020-04-17 20:00:33 +02:00

#79333 was marked as duplicate of this issue

blender-admin commented

2020-04-17 20:00:33 +02:00

#74440 was marked as duplicate of this issue

blender-admin commented

2020-04-17 20:00:33 +02:00

#78727 was marked as duplicate of this issue

Brecht Van Lommel commented

2020-04-17 20:16:28 +02:00

Added subscriber: @brecht

Brecht Van Lommel commented

2020-04-17 20:16:28 +02:00

There is also difference in color space, blending in sRGB or linear color will not give the same results.

Roman Kornev commented

2020-04-18 21:10:06 +02:00

Added subscriber: @RomanKornev

Roman Kornev commented

2020-04-18 21:10:06 +02:00

Related #75875

Richard Antalik commented

2020-04-20 11:30:47 +02:00

In #75844#912453, @brecht wrote:
There is also difference in color space

Not really in example file you are comparing results in the same colorspace.

Math for blend_color_linearburn_byte():

      const int temp = max_ii(src1[i] + src2[i] - 255, 0);
      dst[i] = (uchar)((temp * fac + src1[i] * mfac) / 255);

And float blend_color_linearburn_float():

      const float temp = max_ff(src1[i] + src2[i] - 1.0f, 0.0f);
      dst[i] = (temp * fac + src1[i] * mfac);

Regardless of corectness, both should output consistent result within bounds precision.

Following texture paint code, color values of brush are multiplied by "Strength" or "blend factor" in IMB_rectblend(), but this will be done once again in blend_color_linearburn_float() so result is wrong.
With some tweaking(multiplying only alpha, but for input2) to VSE code - apply_blend_function_float() and using Color Mix strip, output of byte and float versions are consistent.

But then looking at blend_color_add_float() for example, src1- [x] is used for premultiplication. This very much looks like a mistake. Condition expression suggests that as well. Either src2- [x] should be used, or it should be assumed that src2 color is already mutiplied by "blend factor"

  if (src2[3] != 0.0f) {
    /* unpremul > add > premul, simplified */
    dst[0] = src1[0] + src2[0] * src1[3];
    dst[1] = src1[1] + src2[1] * src1[3];
    dst[2] = src1[2] + src2[2] * src1[3];
    dst[3] = src1[3];
  }

Resolution proposal:

Always apply "blend factor" to src2 and expect it there.
Decide if "blend factor" is applied in float blend functions or not. Is there convention for this?

Note: these functions seems to be only used for VSE and texture paint.

> In #75844#912453, @brecht wrote: > There is also difference in color space Not really in example file you are comparing results in the same colorspace. Math for `blend_color_linearburn_byte()`: ``` const int temp = max_ii(src1[i] + src2[i] - 255, 0); dst[i] = (uchar)((temp * fac + src1[i] * mfac) / 255); ``` And float `blend_color_linearburn_float()`: ``` const float temp = max_ff(src1[i] + src2[i] - 1.0f, 0.0f); dst[i] = (temp * fac + src1[i] * mfac); ``` Regardless of corectness, both should output consistent result within bounds precision. Following texture paint code, color values of brush are multiplied by "Strength" or "blend factor" in `IMB_rectblend()`, but this will be done once again in `blend_color_linearburn_float()` so result is wrong. With some tweaking(multiplying only alpha, but for input2) to VSE code - `apply_blend_function_float()` and using Color Mix strip, output of byte and float versions are consistent. But then looking at `blend_color_add_float()` for example, `src1- [x]` is used for premultiplication. This very much looks like a mistake. Condition expression suggests that as well. Either `src2- [x]` should be used, or it should be assumed that src2 color is already mutiplied by "blend factor" ``` if (src2[3] != 0.0f) { /* unpremul > add > premul, simplified */ dst[0] = src1[0] + src2[0] * src1[3]; dst[1] = src1[1] + src2[1] * src1[3]; dst[2] = src1[2] + src2[2] * src1[3]; dst[3] = src1[3]; } ``` Resolution proposal: - Always apply "blend factor" to `src2` and expect it there. - Decide if "blend factor" is applied in float blend functions or not. Is there convention for this? Note: these functions seems to be only used for VSE and texture paint.

