Add Lacunarity and Normalize Inputs to the Noise Texture Node #110839

Hoshinova · 2023-08-05T19:58:06+02:00

Hoshinova commented

2023-08-05 19:58:06 +02:00

This PR adds the Lacunarity and Normalize inputs from the Voronoi Texture node to the Noise Texture node.

The Lacunarity input controls by which factor each successive Perlin noise octave is scaled. Currently it is hard coded to a value of 2.0. This input makes it changeable to the user.

The Noise node normalizes it's output by default. The Normalize input makes it possible for the user to turn it off.
To keep the behavior consistent with past versions it is on by default.

This PR adds the Lacunarity and Normalize inputs from the [Voronoi Texture node](https://projects.blender.org/blender/blender/pulls/106827) to the Noise Texture node. The Lacunarity input controls by which factor each successive Perlin noise octave is scaled. Currently it is hard coded to a value of `2.0`. This input makes it changeable to the user. The Noise node normalizes it's output by default. The Normalize input makes it possible for the user to turn it off. To keep the behavior consistent with past versions it is on by default.

❤️ 2 🚀 3

Hoshinova added 2 commits 2023-08-05 19:58:16 +02:00

e91b30804f -- Add Lacunarity and Normalize inputs

buildbot/vexp-code-patch-coordinator Build done. Details

96980b24b7 -- Rename variable "normalize" to "use_normalize" for OSL to keep consistent with Voronoi.

Hoshinova requested review from Jacques Lucke 2023-08-05 20:01:49 +02:00

Hoshinova requested review from Omar Emara 2023-08-05 20:01:49 +02:00

Hoshinova commented

2023-08-05 20:04:07 +02:00

I've only added the inputs for now but knowing how long the review process takes, I created the PR so design questions can already be discussed.

Hoshinova commented

2023-08-05 20:08:02 +02:00

Code wise the fractal_voronoi_x_fx function from 46a17df415/intern/cycles/kernel/svm/voronoi.h (L883) has exactly the same fractalization logic as the one used for the Noise Texture node with the additional inputs already added, so I plan to replace the corresponding Noise Texture functions with a modified version of that.

Code wise the `fractal_voronoi_x_fx` function from https://projects.blender.org/blender/blender/src/commit/46a17df41546d90f74485378bfddec620d0a0f93/intern/cycles/kernel/svm/voronoi.h#L883 has exactly the same fractalization logic as the one used for the Noise Texture node with the additional inputs already added, so I plan to replace the corresponding Noise Texture functions with a modified version of that.

Omar Emara commented

2023-08-05 21:48:37 +02:00

What is a practical example for turning off normalization? That doesn't seem very useful to me.
Also adding Lacunarity kind of turns the noise texture into a Musgrave texture doesn't it?

What is a practical example for turning off normalization? That doesn't seem very useful to me. Also adding Lacunarity kind of turns the noise texture into a Musgrave texture doesn't it?

Hoshinova commented

2023-08-06 01:01:53 +02:00

What is a practical example for turning off normalization?

Pretty mich all cases where the height of the Fractal Perlin noise actually matters, like with procedural terrain generation.

Also adding Lacunarity kind of turns the noise texture into a Musgrave texture doesn't it?

Surprisingly no. The Dimension input of the Musgrave Texture does a very weird thing, where it acts like the Roughness inputs but is the value is taken as Roughness = Lacunarity^(-Dimension) d870f9e841/intern/cycles/kernel/svm/musgrave.h (L202), which is completely unintuitive to even the most technical users, which is probably why I see many artists shy away from it.
So while the Musgrave Texture has a Lacunarity input it effectively has no usable Roughness input.

But you're bringing up an interesting point. I've always wondered why Blender has 2 Perlin noise nodes, neither of which being called Perlin noise, but both of which having features the other one doesn't.

Ideally my plan is to merge the Musgrave Texture node and the Noise Texture node into a single Perlin noise node, but implementation wise that still has many open questions regarding backwards compatibility.

Which is why for now I decided to at least make the Noise Texture have it's full set of features.

