What is a good amount of triangles/vertices/faces for a mesh body?

[TaleTitan]

Hi SL. I'm new.

What's a good range of triangles/vertices/faces for a custom mesh body?

I don't want to look like a blob to most people, lmao.

Thanks.

[Kyrah Abattoir]

As low as you can get away with.

[FinnfinnLost]

2 hours ago, Kyrah Abattoir said:

As low as you can get away with.

This.

I'd say stick below 10k. I'd recommend making a really low-poly version as your first "production-ready" version. Enable smooth shading and see if it resembles the shape you're going for. Rig it and see if it deforms properly.

With your basic topology in place, add more geometry and detail until you're happy or hit your maximum tri count. More subtle details can be added via normal maps without increasing the amount of geometry.

Are you using Blender? If so, I can share some personal experience that could help.  

[Gabriele Graves]

Just for reference does any one know what the SL system body is worth in triangles?

[Aquila Kytori]

1 hour ago, Gabriele Graves said:

Just for reference does any one know what the SL system body is worth in triangles?

For the complete avatar,  7,186 triangles :

 

Model from the  Machinimatrix web site,  Avatar Workbench-2.71.blend file.

https://blog.machinimatrix.org/avatar-workbench/

 

Head = 1,844 tris.

UpperBody = 3,688 tris.

LowerBody = 1,654 tris

Edited April 2, 2021 by Aquila Kytori

[Gabriele Graves]

Unless the SL system body is in fact badly designed and could be majorly improved on by having the angular areas smoothed out without increasing the triangle count, I wouldn't consider this amount of triangles sufficient any longer for a similar mesh body for my own uses.  I'm not a mesh expert so I don't know what is possible with that but assuming that it isn't really possible to do much better, it's one of those YMMV things because there are many people very happy still with the system body and it's limitations and there may well be many people happy with mesh bodies at the same level.  So if you want to create something of a similar quality and efficiency, this is would be perfect but there are few human mesh bodies on the grid that have this few triangles and while many bodies are definitely inefficient, it is unclear to me what is the absolute minimum number of triangles it would take to recreate something at the same level of quality of say the Maitreya Lara body by the most expert of mesh artists.

Edited April 2, 2021 by Gabriele Graves

[ChinRey]

1 hour ago, Gabriele Graves said:

Unless the SL system body is in fact badly designed and could be majorly improved on by having the angular areas smoothed out without increasing the triangle count, I wouldn't consider this amount of triangles sufficient any longer for a similar mesh body for my own uses.

As much as I support the idea of reducing the polycount, I agree. 30,000 for a modern mesh body, including head and everything would be excellent. 50,000 would still be good and far better than any existing mesh body/head combo I know of.

Then do some serious and intellingent simplification for the LoD models, just in case LL some day gets around to fixing the fitmesh LoD bugs. It might be possible to get the mid model down below the 7,186 with no noticeable distortions, certainly below 15,000, and the low and lowest could probably do well with less than 1,000.

Edit because I can't resist it:

In a virtual world where people stayed with the default 1 or 1.25 LoD factor, where there weren't any fitmesh LoD bugs and where content creators made good LoD models, it would be perfectly safe to load up an avatar with a quarter of a million tris worth of lovely details (although I can't imagine that much is ever actually needed). The reason is that there would never be that many avatars close enough to be render at high LoD anyway and 5,000-15,000 tri mid LoD would still be enough to not cause any noticeable distortion to the avatars a little bit further away.

Unfortunately Second Life is not that kind of virtual world.

Edited April 2, 2021 by ChinRey

[TaleTitan]

5 hours ago, FinnfinnLost said:

This.

I'd say stick below 10k. I'd recommend making a really low-poly version as your first "production-ready" version. Enable smooth shading and see if it resembles the shape you're going for. Rig it and see if it deforms properly.

With your basic topology in place, add more geometry and detail until you're happy or hit your maximum tri count. More subtle details can be added via normal maps without increasing the amount of geometry.

Are you using Blender? If so, I can share some personal experience that could help.  

Hey thanks Finn!

Yeah, I'm using Blender. I'm still a pretty basic user but I can make decent things. I want to make a real coolass armor for myself eventually, but I'd also look forwards to making stuff like giant butts in tiny shorts (and selling it to people).

