Need advice from experience devs on tessellation and the megascans library

pixelquaternion

Hi guy's,

After reading for almost a complete day i am still at lost on tessellation performance issues since many threads on various forums seem to indicate that tessellation is still a major issue in UE4.

I am currently  starting to work on my terrain and i was tempted to use the megascans library but as tessellation seem to be the only way to get good looking asset with this library then the UE4 major performance issues is telling me that it is a no go currently.

My understanding so far is that using every megascans library asset on many game objects would quickly put any decent pc on it's knees!

My game is mainly a single player with a simple pvp online component targeting pc and probably consoles later on so i wonder what would be the best approach to keep good performance with best visual?

Obscura

As always, you can just tick it on and see how does it perform on certain hardwares. It is still not that regular to use tessellation in games, but there are some examples here and there, definitely. They use it on very few assets, and only where it makes sense. Landscape already uses tessellation though, its used for its lods, so technically you would have double tessellation if you would use it on a landscape, lol.

pixelquaternion

Obscura said:

As always, you can just tick it on and see how does it perform on certain hardwares. It is still not that regular to use tessellation in games, but there are some examples here and there, definitely. They use it on very few assets, and only where it makes sense. Landscape already uses tessellation though, its used for its lods, so technically you would have double tessellation if you would use it on a landscape, lol.

Hi Obscura,
So basically i was right when i mention using many assets from megascans would seriously impact the performance! Like you say i will have to test a few thing to see how fps get hit.

Also it's seem unclear if the tessellation issue has been solved in UE4 even after reading a few posts on this?

poopipe

It's solved in the sense that generating and rendering a shit load of geometry is expensive so you try to avoid it where possible..
The last place I worked had a tool for baking tessellation down to meshes because that was cheaper than doing it at runtime (not ue4) 

You could try parallax occlusion mapping if it's for relatively small scale (rubble etc),  that's not cheap but it does a very good job of breakiing up silhouette,  supports self shadowing and you can at least disable bits of the effect to make it go faster.

Anything more substantial than that is going to require props or detail objects if you want it to run right. 

pixelquaternion

poopipe said:

It's solved in the sense that generating and rendering a shit load of geometry is expensive so you try to avoid it where possible..
The last place I worked had a tool for baking tessellation down to meshes because that was cheaper than doing it at runtime (not ue4) 

You could try parallax occlusion mapping if it's for relatively small scale (rubble etc),  that's not cheap but it does a very good job of breakiing up silhouette,  supports self shadowing and you can at least disable bits of the effect to make it go faster.

Anything more substantial than that is going to require props or detail objects if you want it to run right. 

Thank for clarifying the matter since when you currently look at the trend it's seem like everybody are tossing tessellation at everything in game but in reality the tech is not yet there to use it and that's a shame since some assets get sale with this false impression that you can use it as you wish.

poopipe

I share the same concerns about a lot of substance materials I see on artstation etc. 
In most game production  situations  you need to cover quite large areas (4m plus) with a tileable texture so the 30cm square, ultra high detail, tessellated materials everyone seems so keen on producing are of no practical use as they'll tile like crap at that scale and 90% of the information will be lost.

That's not to say megascans are useless,  you can certainly use the results to generate props and texture sets as long as you're  aware you'll have to put some manual work into supporting geometry to get it to run right and look amazing. 

pixelquaternion

poopipe said:

I share the same concerns about a lot of substance materials I see on artstation etc. 
In most game production  situations  you need to cover quite large areas (4m plus) with a tileable texture so the 30cm square, ultra high detail, tessellated materials everyone seems so keen on producing are of no practical use as they'll tile like crap at that scale and 90% of the information will be lost.

That's not to say megascans are useless,  you can certainly use the results to generate props and texture sets as long as you're  aware you'll have to put some manual work into supporting geometry to get it to run right and look amazing. 

Agree and too many tutorials are wrongfully  leading artists to believe they can put zillions of nodes in SD and use tessellation all over the place without repercussion!

Even with today video cards and CPU we still need to work almost like if we were 5 years behind to make sure performance will not be compromised.

I don't mind tutorials showing zillions nodes and bad optimization if they are made to make people learn the various nodes but they should start their tutorials with warning about not using this workflow for a real world pipeline.

Regarding megascans i think they should lower the price a bit and then some serious developers could use them in a more performance wise way.

Mant1k0re

Ok, I've got no idea what you're getting at about Megascan assets not being usable? Did you mean 3D meshes? Because textures have a reasonable selection of 4m surface that are working very well with landscapes in Unreal in my experience. I'm just confused here.

Also who cares about messy unoptimized graphs if you're going to generate textures from it in the end? I don't think anyone uses live substances in engine for games today.

Obscura

I also didnt get the comment about the substance nodes... even if you import and work with the raw substance inside unreal, instead of exported textures, the substance instance practically bakes you a texture. It is not calculated in realtime at runtime. It wouldnt make sense. Having said that, it really doesnt matter how big is your substance graph, as long as you can work on an acceptable speed with it, inside substance. Though I absolutely agree about the artists all showing tessellated materials around artstation. Same about "game ready" meshes. There are countless amount of examples of "medium poly" but very wasteful and unoptimized meshes in the portfolios out there which is very misleading in my opinion.

pixelquaternion

Obscura said:

I also didnt get the comment about the substance nodes... even if you import and work with the raw substance inside unreal, instead of exported textures, the substance instance practically bakes you a texture. It is not calculated in realtime at runtime. It wouldnt make sense. Having said that, it really doesnt matter how big is your substance graph, as long as you can work on an acceptable speed with it, inside substance. Though I absolutely agree about the artists all showing tessellated materials around artstation. Same about "game ready" meshes. There are countless amount of examples of "medium poly" but very wasteful and unoptimized meshes in the portfolios out there which is very misleading in my opinion.

I was mainly referring about a few tutorials where people said they were using huge graph for real time fx from SD in UE4. But anyway the main point is that there is a clear misleading trend on many website and video channels.