> What is a practical example for turning off normalization? Pretty mich all cases where the height of the Fractal Perlin noise actually matters, like with procedural terrain generation. > Also adding Lacunarity kind of turns the noise texture into a Musgrave texture doesn't it? Surprisingly no. The Dimension input of the Musgrave Texture does a very weird thing, where it acts like the Roughness inputs but is the value is taken as Roughness = Lacunarity^(-Dimension) https://projects.blender.org/blender/blender/src/commit/d870f9e841a5ea8c289414808e9b8390bb4825b2/intern/cycles/kernel/svm/musgrave.h#L202, which is completely unintuitive to even the most technical users, which is probably why I see many artists shy away from it. So while the Musgrave Texture has a Lacunarity input it effectively has no usable Roughness input. But you're bringing up an interesting point. I've always wondered why Blender has 2 Perlin noise nodes, neither of which being called Perlin noise, but both of which having features the other one doesn't. Ideally my plan is to merge the Musgrave Texture node and the Noise Texture node into a single Perlin noise node, but implementation wise that still has many open questions regarding backwards compatibility. Which is why for now I decided to at least make the Noise Texture have it's full set of features.

Hoshinova commented

2023-08-06 01:30:11 +02:00

Also code wise I noticed, that there is actually no need to replace any functions. Simple replacing 2.0 with lacunarity already does the trick. And for the normalization an if statement around the corresponding code should also be enough.

So I don't see any downsides to this, neither user experience nor code wise.

Also code wise I noticed, that there is actually no need to replace any functions. Simple replacing 2.0 with lacunarity already does the trick. And for the normalization an if statement around the corresponding code should also be enough. So I don't see any downsides to this, neither user experience nor code wise.

Omar Emara commented

2023-08-06 07:54:01 +02:00

Since work to bring features from Musgrave texture to the noise texture was approved before in D7065, I see no reason to block the addition of Lacunarity.

I still don't see how normalization could be useful, Perlin noise intrinsically has no reasonable maximum, so leaving it to its own maximum brings us what value? Can you demonstrate a visual practical example?

Since work to bring features from Musgrave texture to the noise texture was approved before in D7065, I see no reason to block the addition of Lacunarity. I still don't see how normalization could be useful, Perlin noise intrinsically has no reasonable maximum, so leaving it to its own maximum brings us what value? Can you demonstrate a visual practical example?

Raiko commented

2023-08-06 15:22:00 +02:00

First-time contributor

A while ago I posted this picture in the Pull Request for Fractal Voronoi noise:

Both the water and the fire were created using a combination of Voronoi Texture node and Noise Texture node.
The fire and the water were created using two separate node trees and as you can see they are aligned such that the fire sits perfectly on top of the water.
The fact that the Noise Texture node normalizes the Fractal Perlin noise makes that alignment much more difficult, because it introduces an offset to the midlevel of the noise which scales along as you scale the noise output.
To combat this issue I had to manually un-normalize the output for every node, which gets tedious if the node setup is complex like in this case.

Alternatively I could've used the Musgrave Texture node which unlike the Noise Texture node doesn't normalize it's output, but found the Dimension input to be hard to work with, just as Hoshinova described.

Apart from that the Noise Texture node normalization tends to overcompensate for higher Roughness values making the entire output just 0.5ish.

A while ago I posted this picture in the Pull Request for Fractal Voronoi noise: ![250.png](/attachments/e377bafc-9c88-4587-ae7e-bc06679cdd8b) Both the water and the fire were created using a combination of Voronoi Texture node and Noise Texture node. The fire and the water were created using two separate node trees and as you can see they are aligned such that the fire sits perfectly on top of the water. The fact that the Noise Texture node normalizes the Fractal Perlin noise makes that alignment much more difficult, because it introduces an offset to the midlevel of the noise which scales along as you scale the noise output. To combat this issue I had to manually un-normalize the output for every node, which gets tedious if the node setup is complex like in this case. Alternatively I could've used the Musgrave Texture node which unlike the Noise Texture node doesn't normalize it's output, but found the Dimension input to be hard to work with, just as Hoshinova described. Apart from that the Noise Texture node normalization tends to overcompensate for higher Roughness values making the entire output just 0.5ish.