Got Discord? Feel free to add me at TaleTitan#5366

Edited April 2, 2021 by TaleTitan

[Quarrel Kukulcan]

8 hours ago, Aquila Kytori said:

Head = 1,844 tris.

UpperBody = 3,688 tris.

LowerBody = 1,654 tris

What about the skirt? 😁

[Wulfie Reanimator]

The bar for improvement is pretty low when the current mainstream bodies are in the 300-750k range, not including the head.

[Gabriele Graves]

4 hours ago, Wulfie Reanimator said:

The bar for improvement is pretty low when the current mainstream bodies are in the 300-750k range, not including the head.

Not all are that bad.  Maitreya is 176K but your point still stands.  There is definitely scope for someone to step up and show us all how it is done for a mainstream body.

[FinnfinnLost]

1 hour ago, Gabriele Graves said:

Not all are that bad.  Maitreya is 176K but your point still stands.  There is definitely scope for someone to step up and show us all how it is done for a mainstream body.

... this is not acceptable. At all. I usually try to stay friendly and any newbie to 3D is going to produce way too complex models, I was the same. So, as a disclaimer: If you produce models like these, but are trying to improve, don't feel bad. I'm still working on it as well and am nowhere near the optimization skills I desire.

But this is a character model that the creator is generating no small amount of income from. If I purchase a model for a videogame, I can expect a certain quality. Yes, I've looked at the body and yes, it's a very pretty model. But this is ludicrous. She has about the same amount of detail to her as Adam Jensen from Deus Ex: Human Revolution (the full character, that is). And he's just shy of 26k.

On the bright side, you just motivated me to resume work on my female character mesh. Without arms, hands and hair, she's currently clocking in at 13k without optimization (read: baking finer details as a normal map). She still has a cute butt and rather elaborate boobs: https://paste.pics/C3CAT

I don't mean to brag, really, I don't. I just want to point out that if my FIRST HUMAN MESH EVER, and I'm not doing this for a living, achieves this tri count, there's no reason for someone who has a large customer base to produce something like that. Maybe I'm totally wrong though. If anyone can think of a reason, I'd be glad to hear it.

[Gabriele Graves]

I expect one of the reasons for the high poly count is the alpha cuts.  There are 208 cuts on the ML body which at eight faces per mesh object makes 26 objects that make up the body.  Obviously the head isn't included in this.  The main thing to understand about this though is that for a mainstream female body in SL, these kinds of cuts are essential for a large amount of users.
There have been spirited arguments about whether that is a good thing or bad thing but a thing it is and considering this body has more users than all the other mainstream female bodies and the creator knows their market, it actually makes it unlikely in this context for a comparable body to lack this feature and gain the same level of popularity.  The main reason is that a lot of particularly female clothing has complex requirements when it comes to the alpha masks, the effort and skill required to make alpha masks for many people is too much compared to the simplicity of clicking a few HUD sections.  So you simply cannot ignore this aspect.  One notable female body designer who did remove alpha cuts when bringing out their BoM bodies has sunk beneath the waves in terms of market share and as a consequence more and more clothing designers are dropping support for their bodies.

So this brings us to the aspects and considerations of mesh bodies that are unique to SL.  It makes comparisons with game characters invalid because for a game character you would design the clothes to fit perfectly and control every aspect of clothing, animation and movement.  You wouldn't even need alpha masks most likely, the clothing would be one piece per outfit because you don't have to deal with people wanting to mix and match outfit parts and you wouldn't have to deal with multiple creators and permissions.  There probably wouldn't be a limit of eight faces per mesh object - a case where the limitations designed to keep complexity down work against that goal when designers are forced to work around it to implement features that people need.  Most of all, from the videos I have seen of most modern video games, the characters do not seem as well finished as an SL mesh body avatar overall.  They all have some noticeable areas that haven't been smoothed, or textures that stretch badly from what I can see or optimisations on the clothing that make them appear more painted than 3D.  I'll admit I haven't seen any game videos more recent than about 6 months ago though.

Due to SL's less controlled, more dynamic environment and particularly the way animations and the way they affect rigged attachments (i.e. the mesh clothing), it brings along considerations you just wouldn't have to worry about in games.

