# Fields and Anonymous Attributes [Proposal]

This document proposes new concepts for geometry nodes that solve many of the existing usability problems and add a lot of new flexibility.

## Fields

### Definition

We have different socket type for different primitive data types. For example, there is a float, a vector and a boolean socket type. What these types have in common is that they always just carry a single value of their data type. This is what this proposal wants to change. Instead of always carrying a single value, now these sockets carry a âfieldâ (the naming is of course open to debate, but Iâll stick to that for now). Everything else about the sockets stays the same. Nodes using these sockets also donât really change. The math node shown in the image below looks exactly the same with and without this proposal.

So what is a field? A field is a function that outputs a single value of the corresponding data type. A float field computes a float and a boolean field computes a boolean. While the field can only have a single output, it is allowed to have an arbitrary number of inputs/dependencies. These inputs are provided by the node evaluating the field. A field that has no inputs always outputs the same value. Therefore, it is a constant and works exactly the same as the sockets we have today.

Every dependency of a field should also have a default value. This way it can be evaluated even when not all dependencies are available.

Note, this definition of a field is different from definitions in mathematics and physics. While the term is inspired by the fields used in physics, they can depend on more than just space and time in this proposal.

### Usage

One of the simplest demonstrations of where this can be used is in a node that displaces all points in some direction. In the image below the Attribute node is optional, because if the vector input to the noise node is not connected, it implicitly depends on the position attribute (this is the same behavior as in shader nodes). Note how it would be possible to use the same texture node group in shader and geometry nodes in this setup.

In the next example, I want to build a node group that first subdivides the incoming geometry and then distributes points on the new geometry. Furthermore, it modifies the incoming density weights a bit. Note how it is possible to edit the geometry without making the fields incompatible in most cases.

So far we only depended on attributes that exist on the geometry. However, it is also possible to depend on other things like the index. Note that in the example below, the Delete Geometry node would need a domain dropdown, which seems reasonable.

With fields one can build node groups that compose well and encapsulate high level concepts for the user. For example, one can build âobject falloffsâ, i.e. float fields that are one very close to an object and are closer to zero the further you are away.

In the above example, the falloff float field contains a reference to the objectâs geometry. Otherwise it would not be able to compute the distance to it. This solution opens up even more possibilities with respect to attribute transfer between geometries. In this example, the `weight` attribute from a different geometry is used to distribute points. The âtransfer fieldâ will find the closest point on the surface of the geometry and output the attribute value on that point. The same approach can work for volume grids as well.

To store a field as an attribute on a geometry, the Attribute Set node can be used. This way it can be passed e.g. to the renderer.

## Anonymous Attributes

### Definition

An anonymous attribute is like normal attributes with a few key differences:

• It does not have a user provided name. Instead it has some randomly generated name that the user should never see. It might also have a âdebug nameâ that helps understand where the attribute comes from, but this name cannot be used as identifier.
• Its generated name is reference counted.
• Once no one has a reference to the name anymore, the attribute can be freed automatically.

### Usage

As a first example, letâs have a look at the Point Distribute node. Instead of generating attributes with hardcoded names, it generates anonymous attributes for rotation and normal. Furthermore, it outputs two vector fields which depend on the anonymous attributes.

Note, the Point Distribute node can see that the normal output is not used, so it does not have to be generated. Also, once the Point Instance node has evaluated the rotation input, this vector field is freed. This results in the reference count of the anonymous attribute to go to zero. Therefore the anonymous rotation attribute is not transferred to the new points (if we would support that).

The next example shows how anonymous attributes and fields can be used for procedural mesh modelling. Note how it is possible to subdivide some faces without making the other selections incompatible. Another nice aspect is that nodes only generate the selections that are needed and they will be freed automatically once they are not needed anymore.

Itâs possible to combine selections as well. E.g. in this example every third side face of the cone is extruded.

Expanding an existing selection can look as shown below. Note that in order to work well this node needs the geometry.

If the user wants to give an output attribute a name explicitly, an Attribute Set node can be used.

## Debugging

Fields with dependencies are harder to debug than attributes that are stored on a geometry. That is because such fields need more context (usually a geometry) to be evaluated to actual values.

### Node Warnings

The first level of defense against wrong use of fields are warnings on nodes. Whenever the user passes a field to a node that cannot provide all inputs to the field a warning should be shown. This typically happens when the field depends on an attribute that does not exist on the geometry or when the node only allows a constant value for an input.