Philipp Oeser commented

2020-06-17 14:23:39 +02:00

Added subscriber: @lichtwerk

Philipp Oeser commented

2020-06-17 14:23:39 +02:00

Changed status from 'Needs Triage' to: 'Confirmed'

Philipp Oeser commented

2020-06-17 14:23:39 +02:00

I guess this can be confirmed.

Jehan Pagès commented

2020-07-09 08:55:42 +02:00

Added subscriber: @jehan

Richard Antalik commented

2020-07-10 11:30:09 +02:00

Added subscribers: @Aryeom, @Sergey, @Jeroen-Bakker

Philipp Oeser commented

2020-07-28 10:59:53 +02:00

Added subscribers: @tintwotin, @XB982

Philipp Oeser commented

2020-07-28 11:00:31 +02:00

Added subscriber: @febby

Philipp Oeser commented

2020-07-28 11:07:22 +02:00

And since this is a regression in a way (partially caused by fdee84fd56) I dare setting this to High -- @iss, correct me if I am wrong...

And since this is a regression in a way (partially caused by fdee84fd56) I dare setting this to High -- @iss, correct me if I am wrong...

tintwotin commented

2020-07-28 11:16:55 +02:00

Removed subscriber: @tintwotin

tintwotin commented

2020-07-28 11:22:47 +02:00

A detailed analyse of effect strips and strip settings duplicates(and defunct blend modes and other bugs):
https://devtalk.blender.org/t/design-proposal-effect-strips-vs-strip-settings/11987

On top of this, most effect strips doesn't seem to be multi-threaded:
https://blender.community/c/rightclickselect/QLfbbc/

A detailed analyse of effect strips and strip settings duplicates(and defunct blend modes and other bugs): https://devtalk.blender.org/t/design-proposal-effect-strips-vs-strip-settings/11987 On top of this, most effect strips doesn't seem to be multi-threaded: https://blender.community/c/rightclickselect/QLfbbc/

Richard Antalik commented

2020-07-28 13:03:05 +02:00

In #75844#986479, @lichtwerk wrote:
And since this is a regression in a way (partially caused by fdee84fd56) I dare setting this to High -- @iss, correct me if I am wrong...

I mean it is core functionality, that is broken, but it is broken for quite some time. I will try to fix this soon

> In #75844#986479, @lichtwerk wrote: > And since this is a regression in a way (partially caused by fdee84fd56) I dare setting this to High -- @iss, correct me if I am wrong... I mean it is core functionality, that is broken, but it is broken for quite some time. I will try to fix this soon

Richard Antalik self-assigned this 2020-08-10 11:16:59 +02:00

Sergey Sharybin commented

2020-08-10 12:27:23 +02:00

@iss, do you think fixing this issue is something doable/safe for 2.90?

Richard Antalik commented

2020-08-10 12:39:48 +02:00

In #75844#993763, @Sergey wrote:
@iss, do you think fixing this issue is something doable/safe for 2.90?

I think so, will continue with fix today or tomorrow and see how review will go. I don't think it would be catastrophic if it gets committed only in 2.91 though.

> In #75844#993763, @Sergey wrote: > @iss, do you think fixing this issue is something doable/safe for 2.90? I think so, will continue with fix today or tomorrow and see how review will go. I don't think it would be catastrophic if it gets committed only in 2.91 though.

Sergey Sharybin commented

2020-08-10 14:48:28 +02:00

Ok, for now putting to 2.90 workboard.
If the fix goes deeper or requires more time, will re-schedule the release and priority.

Ok, for now putting to 2.90 workboard. If the fix goes deeper or requires more time, will re-schedule the release and priority.