It is easy to decry that this is because of the insane amount of triangles going into mesh avatars and that this shouldn't be happening.  I can accept that but I am also told that this can all be optimised without noticeable quality loss too so I am not sure what to believe there.  However the facts are facts, to gain mainstream popularity in the SL female body market it is going to take a body that cannot be distinguished in quality (and by quality I mean smoothness with curves that look like curves) from other bodies, retaining features people want and do all this whilst keeping the triangle count much lower than 176K.  All in all, for an equivalent body quality with these considerations, it would be interesting to see what would be the absolute minimum it would take, especially with regards to the alpha cuts.

Looking at your picture and I mean no offence because making a body is way beyond my talents so kudos to you for what you have achieved so far but from what I can see the bum on that model is way more angular than on my ML body.

Edited April 3, 2021 by Gabriele Graves

[FinnfinnLost]

I appreciate the elaboration. And yes, I can see how alpha cuts would increase the complexity.

HOWEVER: Stretching textures have little to do with the amount of geometry. I can stretch a triangle to oblivion and back and as long as I have a long texture with good UVs, it's going to look crisp. So you can ignore those stretching textures for the purpose of this discussion. I looked at Ellie from The Last of Us and all I see is lower resolution due to technical limitations, no stretching due to faulty UVs. This leaves features specific to SL. With the example Ellie, with her clocking in at ~30k polygons (with decent topology AT MOST 60k tris assuming none of these polygons are tris already), this would mean that for the added detail, which is negligible and could have been done using normal maps, and SL specific features you would need enough geometry to model her thrice.

Regarding attachments and animations... you rig clothing and attachments to the existing armature and fit them for a mesh body of your choice. This does not require additional geometry on the body unless the topology on areas with high deformation is severely screwed up. This would make any RPG with an equipment system playable only on the most cutting edge PCs, which would not include mine.

About my model, you may wish to read my description again. There is no optimization and no normal mapping going on. But I just attempted to throw raw amounts of geometry at her which would allow me to smooth the crap out of it. For the torso+legs+feet I would go from 4.8k to 5.5k. And if I wanted to smooth the whole body without giving any thought to optimization, I can actually subdivide the whole thing, which means turning every single quad into four ones, which would push it to ~20000: https://paste.pics/C3D6N

Let's assume for a moment that I decide to quit optimizing right now and add hands and arms. If I estimate this using the amount of geometry required for legs and feet, this would put the torso at ~6.5k, again, assuming no optimization of the mesh, and the whole body at, rounding up, 16k. Let's add hair for the sake of getting the whole thing to 20k without subdivision.

If I add a subdivision after that, I'd be at 80k for a whole naked body. A result which would get me slapped by any 3D artist for modern games. Can anyone confirm that, just to make it work with common SL features, I would need MORE THAN TWICE the amount of geometry on that character?

[Gabriele Graves]

42 minutes ago, FinnfinnLost said:

HOWEVER: Stretching textures have little to do with the amount of geometry.

OK, I took this effect to mean that the mesh didn't have enough triangles to stretch in a believable fashion and that was due to reducing the triangles of the body.

Quote

About my model, you may wish to read my description again. There is no optimization and no normal mapping going on. But I just attempted to throw raw amounts of geometry at her which would allow me to smooth the crap out of it. For the torso+legs+feet I would go from 4.8k to 5.5k. And if I wanted to smooth the whole body without giving any thought to optimization, I can actually subdivide the whole thing, which means turning every single quad into four ones, which would push it to ~20000: https://paste.pics/C3D6N

Let's assume for a moment that I decide to quit optimizing right now and add hands and arms. If I estimate this using the amount of geometry required for legs and feet, this would put the torso at ~6.5k, again, assuming no optimization of the mesh, and the whole body at, rounding up, 16k. Let's add hair for the sake of getting the whole thing to 20k without subdivision.

If I add a subdivision after that, I'd be at 80k for a whole naked body. A result which would get me slapped by any 3D artist for modern games. Can anyone confirm that, just to make it work with common SL features, I would need MORE THAN TWICE the amount of geometry on that character?

I am not a mesh expert so when you say that it is unoptimised that didn't really give me the impression that the angular butt would be sorted out during optimisation.  In my mind, I would have expected to model to have more triangles to start with and that this would be reduced by eliminating triangles that aren't necessary and so I expected the curves to be as curvy at the start as they would ever be.  It never occurred to me that optimisation meant making the body have more triangles than less.

42 minutes ago, FinnfinnLost said:

Regarding attachments and animations... you rig clothing and attachments to the existing armature and fit them for a mesh body of your choice. This does not require additional geometry on the body unless the topology on areas with high deformation is severely screwed up. This would make any RPG with an equipment system playable only on the most cutting edge PCs, which would not include mine.