### Socket Inspection

When hovering over a socket we can show the last evaluated data that passed through the socket. This is planned for some time already and still has to be implemented. For fields socket inspection should show the fields dependencies. In some cases it could also show the entire field as an expression. For anonymous attributes the debug name is shown.

Showing fields in the spreadsheet next to other geometry data can be very useful. Unfortunately, it is a bit more difficult, because a field is not just a list of values. Unless the field is a constant or only depends on the index, it is necessary to provide context for the field evaluation. This context is typically given in the form of a geometry that the field is evaluated on.

One solution that I currently like is to extend the concept of the viewer node. Instead of just accepting a geometry, the viewer could also have a field as input. In this case the first column of the spreadsheet can become the evaluated field.
Different shortcuts can be assigned to connecting a geometry or field to the viewer node. This way one can set one geometry as context and then quickly select different fields to show in the spreadsheet. Anonymous attributes can be shown in the spreadsheet like any other field.

### Viewport

Once the viewer node can display geometry in the viewport, we can also show the field that is connected to the viewer in the viewport.

## Evaluation

The fields and anonymous attribute concepts do make the socket value and attribute concepts more complex. For me, these complications are well worth it given the problems that they solve and the additional flexibility they provide. With the mentioned debugging solutions, exploring what fields are should become much simpler.

Here are some benefits of this proposal:

• Attribute processing is not linear anymore.
• Temporary attributes are not needed anymore.
• Compared to lists, fields donât become incompatible when geometry changes in most cases.
• Temporary data that has to be stored on the geometry (e.g. selections) does not need to be named explicitly and is freed automatically.
• Node groups can be shared between geometry and shading nodes.
• Texturing concepts from shading nodes map directly to geometry nodes, so less relearning is required.
• Instead of having options in nodes to determine which output attributes should be generated, this is entirely determined by what output sockets are used.

## Related discussion points

This proposal also benefits from more development regarding compact nodes, socket type inferencing (automatically changing socket types depending on what is selected) and list sockets. However, those topics are orthogonal to what is presented here.

### Constant Sockets UI

It might be useful to be able to show e.g. integer sockets differently if they do not allow a field with dependencies. Typical cases for such inputs are seed sockets and inputs to mesh primitive nodes. Generally, it should be obvious from the context when a socket does not support a field, but maybe not always.

One proposal in T74967 was to show these inputs as rectangles instead of circles. Getting this to work on built-in nodes is fairly straight forward, but making it work in more general becomes significantly harder because it requires something like type inferencing that we donât have yet. In the image below, all the highlighted sockets should be drawn differently because they only allow constant values.

Given the complexity of a good implementation and the fact that in most cases this should be clear from context, I donât consider this to be a blocking issue for this proposal.

### Immediate Attributes and Expressions

Currently, it is possible to write an attribute name directly into the node that is using the attribute. In the current state of the proposal that is not really possible, because the enum to switch between value and attribute is not necessary anymore (both can be passed into the float socket). It can still be very convenient and less verbose to type the name directly into the node though. Extending the concept further, it makes sense to not only be able to type attribute names but entire expressions into the socket.

My currently preferred solution is that one can just click on the float socket and start typing an expression. Not exactly sure how that would work with boolean sockets yet though. This is part of the compact nodes discussion.

48 Likes

That sounds wonderful. One concern I have in regard to procedural modeling is : output attributes (such as the âtop faceâ of a box primitive) are freed if not used right after the primitive node -do I understand this correctly? So what if you want to extrude that top face, but not immediately? You want to first set down a subdivide node (which I assume interpolates the top face attribute along the newly created âtop facesâ), and add the extrude node after that. Does that mean we need to set attribute the top face so that is stays there permanently, and can be used after the subdivide node?

Iâm not sure if Iâm being clear, to reformulate : the process of modeling an object might involve intermediate operations before needing to pass a given output attribute. Does that mean that attribute has to be âsavedâ explicitly so that it can be used later?

Apart from that, this sounds great. And I finally understand what this texture node group is aboutâŚ it does make sense being able to use it between shading and geometryâŚ cool idea.

Second question that just sprang to mind : what if I want to use a fixed vector as input, not a field? Do I have to connect a fixed vector to the vector field socket? I canât just leave it blank otherwise itâll use the position values, right?

The top face attribute is only be removed when it is not referenced anymore. It is not necessary that the selection is used immediatly in the next node. This is shown in one of the examples with a cone where the selection is used after the subdivision.

