Medieval Castle and Geometry Nodes Workflow

[Gorgious]

I started a new castle project. Similarly to my previous one I’d like to document my workflow which is primarily procedural and Geometry Nodes-based and showcase the work in progress for anyone interested.

The project is a castle surrounded by and towering over a river.

As always it’s an iterative process and things are expected to change pretty drastically. For now I’m working on improving my wall Geometry Nodes tree which I still feel is a bit lackluster in terms or realism, especially on the top with the crenels.

I estimate this will take around 1 or 2 months to complete. I’ll release everything I can under CC0 license once completed.

22.03.21

22.03.22

22.03.24 (Already 3,7 M tris, yay !)

[Gorgious]

A new wooden scaffolding I’m trying to design. Still a few kinks to iron out but it’s shaping up.

Blender : WIP wooden scaffolding using Geometry Nodes

[sozap]

Awesome !!

[Gorgious]

Thanks that means a lot !

Here’s a snippet of small progress

Adding some features to the wooden scaffolding

[sozap]

So cool !

These scaffolding looks really great !
In the demo I really like the result at first, at some point it get a bit too bendy, maybe there is a way to resample and force straight lines.
It was on my todo to try to make a castle , should be really fun to do and looking at your work it raise the bar !
I really looking forward how it will turn out !

[Gorgious]

Yup I cranked the “destruction” slider a bit too far you’re right ^^

Procedural damage / aging is really rewarding in terms of asset variation but also hard to get right… I take a lot of inspiration from the projects you recently showcased

[sozap]

Cool !
Yes it’s hard to get right ! it’s either highly distorted because it’s collapsing, subtly distorted because it’s hand made, or made in a hurry. Aging distort stuff also, but it’s subtle. When it’s too much distorted for no reason it gives the impression that it’s made by drunk people. You can also add some distortions to get a cartoony look, but it’s a bit different, it’s more in the proportions than distortions.
And of course it changes depending on how far/close we are.
But people who build the thing wanted desperately that everything look straight and clean Somehow they fail a bit …

If I look for realism I try to add some human model next to what I do, it helps a lot to get the right amount of distortion and keeping it believable.
You can also try to make some distant and close shots to get a good balance.

All that said, it’s really great what you’ve done, I just noticed that, like me, you like distortion
I’m curious about how you made the cross sections :

Can you share a bit how you get these ?

keep up the good work , it’s awesome to see all this being build bit by bit !

[Gorgious]

Hehe I’m glad you picked up on that because I scratched my head for a bit figuring that out. In the end I came up with a hacky solution, but that seems to do the trick for now. I’ll certainly redo it more cleanly at some point, but here’s the workflow :

I begin with a simple curve

Then I offset it along its normal in the local right and left directions, making sure to factor in the control points radius

Then I instance in on a Mesh line along the global Z axis.

I resample the curve every few meters.

I then instance a Mesh line with 2 points on each resampled control point, setting the second vertex at an absurdly low Z coordinate to be able to filter it in later when I change its position.

Then I compute the position of the next control point by index, and set the position of the second vertex of each of the mesh line instances (now realized) to the position of the next CP which I computed the position with the Field at Index node.

It basically recreates the curve with a small caveat, since it’s a single curve with multiple splines, the last CP of each spline is connected to the first CP of the next spline. And of course, the last CP can’t compute the position of the next CP so it defaults to object origin.

Then I offset every selected vertex by the storey height :

Here’s that part’s node tree snippet for reference

Until then I think it’s a pretty neat procedure. The hacky part in the end was how to remove the very long stretches that shouldn’t be there.

Since these unwanted lines are longer than the diagonals I’m trying to get, I figured after spending too much time trying to juggle between point indices that I could delete them based on their length

Then I copy/pasted the same procedure, but taking the position at the previous index rather than the next one, and merged both sets of splines

And after a Curve to Mesh node

Here’s the final output after deleting all geometry below Z = 0

[sozap]

Damn !! That’s great !!
Thanks a lot for sharing your process, it’s quite inspiring !
I was thinking you may find a simple solution for that, but anyway ! that’s really cool ! well done !

