help with poly-reduction strategy - Japanese castle

umbrellabuddha

Hello, I've been working on a Japanese mountain-top castle scene in Unity and I wanted to get your guys's advice on the castle.. who's geometry alone currently exceeds 50k tris... and it's all in the roof detail. Let me show you my reference:




and the current castle geo:



My biggest concern is KEEPING the interesting silhouette of the edges of the roof sections. For now I basically just used a single 6-sided cylinder(18 tris each piece) with an interlocking convex-shaped piece(12 tris each piece), and a simple cube-shaped piece underneath(8 tris each piece)all of which I duplicated, transformed and carefully positioned across all the roofs. All 3 of the pieces have the same recycled UVS per each instance.



I considered having the inner parts of the roofs a simple normal-mapped plane as I illustrated, but it wouldn't actually reduce anything anyway since each cylinder has only one span in length.



I really want that detail from the reference photos... am I going about it all wrong, or can I somehow get away with what I have? I really appreciate your input!

Hayden Zammit

I'm working on something similar, and I just normal mapped all the cylinders onto the roof. I can't think of a way you'd be able to keep all those cylinders on there for detail and still keep it at a good tri count.

I doubt the player would even care or notice. Especially for the ones closer to the top of the castle.

Still, maybe one of the more veteran artists on here might know a good way.

keres

You could bust up the model into pieces and let the occlusion culling take care of performance. 50,000 for an entire building such as that should be acceptable, in my opinion, as it should be the majority of your scene (by that I mean what the player is in at the time.)

I suggest to bust it up into pieces, then LOD like crazy so it doesn't draw unnecessary amounts of geometry at farther distances. That model does not look like it can be remade with less geometry and the same amount of detail.

Good luck, and merry Christmas.
-Tom

umbrellabuddha

I didn't think to cut it into LODS that sounds good. I found places I can use normal mapped planes that don't have their own silhouette also. Thanks for suggestions!

Hayden Zammit

What the heck? Is 50k tris really okay for a building these days?

EMC3D

Hero building maybe?

Snader

My advice? don't try to model all the separate pieces.


12 pieces, so about 15 triangles per piece, instead of 30.
Top one has a nicer silhouette, bottom one is easier to make modular (due to the top one having a big box for the inbetweens).

Acumen

Also if we talk about the building being mostly visible from the ground, you could also weld the end vertices of Snader's example so you have the modeled detail on the edge of the roof that kinda blends into the normal mapped version on the inner parts, if you know what I mean.
Maybe have a look at Assassin's Creed how they approach roofs, since it's an quite current-gen-game which features roof access as a prominent gameplay feature. Yes it's not in the japanese-castle scenario (yet ) but some shingles and other detailled stuff sure should have similar approaches or level of detail comparable to your project.

What would really interest me, is for what kind of game this building would be used.
Apart from that I wouldn't call the silhouette interesting. It looks reeeeeaaally repetetive and boring, imo.
I think I'd choose to not model out every single piece of detail on that roof, but instead invest some polygons in creating interesting flows and shapes that support the silhouette much more with adding imperfection and whatnot.

nullfed

As Hayden suggested, its down to what the player would actually see, surely? maintaining all of the detail across the whole roof seems wasteful if the game camera is an FPS and the player may catch a glimpse of it as he enters the building.
Apart from some key areas like that (although I would be much more interested in seeing some of the carved faces and medallion details rather than the corrugation effect) you could get a lot of the details using normal maps with alphatest fringes.