1 Like

There is only a hand full of nodes that implicitly depend on the position attribute if nothing is connected. You can see that because you cannot change the vector value of the input socket. For those nodes youâd indeed have to use a separate vector input node if you wanted to use the node with a constant vector.

I see! the attribute being passed references a mesh thatâs not subdivided yet, so Iâm surprised the next node knows how to apply it correctly to the subdivided meshâŚ but if that works, Iâm all for it! The only downside I can see in that case : thereâs quite a bit more socket connections necessary.

Does that proposal replace the attribute processor or does it complement it?

Fair enough, that makes sense. Plus if the sockets are shaped differently I guess thereâs no room for error.

Important difference, the selection field does not reference the mesh, only the attribute name. So the mesh can change as a long as its attributes stay intact.

More socket connections are required indeed, which is kind of the goal (instead of relying on matching strings in both nodes).

This is currently designed as a replacement for the attribute processor, but both proposals donât really conflict with each other.

Ok ! thanks Jacques for answering my concerns. Another question comes to mind, can we still reference attributes by name, or is socket linking the only way to use them now?

Oh, and regarding the other thread by @Erindale where he suggested having specific nodes to work on lists : would those still be relevant? Since with fields we can work on lists just like we work with unique values? (using the same nodes, etc.)

And finally, where do loops fit in this?

That will add quite a high level of abstraction into the mix

1 Like

Would you be able to âaddâ an anonymous attribute to a ânormalâ attribute? Like if I have a selection based on normal angle through attribute compare, then want to add it to an extruded âtop faceâ or something like that.

Once you talked about selections, I would like to ask something. How would this work with user controlled selections? Select the side of a cone, or each nth face, is one thing. But how one could select an specific face to be extruded? Can the anonymous attributes be used to select an user selected procedurally generated face?

Thanks and keep up the amazing work!

1 Like

Sounds really promising! I love the idea of being able to share groups between shaders and geo nodes. It would be a tall order but perhaps at some point a way down the line, the various node trees Blender gains can be merged into a single scene graph where artists can create and control and shade multiple assets procedurally (that might lend itself to the old pages prototype as well if different graph levels were actually functionally different).

I really like that things look more visual in this mockup.

Does this mean that the fields system wouldnât be expected to generate new geometry?

I might be missing something but arenât fields and lists essentially the same? A list is going to be generated when evaluating the field at the geometry? Like in shaders, the noise texture exists in 3D space and we only see it where itâs evaluated intersecting the mesh. In that regard it seems like a solid system to work with some manual list fields for eg, making a selection of vertices 2, 5, and 13 for every mesh island. Rather than doing maths on the index field, youâd just analyze where the index field was equal to your list field and add those points to your selection.

I think it would be worthwhile to get a prototype for the fields up to a point where you can see how it works logistically for quite specific non-linear modelling processes.

1 Like

It looks like a good proposal to me.

How were you thinking of implementing fields? It kind of feels that instead of always having nodes retrieve values from their inputs, they now may have to call some function, and it will have to be clear what dependencies have to be passed to the function. Or am I completely misunderstamding?

Are you confident that the cases that donât need field evaluation will not suffer performancewise?

So thatâs basically a whole rework of everything we had so far?

## that mean that every work made with geonode will be obsolete soon??

all the infos/videos/nodegroups already available online? all will be broken ?

3.0 is the release to break things. Hopefully we wonât need more of these breaking changes in 3.1.

No, whatever happens there will be versioning to convert node trees for 3.0, unless that proves impossible, which is quite unlikely.

The idea of texture nodes influencing both geometry and shaders sounds great!

1 Like

I really like this proposal. I love fallofs in animation nodes, very useful for motion graphics, it would be great to have the same here. And the compatibility with sader nodes would be great.

1 Like

Sure hope so
it seems quite weird to change all structure of a newly implemented procedural system only 7 months after itâs release

3.0 is the release to break things. Hopefully we wonât need more of these breaking changes in 3.1.

Yeah but breaking things right after creating them directly imply that their gestation period was completely rushed out.

1 Like

It would be even harder later on, these important design decisions can happen more easily early on and I think itâs a good thing that Geometry Nodes now is still flexible enough for this level of changes.

14 Likes

Great Idea.Tracing attribute is no longer a troubleă
Now we need to improve our socket UI to separate the filed(or something else to draw the value just in the node tree panel?)