The initial cost for tessellationon landscapes (before adding any more geometry to the landscape yet) is too high that it makes the feature unusable.
Here is an empty material with no tessellation
Here is the result when it’s applied to a landscape
Now we enable the tessellation chekcbox, but there is no displacement map, no tessellation multiplier yet
Here is the result when it’s applied
The initial cost is about 10ms !!! and this depends on the landscape size so the cost changes from one landscape size to another. The example setup is on an 8K landscape.
Star Wars Battlefront makes extremely good use of distance based tessellation on terrain and nearby objects, which REALLY boosts the quality and depth of the meshes. Dice mentioned on their photogrammetry talk that it was a very trivial cost in terms of performance since they faded the effect out completely after a few meters. This isn’t possible with the current implementation in UE4, as even if you do fade the amount of tessellation by distance, there is still a huge initial cost of just having it on.
*Hey Deathrey,
We are aware of the performance issues with tessellated landscapes and have a report in to improve this feature (UE-14253). I have increased the community interest and added a link to this post to the report. We appreciate you taking the time to investigate and provide additional posts where users are reporting a similar issue.
Let me know if you have further questions or need additional assistance.
Thank you,
*
Do we have a time frame for this feature improvement? It’s not like we can use it now and just accept the performance hit, hoping you’ll have the fix before we launch. We have to do performance testing with each iteration of our game, and having this throwing us off isn’t good.
This topic doesn’t get that much attention because there simply aren’t enough forum users who are advanced enough to dig into these problems and post about them. Most UE4 forum users are happy to take the 10ms hit and assume it’s normal (and most UE4 forum users don’t actually ship games and therefore don’t realize the implications).
Having the issue as well. Abandoned landscape tessellation because -to my horror- it just doesn’t work properly,
Even with height-fade and distance-fade, there’s no justification for the tremendous performance-drop (which can’t be optimized on ‘average user’ cases it seems. Very much hoping there’s going to be a solution for this soon.
Tessellation on landscape is just too important to write off as ‘not doable’.
Yep, same here. It looked fantastic but just wasn’t usable. It couldn’t even be a feature for the highest settings as it’s obviously broken and the cost is way to high. Having tessellation completely fade out after just a few meters hardly helps either.
Same goes for tessellating any amount of foliage with LODS, performance tanks immediately despite having only a few meshes and not using tessellation on anything but LOD0.
Tessellation can make a tremendous difference in how a scene looks, it can really set it apart. I thought Epic would want to encourage people to use it as it would give them lots of free advertisement on how well UE games can look by having people use it and post pictures of their projects.
But alas, it seems to be one of those hush topics we pretend doesn’t exist.
Why implement it at all if you’re going to leave it unusable? Is it just so you can tick it off a list and have another sale argument?
+1 for this! It’s a real shame that it isn’t usable because it really does boost the quality of the terrain, especially when trying to increase detail. POM and other techniques don’t match the quality of tessellation without problems of their own, and it seems the best way for Epic to handle this is to figure out a nice distance based, performant solution! It will benefit everyone including their own projects! Battlefront is a perfect example of a game that runs exceptionally well but still has beautiful visuals and tessellation on all nearby terrain and trees/rocks etc. It’s a huge quality boost. If you look closely in Battlefront, the tessellation smoothly “slides” in as the player gets closer to the object- could be cool to have this automatically happen on landscape tessellation at a distance set by the user, like the awesome instanced foliage LOD fade in that has already been implemented by Epic!
Yes Xosh, thanks for picthing in (seriously) but we know all about that stuff. We didn’t come here to cry as a bunch of schoolkids about our precious tessellation not working at all. The point is that, even with mega optimization that makes one wonder what the hell is necessary to get 4 meters of good looking landscape geometry, it still impacts performance to such a demanding level, it’s just not a valid strategy for anything but the odd portfolio piece.
Hey guys, they are some technique to improve the tessallation. I personally fade it with the camera position. I have really better result by using this technique. It can be use for a many object.
Here a screenshot of mine.
Take time to play with the value, I suggest you to put the offset closer to 0. I know that doesn’t completly resolve the issue, but can make the tessallation on landscape a lot more usable. I wish this can help you and have a good day!
Thanks but it does not solve the issue, it just make it a little less performance heavy. It is not enough to be production ready solution, unfortunately
There is a bug report that deals with assessing landscape tessellation performance as we are aware of the impact it has on projects. For tracking purposes the bug report entered is UE-14253. I do not have a time frame as to when this issue will be resolved as it is currently backlogged at the moment.
For now, I would do your best to minimize the use of tessellation, and use the known distance based tessellation solutions to reduce the impact on performance.
Update: We are still investigating Landscape Tessellation and performance. We did submit a fix which resolved some of the performance issue in regards to this issue, but there are still some lingering performance impacts we are currently investigating. For reference, I have entered a new bug which can be found below.
Not at all. This is something we are looking into, but currently resources are dedicated elsewhere. Also, this fix is not something that is going to be super simple, so it will take a deeper investigation as to what needs to change to help improve performance.