Something simpler could be, extruding the curves, then extrude each faces to make a “poked face” by scaling at 0 the extrusion, that should give a similar pattern, after that you need to get rid of unwanted geo to keep only the diagonals. That should be easy because it’s part of the extrude selection. Not sure it will work in practice, but your solution is quite clever !

I’ve got another question for you :
Right now I’m having issues to align the curve tilt according to another mesh normals :

It’s the red parts in the capture, they are based on some curves.
I use a dot product on the mesh normal (that is a cubic shape not seen in the capture) and the curve normal and plug that into a arcosine to get the angle.
As you see, it can work but it’s failing at some angles, and if I start to distort the curves it turns into a big mess.
Have you find a solution to deals with situation like that ?

[joseph (J)]

That is genius, well done figuring out that solution! I’m definitely bookmarking that!

[Gorgious]

Oh man I was sure I would end up finding something better. Thanks for the pointers !

At first I was wondering how to poke the faces like you suggested but I remembered something I read on BSE. Extrude faces and scale the top face to 0 to get poked faces.

So to reiterate :

1. Resample curves
2. Curve to mesh, but you can’t just use the direction inside the Curve Line node because it will be extruded along the local normal (or tangent ? didn’t check)

(Edit : the Curve line direction should be (0, -1, 0) but the problem doesn’t change)

So my solution was to instance a curve with 2 duplicated control points, then offset every other point using a modulo with the index.

3. Poke the mesh using an Extrude and Scale Elements with 0 scale

Tha was a piece of cake ! Note I used a slight offset in the Extrude because I use the edge angles to filter out the edges that are not diagonals later.

4. Now we need to heat up the brain cells a bit to filter in only the diagonal edges. The selection is threefold.

First we remove the “horizontal” edges by checking the unsinged edge angle (honestly that was a shot in the dark, I can’t really explain why it works)

Then we get rid of the vertical lines by checking that the two verts of an edge’s X and Y components are the same

Then we get rid of the edges that are located below the local Z = 0, that will aleviate some computation in the boolean later on.

5. And now we split the edges to get individual splines later on, convert them to curves, resample then using the LOD input slider, and transform it to mesh with a quadrilateral curve !

Pretty much the same result as earlier, but with way less overhead in nodes.

Here’s the updated node tree

However I did notice that this performs worse than the first solution i nterms of computation, roughly * 2 time, but it’s nowhere near the bottleneck of this tree so I’ll stick with it since it’s so much simpler !

Concerning your question, that’s really funny, I think I’m stuck on a very similar thing ! Vector maths is not my forte so I usually play with every setting until it works but in my case it didn’t.

I was trying to instance ladders which are themselves a node tree I’ve showcased earlier.

It looks Ok-ish when the curve is along the X axis

But pretty bad when it turns… i can’t figure out the correct rotation

And I can’t use the Align Euler to Vector trick because the ladders are instanced using curve lines and a curve line doesn’t really have a notion of rotation… I’m trying to play with the tilt but can’t get my head around it for now !

So I’m afraid you seem to be way beyond my level on this topic. I’ll poke around though and see what takes !

@joseph cheers, happy to share ! Don’t hesitate to ask questions or maybe show what you’re working on, inspiration goes both ways

[joseph (J)]

I haven’t done much with geo nodes recently, to be honest, but with an invitation like that, I’m going to have to whip something up to apply what I’ve learned from this thread

[zeroskilz]

@Gorgious

curve line doesn’t really have a notion of rotation

Not really, curves respect the notion of rotation… tilt is just rotation along the tangent on-top-of that rotation:

e.g. here just instancing a line onto the surface and rotating the line using regular align Euler to Vector…

…curve-to-mesh after this rotation will result in the expected rotation (even though profile curves have constant offset)

So, it should be possible to align those ladders if you can capture the underlying geometry’s normal.