250.png

8.8 MiB

👍 2

Hoshinova commented

2023-08-06 16:03:55 +02:00

@OmarEmaraDev So there you have it.
Thanks again @Raiko for your explanation!

@OmarEmaraDev So there you have it. Thanks again @Raiko for your explanation!

Hoshinova changed title from ~~Add Lacunarity and Normalize Inputs to the Noise Texture Node~~ to WIP: Add Lacunarity and Normalize Inputs to the Noise Texture Node

2023-08-06 16:04:49 +02:00

Hoshinova requested review from Brecht Van Lommel 2023-08-06 16:04:57 +02:00

Brecht Van Lommel commented

2023-08-08 18:03:38 +02:00

@blender-bot package

Blender Bot commented

2023-08-08 18:03:41 +02:00

Package build started. Download here when ready.

Package build started. [Download here](https://builder.blender.org/download/patch/PR110839) when ready.

Brecht Van Lommel requested changes 2023-08-08 18:06:38 +02:00

Brecht Van Lommel left a comment

The functionality and Cycles implementation look fine to me.

I guess the Eevee implementation is still missing since it does not use the parameters.

The functionality and Cycles implementation look fine to me. I guess the Eevee implementation is still missing since it does not use the parameters.

Hoshinova added 1 commit 2023-08-09 15:27:24 +02:00

buildbot/vexp-code-patch-coordinator Build done. Details

e1b82960d3 -- Implement new inputs to shader and geometry nodes.

Hoshinova commented

2023-08-09 16:37:34 +02:00

As you may have noticed I changed the way normalization is done.
In the old code the octaves are first normalized individually then added together then normalized again.
In the new code both steps are combined into one normalization step after the fractalization process.
This makes the new code slightly faster as it saves a few additions and multiplications but more importantly it separates the normalization step from the fractalization process, making it possible to easily turn on or off.

To dispel any possible doubts about whether or not the new code always computes the same output as the old code and for future reference here is a quick proof:

The way the old code computes the output can be explicitly written as:

output =\frac{ \sum_{i=0}^{detail} ((0.5*snoise(p*lacunarity^i)+0.5)*roughness^i) }{maxamp}

where maxamp = \sum_{i=0}^{detail}(roughness^i)

The way the new code computes the output can be explicitly written as:

output =0.5*\frac{ \sum_{i=0}^{detail} (snoise(p*lacunarity^i)*roughness^i) }{maxamp}+0.5

The old and new code always compute the same output iff the expression on the right side of both equations are equivalent.
Thus:

\frac{ \sum_{i=0}^{detail} ((0.5*snoise(p*lacunarity^i)+0.5)*roughness^i) }{maxamp}

=\frac{ \sum_{i=0}^{detail} (0.5*snoise(p*lacunarity^i)*roughness^i+0.5*roughness^i) }{maxamp}

=\frac{ \sum_{i=0}^{detail} (0.5*snoise(p*lacunarity^i)*roughness^i) }{maxamp} + \frac{ \sum_{i=0}^{detail} (0.5*roughness^i) }{maxamp}

=0.5*\frac{ \sum_{i=0}^{detail} (snoise(p*lacunarity^i)*roughness^i) }{maxamp} + 0.5*\frac{ \sum_{i=0}^{detail} (roughness^i) }{maxamp}

=0.5*\frac{ \sum_{i=0}^{detail} (snoise(p*lacunarity^i)*roughness^i) }{maxamp} + 0.5*\frac{ \sum_{i=0}^{detail} (roughness^i) }{ \sum_{i=0}^{detail}(roughness^i)}

=0.5*\frac{ \sum_{i=0}^{detail} (snoise(p*lacunarity^i)*roughness^i) }{maxamp}+0.5

This concludes the proof.

