Hi PaQ, so that's basically that same "low polygon" density kind of problem that I was describing earlier - the fillet piece there gets a low number of divisions along one direction despite the angle setting of 6 because the angle setting is based purely off of curvature and will not divide things that are very shallowly curved but long.

What you get in that case is an n-gon that's pretty non-planar and then introducing a centroid point in the middle of that does not work very well.

The real fix is probably more like adding in some additional refinement mechanism that kicks in automatically to avoid making a pretty non-planar n-gon to begin with. Basically if your n-gons are not torqued then centroid triangulation would not cause that kind of problem.

Right now for that kind of geometry you have to use those various additional settings like "Divide larger than" to force that fillet to get subdivided more in the long direction than what is currently happening by the angle measurement alone. On shapes that are shallowly curved like that you basically will get low-poly artifacts even with a high angle setting.

Non-centroid triangulation does tend to not mess things up on non-planar n-gons as much as centroid ones do, but it's still better for those types of things to get subdivided more overall so that the final n-gons that get triangulated are not so non-planar in the first place.

But part of the difficulty though is that the non-planarity of the n-gons isn't really there during the subdivision process itself which works only on "underlying surfaces". In a case like you have here the non-planar n-gon is formed only after the trimming boundaries are introduced onto the surface, that happens after the subdivision process is all complete so it's not very easy for the subdivider to take that in particular into account at least with the current architecture. But some other metric that would look for things that were fairly long but only shallowly curved and divides those surfaces additionally would help, it's just that some of these things like what exactly "fairly long" means in some algorithmic way that applies to any kind of model is kind of difficult to quantify. And also it's a problem for actual desired low poly output if something kicks in with too many subdivisions totally automatically.


Anyway, to make a long story short - in your model here your fillets are getting low-poly artifacts on them because they are not divided enough in their long directions, they need to get divided more to make the n-gons less warped and once they are divided more either centroid or non-centroid triangulation will work ok on them. Right now you have to force such subdivisions yourself by using some of those other settings, I do not yet have an automatic way of ensuring that this kind of situation gets automatically subdivided.

- Michael