Gilberto Rodrigues commented

2020-10-16 05:16:15 +02:00

Added subscriber: @Gilberto.R

Ahryeom HAN commented

2020-10-25 17:08:49 +01:00

Thanks for working on this.
Will the fix also take care of the common non-linear blending issue?
Currently in blender, with alpha over blend mode and with a gaussian blur effect :

Thanks for working on this. Will the fix also take care of the common non-linear blending issue? Currently in blender, with alpha over blend mode and with a gaussian blur effect : ![blur black silhouette.jpg](https://archive.blender.org/developer/F9088596/blur_black_silhouette.jpg)

Evan Wilson commented

2020-12-23 18:32:07 +01:00

Added subscriber: @EAW

Philipp Oeser commented

2021-03-08 23:44:02 +01:00

Added subscribers: @Gurra, @mano-wii, @tintwotin

Philipp Oeser commented

2021-03-08 23:44:28 +01:00

Added subscriber: @DMan

Philipp Oeser commented

2021-03-08 23:45:05 +01:00

@iss: did you have a chance to give this another look?

Philipp Oeser commented

2021-05-21 12:41:22 +02:00

Added subscriber: @DragonLEE

Philipp Oeser commented

2021-05-21 12:41:42 +02:00

Yet another release with this in...

Seyit Ali commented

2021-05-21 15:40:06 +02:00

UPBGE version working
https://dev-files.blender.org/file/data/annrdtbtae7neo4zestd/PHID-FILE-yf6umhiziguhfmepqtdl/Blend_Modes.jpg

UPBGE version working https://dev-files.blender.org/file/data/annrdtbtae7neo4zestd/PHID-FILE-yf6umhiziguhfmepqtdl/Blend_Modes.jpg

Steffen Hartmann commented

2021-05-21 18:32:28 +02:00

Added subscriber: @Harti

Steffen Hartmann commented

2021-05-21 18:32:28 +02:00

Now that is a sad story overall.

Juan commented

2021-09-29 16:39:07 +02:00

Added subscriber: @jc4d

Juan commented

2021-09-29 16:39:07 +02:00

Hi, it's still broken in 2.93.4.

Kristian Holmgren commented

2021-09-29 16:58:22 +02:00

And, for that matter, in 3.0.0 (2021-09-29, 367775ac6a)

In #75844#1228187, @jc4d wrote:
Hi, it's still broken in 2.93.4.

And, for that matter, in 3.0.0 (2021-09-29, 367775ac6a2d) > In #75844#1228187, @jc4d wrote: > Hi, it's still broken in 2.93.4.

Yevgeny Makarov commented

2021-12-19 22:23:24 +01:00

Added subscriber: @jenkm

Tora Stiefel commented

2022-01-02 15:36:45 +01:00

Added subscriber: @Tora-Victoria

Tora Stiefel commented

2022-01-02 15:36:45 +01:00

Still broken in 3.0

Only possible to adjust opacity in cross - not in difference, add etc. Just kicks in - no transition.

Still broken in 3.0 Only possible to adjust opacity in cross - not in difference, add etc. Just kicks in - no transition.

Kristian Holmgren commented

2022-01-02 20:13:03 +01:00

As I wrote in September, it didn't work for me either in 3.0.0, but it seems to be finally working now in today's build of 3.1.0.

Tora Stiefel commented

2022-01-02 20:17:35 +01:00

Well I am using 3.1.0 Alpha and the transition is not working there - just kicks in, no transition if you choose other than cross blending mode

Kristian Holmgren commented

2022-01-02 21:08:48 +01:00

Have you tried building it today? I'm on 3.1.0 Alpha (2022-01-02, 8be217ada5) on Linux, and for the first time ever Overlay, Difference, Add, etc. do come in gradually, instead of just 'kicking in,' like you say.

Have you tried building it today? I'm on 3.1.0 Alpha (2022-01-02, 8be217ada575) on Linux, and for the first time ever Overlay, Difference, Add, etc. do come in gradually, instead of just 'kicking in,' like you say.

Kristian Holmgren commented

2022-01-02 21:11:01 +01:00

I should perhaps say that I am checking this in the VSE, perhaps this hasn't been fixed in other places where blend modes are used...

Tora Stiefel commented

2022-01-02 21:13:38 +01:00

I am using mac and video editing. I just tried this today. Using Blender as VSE for the first time coming from Premiere.