As you may have noticed I changed the way normalization is done. In the old code the octaves are first normalized individually then added together then normalized again. In the new code both steps are combined into one normalization step after the fractalization process. This makes the new code slightly faster as it saves a few additions and multiplications but more importantly it separates the normalization step from the fractalization process, making it possible to easily turn on or off. To dispel any possible doubts about whether or not the new code always computes the same output as the old code and for future reference here is a quick proof: The way the old code computes the output can be explicitly written as: $$output =\frac{ \sum_{i=0}^{detail} ((0.5*snoise(p*lacunarity^i)+0.5)*roughness^i) }{maxamp}$$ where $maxamp = \sum_{i=0}^{detail}(roughness^i)$ The way the new code computes the output can be explicitly written as: $$output =0.5*\frac{ \sum_{i=0}^{detail} (snoise(p*lacunarity^i)*roughness^i) }{maxamp}+0.5$$ The old and new code always compute the same output iff the expression on the right side of both equations are equivalent. Thus: $$\frac{ \sum_{i=0}^{detail} ((0.5*snoise(p*lacunarity^i)+0.5)*roughness^i) }{maxamp}$$ $$=\frac{ \sum_{i=0}^{detail} (0.5*snoise(p*lacunarity^i)*roughness^i+0.5*roughness^i) }{maxamp}$$ $$=\frac{ \sum_{i=0}^{detail} (0.5*snoise(p*lacunarity^i)*roughness^i) }{maxamp} + \frac{ \sum_{i=0}^{detail} (0.5*roughness^i) }{maxamp}$$ $$=0.5*\frac{ \sum_{i=0}^{detail} (snoise(p*lacunarity^i)*roughness^i) }{maxamp} + 0.5*\frac{ \sum_{i=0}^{detail} (roughness^i) }{maxamp}$$ $$=0.5*\frac{ \sum_{i=0}^{detail} (snoise(p*lacunarity^i)*roughness^i) }{maxamp} + 0.5*\frac{ \sum_{i=0}^{detail} (roughness^i) }{ \sum_{i=0}^{detail}(roughness^i)}$$ $$=0.5*\frac{ \sum_{i=0}^{detail} (snoise(p*lacunarity^i)*roughness^i) }{maxamp}+0.5$$ This concludes the proof.

Thomas Dinges commented

2023-08-09 16:45:12 +02:00

@blender-bot package

Blender Bot commented

2023-08-09 16:45:15 +02:00

Package build started. Download here when ready.

Package build started. [Download here](https://builder.blender.org/download/patch/PR110839) when ready.

Hoshinova added 1 commit 2023-08-09 18:14:16 +02:00

buildbot/vexp-code-patch-coordinator Build done. Details

893b425dcc Merge branch 'main' into add-noise-inputs

Ray molenkamp commented

2023-08-09 18:26:53 +02:00

@blender-bot package

Blender Bot commented

2023-08-09 18:26:56 +02:00

Package build started. Download here when ready.

Package build started. [Download here](https://builder.blender.org/download/patch/PR110839) when ready.

Hoshinova commented

2023-08-09 18:31:06 +02:00

Here's a test file to show that the new behavior is same for all shader node implementations:
Noise_Test_File.blend

The output of the new Noise Texture node should be identical with the old one when Normalize is turned on and Lacunarity is set to 2.0.

Here's a test file to show that the new behavior is same for all shader node implementations: [Noise_Test_File.blend](/attachments/ad90643e-26f4-4f6e-84c6-cc001b85112a) The output of the new Noise Texture node should be identical with the old one when `Normalize` is turned on and `Lacunarity` is set to 2.0.

Noise_Test_File.blend

4.3 MiB

Jacques Lucke requested changes 2023-08-09 21:44:07 +02:00

Jacques Lucke left a comment

The code looks fine to me too, but it's breaking a bunch of tests currently. That should be fixed.
I haven't looked at it in detail yet, but the reason might just be missing versioning code for the normalize property.

