Sure: D15445: Fix T99667: Regression in Delete Geometry node

Although T99664: Delete Geometry node face mode behaves unexpectedly when used after Join Geometry shows a bigger problem, I think this specific regression is caused by this check here:

In T99664#1389633, @Jacques Lucke (JacquesLucke) wrote:

Hm seems reasonable. It feels like it would be better to split up Separation and Deletion a bit more then, and maybe just share some common functions.
It feels wrong to output these extra elements in the Separate node.

I think it makes sense to add them, since this way, when you join the two outputs, you get back the original geometry (except maybe at the boundary). Otherwise, part of the geometry just disappears for seemingly no reason.

I think this is a bug, but it will require some big changes to the delete geometry node to fix.

• Fix wrong node name

If I'm not mistaken, the node is actually slightly transparent, and because the background is dark, the node color appears darker as well.

This bug seems to actually not be present anymore in latest master, so I'll just close this as resolved.

In T93663#1342061, @Hans Goudey (HooglyBoogly) wrote:

@Wannes Malfait (Wannes), I believe part of this problem is because BKE_mesh_new_nomain is used instead of BKE_mesh_new_nomain_from_template, which copies the materials/slots.

Another is probably that face attributes are transferred to vertices, but that isn't possible with material indices. I guess there's probably some way to transfer face attributes to new faces? Though maybe that exposes some arbitrary decisions in the edge cases.

In T94290#1294714, @Shrey Aggarwal (shrey_agg) wrote:

Is this issue solved?

Can't you just use the Alt+⇧+LMB shortcut to connect to the output? I think the o shortcut was disabled for geometry nodes because it provides no extra functionality over Alt+⇧+LMB, and there is no "output node" for geometry nodes, just group output nodes.

I can't seem to reproduce this. What shows up in the SHIFT + S menu? The error message says Mesh Cone which shouldn't be in the menu, so did that show up in the menu or did you select another node? Is there a specific node which you have to select/change to for it to work?

Can't reproduce here either:

**System Information**
Operating system: Windows-10-10.0.19043-SP0 64 Bits
Graphics card: NVIDIA GeForce GTX 1050/PCIe/SSE2 NVIDIA Corporation 4.5.0 NVIDIA 472.39

Using latest build (rB61aaeb3745ad)

• Better comment

If you're looking for the indices overlay in edit mode then you need to go to Preferences > Interface > Display > Developer Extras and turn that on. Then you can find the button in the overlay menu in edit mode. This does seem different from the screenshot you're showing though.

Can confirm. I don't know if this is expected/unavoidable behaviour. Could any of the developers confirm this?

Did you mean the FBX file format? I don't know what you mean with FBH. The issue could just be something to do with floating point errors but I'm not sure.

Could confirm the problem, but it does not seem to have worked in version: 3.1.0 Alpha, branch: master, commit date: 2022-01-23 19:27, hash: rBabf30007abda even though the bug report claims it does.

I was not able to reproduce this. I just rendered the default scene and opened the compositing tab. There I checked Use Nodes added a color balance node and a viewer node. Changing any of the values of the color balance node updated the backdrop.

You can still see which is which by hovering over it, and reading the tooltip. But it is quite annoying to not be able to see the difference at a glance. So I will leave it to the developers to decide if it is something that should be fixed.

I tried it again, because I realised I was rendering with CPU + GPU. When I render with GPU only I can reproduce the same problems. Sorry for the confusion.

I could not reproduce this on my laptop with NVIDIA GeForce GTX 1050 and 2GB VRAM. The VRAM usage was consistent throughout the 400 frames in the attached blend file.

The code works well! Just some picky comments about code style and grammar.

I have looked into it a little bit closer, and I have found the problem (or one of the problems). Basically the current code doesn't detect that the central vertex here is "non-manifold":

Yeah I saw this report and also tried simplifying it a little bit. I think the problem is that there are ngons which have more than 2 edges connected to a single vertex. I did not know that was possible. I tried creating the weirdest types of geometry for my tests, but apparently not weird enough. The current code assumes that each ngon has exactly 1 or 2 edges connected to a vertex, which causes the crash. The assert was there to ensure there was not some weird edge case that was missed for sorting, like this.

No, the capture attribute node won't change anything. In this case you either have to subdivide the plane a lot (which still won't give great results) or just use one node tree.

This is not a bug. Let me explain why:

• When you use one modifier, the random values are calculated for each of the randomly distributed points. The "field" is evaluated only at the Instance on points node.
• When you use two modifiers, the first modifier doesn't return a field but an evaluated field which is stored as an attribute. This field gets evaluated on the "Scatter Plane" and stored on the vertices of the plane (in this case the four corners). At this point there is no longer any field, just some values per vertex on the mesh. When it is used in the second modifier the values at the four corners are interpolated to the distributed points, which is then used by the instance on points modifier to pick the instance.

Can confirm and reproduce in new blend file. Just make sure you turn off 'Outline selected' in the overlays menu to see the effect.

At the moment the documentation just references the shader documentation of the node. This is because there is no real difference in the way the shader and geometry nodes versions work. The same thing is done for other similar nodes like the math node and the noise texture nodes.

I think the change would delete the vertices and edges in this case:

Although it looked like a good change to me at first, I think that it is sadly not as simple as this.
The problem is that this changes the current behaviour of the node. We can only delete everything
if everything is selected. The selection could be true for all the faces, but that doesn't mean that all
edges and all vertices are in the selection. With this change the whole mesh would get deleted even
if some parts aren't "selected". This change only works for the case of domain: ATTR_DOMAIN_POINT
and mode: GEO_NODE_DELETE_GEOMETRY_MODE_ALL.

Thanks for the report.

• Split code into new function and clean up comments
• Cleanup: comments and move check to separate function

If keep shape is true and the extra point is added, does it still get set to cyclic=true as well?

Thank you for your report. Just adding the error message is most likely not going to be enough to solve the issue.
Please provide:

• either a screenshot of a node setup along with steps on how to get the error there.
• or a blend file along with steps to get the error.
• or exact steps to get to the situation form the default blend file.

• Split code into new function and clean up comments

Assigning materials in geometry nodes to the mesh from dual mesh seems to broken.

I don't see this as a bug, more like a feature request in my opinion. When you make a node group, the only thing that's consistent across copies of the node groups, is what's inside the node group. So in that sense the most logical way to make the node group socket names is to take the socket names from what's inside the node group and not from what's outside the node group at the time of creation. Maybe I misunderstood the problem.

I think this should also be added to the separate geometry node. Normally it should be easy to write the code in such a way that it works for both nodes (similar to the current code).

• Remove outdated comment

I also changed the #define in BKE_node.h to the correct number.
It was much higher to prevent merge confilicts will working on
the patch.

• Face attribute propagation for boundary vertices

• Fix corner attribute interpolating
• Cleanup: comments and switch case