Richard Antalik commented

2022-01-03 00:23:18 +01:00

I have fixed issue with blend factor few days ago, there are still some issues though so I haven't closed this task yet. For most cases this could be considered as resolved in 3.1 alpha.

I will add more fixes and close soon.

I have fixed issue with blend factor few days ago, there are still some issues though so I haven't closed this task yet. For most cases this could be considered as resolved in 3.1 alpha. I will add more fixes and close soon.

Richard Antalik commented

2022-01-04 03:06:12 +01:00

Changed status from 'Confirmed' to: 'Resolved'

Richard Antalik closed this issue

2022-01-04 03:06:12 +01:00

Richard Antalik commented

2022-01-04 03:06:12 +01:00

I have pushed fixes to resolve blending issues:
bdcc258305
9c85acf61d
d2cc672b0c

In addition, I have added visual tests to SVN with what I think is good state and also reference images made with GIMP - rBL62783.

There are differences between byte and float blend modes still, but this is mostly due to way how color values and alpha are handled internally - Linear premultiplied vs sRGB straight.

Running tests with GIMP image results as reference will also reveal differences, but this seems to be due to how values are handled there. I would expect EXR's to blend exacly in these 2 programs, but they do not even if blend functions use same formula, so this is not issue with blend modes.

I have pushed fixes to resolve blending issues: bdcc258305 9c85acf61d d2cc672b0c In addition, I have added visual tests to SVN with what I think is good state and also reference images made with GIMP - rBL62783. There are differences between byte and float blend modes still, but this is mostly due to way how color values and alpha are handled internally - Linear premultiplied vs sRGB straight. Running tests with GIMP image results as reference will also reveal differences, but this seems to be due to how values are handled there. I would expect EXR's to blend exacly in these 2 programs, but they do not even if blend functions use same formula, so this is not issue with blend modes.

Tora Stiefel commented

2022-01-04 04:21:40 +01:00

Hi Richard

Perhaps I was not clear about the blend issue. I lust downloaded the newest version of Blender 3.1.0 Alpha and tested the blend feature. Now when loading in images in VSE images are loaded in as Alpha Over. The Alpha Over works fine per se, as do all the other blending modes. The issue here is that only the Alpha Over and Cross blending modes can be animated and get a transition - all the other modes are still either on or off. The transition is still not working. So the issue is not solved.

Tested on a mac with PNG and JPG images.

Hi Richard Perhaps I was not clear about the blend issue. I lust downloaded the newest version of Blender 3.1.0 Alpha and tested the blend feature. Now when loading in images in VSE images are loaded in as Alpha Over. The Alpha Over works fine per se, as do all the other blending modes. The issue here is that only the Alpha Over and Cross blending modes can be animated and get a transition - all the other modes are still either on or off. The transition is still not working. So the issue is not solved. Tested on a mac with PNG and JPG images.

Richard Antalik commented

2022-01-04 05:35:13 +01:00

@Tora-Victoria I would recommend to make new report. Please share build hash (When you click on Help > Report a Bug it will be filled automatically) attach blend file with 2 color strips not blending properly or images if you want. Testing with ea8d749587 I can't see any issues with blend modes in side panel or using colormix effect. So not sure if I am checking same issue as you.

@Tora-Victoria I would recommend to make new report. Please share build hash (When you click on Help > Report a Bug it will be filled automatically) attach blend file with 2 color strips not blending properly or images if you want. Testing with ea8d749587dd I can't see any issues with blend modes in side panel or using colormix effect. So not sure if I am checking same issue as you.

Tora Stiefel commented

2022-01-04 07:16:17 +01:00

I´ll do that - cheers

Tora Stiefel commented

2022-01-04 07:19:42 +01:00

Here is the link if you would like to read my report on the issue:

https://developer.blender.org/T94625

Here is the link if you would like to read my report on the issue: https://developer.blender.org/T94625

Thomas Dinges added this to the 2.93 LTS milestone 2023-02-07 18:46:39 +01:00

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

Blend modes are broken #75844