The code looks fine to me too, but it's breaking a bunch of tests currently. That should be fixed. I haven't looked at it in detail yet, but the reason might just be missing versioning code for the normalize property.

source/blender/makesdna/DNA_node_types.h Outdated

						
				@ -1216,3 +1216,3 @@

				  NodeTexBase base;

				  int dimensions;

				  char _pad[4];

				  int normalize;

Jacques Lucke commented

2023-08-09 21:44:04 +02:00

Use uint8_t for normalize and keep a padding of 3 bytes.

Use `uint8_t` for `normalize` and keep a padding of 3 bytes.

Hoshinova marked this conversation as resolved

Hoshinova added 1 commit 2023-08-10 14:34:29 +02:00

buildbot/vexp-code-patch-coordinator Build done. Details

312ba903f8 -- Do versioning

Jacques Lucke commented

2023-08-10 14:38:08 +02:00

@blender-bot package

Blender Bot commented

2023-08-10 14:38:12 +02:00

Package build started. Download here when ready.

Package build started. [Download here](https://builder.blender.org/download/patch/PR110839) when ready.

Hoshinova added 1 commit 2023-08-10 15:28:19 +02:00

54726d83cf BLENDER_SOURCE_CODE_COMMIT commit message number: 17

~~Charlie Jolly referenced this pull request 2023-08-10 21:50:19 +02:00~~

Node: Gabor Noise Texture #110802

Hoshinova added 1 commit 2023-08-11 11:09:19 +02:00

buildbot/vexp-code-patch-coordinator Build done. Details

e848057d17 BLENDER_SOURCE_CODE_COMMIT commit message number: 18

Jacques Lucke approved these changes 2023-08-15 11:49:29 +02:00

Jacques Lucke left a comment

LGTM.

Jacques Lucke commented

2023-08-15 11:50:59 +02:00

Ah, you still have to increase BLENDER_FILE_SUBVERSION in this patch.

@blender-bot build

Ah, you still have to increase `BLENDER_FILE_SUBVERSION` in this patch. @blender-bot build

Hoshinova added 3 commits 2023-08-15 14:31:24 +02:00

c0977f3d82 Merge branch 'main' into add-noise-inputs

6d5d30f6e5 Merge branch 'main' into add-noise-inputs

buildbot/vexp-code-patch-coordinator Build done. Details

fae5dc1255 -- Increase BLENDER_FILE_SUBVERSION to 16

Hoshinova changed title from ~~WIP: Add Lacunarity and Normalize Inputs to the Noise Texture Node~~ to Add Lacunarity and Normalize Inputs to the Noise Texture Node

2023-08-15 14:33:39 +02:00

Brecht Van Lommel approved these changes 2023-08-15 14:36:17 +02:00

Jacques Lucke commented

2023-08-15 14:44:35 +02:00

@blender-bot build

Jacques Lucke commented

2023-08-15 15:13:52 +02:00

If you don't want us to merge a specific patch, add the WIP: prefix to the commit title. Besides that, I don't know what you expect us to do when we want to merge it?

If you don't want us to merge a specific patch, add the `WIP:` prefix to the commit title. Besides that, I don't know what you expect us to do when we want to merge it?

Omar Emara requested changes 2023-08-15 17:06:23 +02:00

source/blender/makesrna/intern/rna_nodetree.cc Outdated

						
				@ -5157,0 +5157,4 @@

				  prop = RNA_def_property(srna, "normalize", PROP_BOOLEAN, PROP_NONE);

				  RNA_def_property_boolean_sdna(prop, NULL, "normalize", 0);

				  RNA_def_property_ui_text(prop, "Normalize", "Normalize output Distance to 0.0 to 1.0 range");

Omar Emara commented

2023-08-15 17:02:17 +02:00

Misplaced mentioned of Distance.

Hoshinova marked this conversation as resolved

source/blender/nodes/shader/nodes/node_shader_tex_noise.cc Outdated

						
				@ -31,0 +32,4 @@

				      .min(0.0f)

				      .max(1000.0f)

				      .default_value(2.0f)

				      .description("The scale of a Voronoi layer relative to that of the previous layer");