Great work, and good luck!

[sozap]

Hey !

Here is my take on getting the poked lines :

I thought there could be a less hackish way, it’s probably not the optimal in term of performance , but it’s simple to setup : I remove edges by proximity with the original mesh.

And ok for the curve tilt ! I suck at math since school, it’s hard to get back at it, isn’t it ?

If the ladders are instanced it’s indeed simpler as @zeroskilz pointed out, it’s like with regular instances, the curve tilt part shouldn’t comes to play.

But if you instance some curves and build the ladder from that , yes you need to set the tilt.
In my procedural house I added ladders, it’s some curves and I took the end point and snap it to the walls. And from that curve I build the ladder.
In that case I was lucky, because the ladder is oriented toward the wall the ‘tilt’ was set right without other issues. But if it’s straight I get the same issues as you have.

BTW @zeroskilz , you’ve got the jedi level with your ZCurve tilt node ! do you have some hints on how to fix this :

I think I can manage the negative rotation by trying random stuff, but once I activate some distortions on the curves everythings go wild :

Here is the .blend : Curve Tilt.blend (151.3 KB)
In the meantime, I’ll try every nodes combinations and see if that fix the issue ( no that doesn’t sound like a desperate call for help…)

[zeroskilz]

@sozap

you’ve got the jedi level with your ZCurve tilt node

I don’t know about that… I feel stupid 99% of the time… and its rare that I do something useful the other 1% of the time when I don’t feel like a complete moron.

I see now the challenge when your curves and instances get generated from other things its not so easy to rotate them… hence the tilt requirement… The arccosine probably isn’t enough - guessing the arcTan2 will be required, so I don’t have a quick answer for you.

About to go to sleep and busy tomorrow, but will attempt something when I get a chance.

[sozap]

hahaha maybe Luke Skywalker is effective 1% of the time also but that’s enough to save the day !
Thanks a lot for the quick answer, I need to learn more about ArcTan2 !
I’ll post an update if I find something !

[zeroskilz]

Curve Tilt1.blend (98.9 KB)

K, so, its got to do with singularities (when math goes to infinity)… it will all make sense when you consider that under default conditions (0 tilt), no single-segment-curve normal ever has a Z component, and a Z-up line-segment will have a default normal of (1,0,0).

… the solution is to “re-frame” the desired normal onto the plane described by the curve normal and the curve tangent-normal (the cross product (will always be a unit-vector orthogonal to the tangent and curve normal)).

… dot-products give the projections of the desired normal onto these two components for the rize/run parts of the arcTan2 (no magic, that’s all it does, you give it rize/run and it spits out angle - better than arcSin, arcCos and arcTan which has a operational range of -pi/2 to pi/2 whereas arcTan2 has operational range of -pi to pi)

Limitations of this method is that when the desired normal points along the curve tangent, you’ll get wild tilts again (can’t be avoided).

Thanks for this exercise - I think it may lead to other solutions that have been evading me.

Good luck.

[sozap]

Damn, you’re so strong !
I probably need to read it a few time again and play with the setup, but it starts to make sense !
And cheery on top, now I understand what ArcTan2 is doing !
Thanks a lot , this is so much useful !! You’re truly a 3D jedi !

[zeroskilz]

@sozap

You’re truly a 3D jedi !

Jeez, stop saying that, I prefer it when ppl’s expectations are low.

Sorry if I’m not explaining stuff good, feel free to ask if stuff don’t make sense.

[Gorgious]

Haha what a perfect example of Cunningham’s Law.

Thanks for schooling me and giving so much pointers to both of you, I’ll explore it asap and report back ! Cheers

@sozap I also noticed that tilt behaves very strangely with line curves that contain only two control points, but if you add one CP in the middle it becomes consistent and you can play with it.