What's polycount preferable in modern games? (Low and hi poly vs mid poly)

Lemenus

I heard that in modern games mid poly modelling (mid poly + tiny tile texture atlases, anyways...) is much more preferable than classic low poly with hi poly bake. Is that right? And if so - then how hard high polycount hit performance and how much it matters to keep it low today?

Alex_J

First make the asset look like how it's supposed to. Then optimize it as much as you need to.

Andreicus

Nowadays you can achieve the look that you want without worrying about tris limit. This doesn't mean that you should model high poly only like you would do for a still rendering. It means that you should remove unnecessary edges that don't contribute to the shape and also optimize it even further if it's a background object not close to the camera.

Of course, i'm talking about PC and console. For mobile it's a different story.

This is a graph to get an idea of the general tris count for various models. It goes up to the PS4 gen: https://docs.google.com/spreadsheets/d/1xJmSuBZbdMOlIcdXmO9hu46GJ7-wIiWCclJwimIOJa4/htmlview

Benjammin

Sounds like you maybe heard something out of context . How an asset is made is going to depend on a variety of factors, including art style, what it is/how it is used, how much time you have to make it, etc.
High and low polycount are very relative terms, so its like Alex said above. 

radiancef0rge

Lemenus said:

I heard that in modern games mid poly modelling (mid poly + tiny tile texture atlases, anyways...) is much more preferable than classic low poly with hi poly bake. Is that right? And if so - then how hard high polycount hit performance and how much it matters to keep it low today?

i dont know if i would characterize it as 'preferable' but there is a case to be made that polys are easy to crunch and small tiling textures are fast to access.  

performance is relative, so its really impossible to answer your second question with a definitive answer.  meaning, if you have low triangle count but large textures and complex shaders, that can be worse than lots of triangles.  but these days hardware is very good at rendering triangles. 

i can't provide you with a triangle count but you should never have sub-pixel triangles.  (lods are key)

edit: a word

radiancef0rge

Alex Javor said:

First make the asset look like how it's supposed to. Then optimize it as much as you need to.

eh, i disagree.  pretty hard to walk back a prop thats been modeled or textured at 3 or 4x what it should have been budgeted for.  

gnoop

it depends on subject a lot.  Character , props, environment etc.   Whether you model/surface is going to use real time displacement/ tesselation.    

I would say model like it's lod_02  first   and refine the silhouettes later with more geometry  if necessary

Alex_J

radiancef0rge said:

Alex Javor said:

First make the asset look like how it's supposed to. Then optimize it as much as you need to.

eh, i disagree.  pretty hard to walk back a prop thats been modeled or textured at 3 or 4x what it should have been budgeted for.  

Yeah, it is a pretty vague and generic answer. But when you are just starting out and have no clue how to scope from the get-go, it's good mindset to have.

Usually what I find is this: First I need to jsut make the peice of art, get it in engine, and only then can I figure out exactly what it needs to be. Both in visual terms and what the optimization requirements will be.

If I spend energy trying to optimize it from the onset, it's always wasted time and effort. Because I just don't know until the one thing is among the many things in engine. This becomes especially so when it's stuff like props or vegetation -- stuff that can all get combined into atlas or whatever. If I try to organize that atlas from the onset... heartache. Because I am going to change the models a thousand times before the end.

I think even with something like a hero character it's the same thing. No point worrying about the topology and material scheme, because you are going to change that model many times. Best to leave the final optimization measures until the very end.

I am sure somebody with 10-15 years of experience can work with less iterations to arrive at the best result. But I doubt there is games being made by anybody in which the first pass art goes straight to publication. So usually I make something like five times. Each iteration I am trying to improve the art visually, but also figuring out what the optimal optimizations are going to be. And usually this has to take the whole picture into account. Maybe if you got a leader who figures that stuff out they just tell you what to do... but somebody has to have the full picture in mind.

Just my experience, which is limited. But a lot of more experienced people say similar things (of course, who do I learn from besides my own experinece.)

Lemenus

radiancef0rge said:

i can't provide you with a triangle count but you should never have sub-pixel triangles.  (lods are key)

What's sub-pixel triangles? Is it about textures or models?

Ghogiel

Lemenus said:

radiancef0rge said:

i can't provide you with a triangle count but you should never have sub-pixel triangles.  (lods are key)

What's sub-pixel triangles? Is it about textures or models?

I think the pixel being refered to there is screen pixels not texels. Having tons of triangles that are at times going to be crammed into a few screen pixels is far from ideal. ie some very detailed object with lots of tri that can be seen from a far distance, Good LODs go a very long way to reduce these dense hotspot areas on the screen as well as the total triangle count per frame.

poopipe

Lemenus said:

radiancef0rge said:

i can't provide you with a triangle count but you should never have sub-pixel triangles.  (lods are key)

What's sub-pixel triangles? Is it about textures or models?

triangles that are smaller than a screen pixel are not only pointless but also cause seemingly disproportionate inefficiencies when rendering. 

In the simplest terms, every triangle that lives in a pixel causes it to be processed again.   (this is a gross oversimplification before anyone picks me up on it) 

Fwiw @Alex Javor  makes a very solid point.  You should make quick blockouts, place them in game and determine what level of detail the object actually needs before committing to a final model.   Quite apart from the optimisation issues,  how do you know how big it should be if you haven't looked at it from the player's perspective?