No but it could require less geometry on the game bodies presumably because it is possible to avoid have to deal with layering mesh clothing, animating those layers or even having parts of the body that aren't shown.  What I mean by that is that you will always know what parts of the body will always be covered or shown where as in SL this is not true.  What's the point of the underlying mesh body having detailed and smooth body parts if they are never shown through the clothing, only the clothing detail would matter where as in SL the geometry is there for both.  I don't know how games are using mesh avatars but it seems that in that controlled environment there are many more opportunities to keep the body simpler because the parts that are never seen don't need to be there let alone have a smooth curve to them.

Edited April 3, 2021 by Gabriele Graves

[ChinRey]

2 hours ago, Gabriele Graves said:

I expect one of the reasons for the high poly count is the alpha cuts.

Yes, and also the tattoo and clothing layers which are dupliactes of the entire mesh body.

BOM should fix both of those issues though. It's too early to say how well it will work in the long run but I think we're already seeing significant improvements.

Edited April 3, 2021 by ChinRey

[Coffee Pancake]

1 hour ago, FinnfinnLost said:

Can anyone confirm that, just to make it work with common SL features, I would need MORE THAN TWICE the amount of geometry on that character?

To be honest, I don't think anyone really knows,there is a lack of documentation, trade sekrets and more than a little pure dumb luck.

The current approach to making mesh bodies has been to make the most awesome beautiful final render quality model practically possible, and then work out what needs to be sacrificed in order to make it work in SL. People expect every part of their avatar to look perfect at all distances (SL's free camera is a curse). Avatars can also be twisted up into entirely unrealistic poses (especially adults ones) and so on. One thing is clear however, there is a LOT of room for optimization.

Due to the complete lack of documentation, It's probably going to be simpler to approach the problem from the other end and make your own. (and far faster in practical terms)

Start with the most minimalistic single layer stick body you can make for the avatar bones, get it in SL and find & identify the problems from scratch, iterate with more complex models. Document each step, publish your work and do it all in the open (so people can help find problem animations and situations at each step of refinement). Applying the final functional blank onto a beautiful model as the final step.

 

Edited April 3, 2021 by Coffee Pancake

[FinnfinnLost]

@Gabriele Graves Different workflows exist, but you are mostly correct: Overall, optimization will probably lead to less complex models. However, during optimization you can also notice that additional geometry's required to refine certain spots. When reducing geometry you're also able to add more in case you need more detail. Ideally, you'll simplify the mesh to the point where you can afford subdividing the whole model to make it smoother. It may seem counter-intuitive at first, but bear with me:

Subdividing basically means turning a quad (a four-edged face) into four ones, by extension also quadrupling overall complexity. You can do this at the very start to get a fairly good impression of what the model can look like, but without optimization it's just not feasible for performance. However, when you remove geometry from the unsubdivided model, its complexity will be decreased. By extension, any subdivision will also add less geometry in places where you, for example, merged two faces into one. The trick is to find spots where less geometry is needed and remove the surplus. Flat areas that don't deform much are good examples for that. Afterwards, if you add a subdivision, every area will be smoothed out according to the amount of geometry you added.

Further fine-tuning can then be applied.

[Gabriele Graves]

12 minutes ago, ChinRey said:

Yes, and also the tattoo and clothing layers which are dupliactes of the entire mesh body.

BOM should fix both of those issues though. It's too early to say how well it will work in the long run but I think we're already seeing significant improvements.

Even without those on the triangles on just the body are still 176K.

[ChinRey]

24 minutes ago, Gabriele Graves said:

It never occurred to me that optimisation meant making the body have more triangles than less.

Oh, it may well. There's one thing so many people get wrong about optimisation, it's not about simplifying for the sake of simplifying, it's all about making the most out of the resources we have. Sometimes that means investing a little bit extra in one particular field.

 

4 minutes ago, Gabriele Graves said:

Even without those on the triangles on just the body are still 176K.

So, with tattoo layer and two clothes layers we're talking 700 ktris for the body alone, not including head, hair, clothes and accessories? There are top notch computer games with state-of-the-art graphics using fewer tris than that.

I know we keep complaining about how outdated the SL software is and how LL neglects client side performance. We have good reasons to but to b e fair, they must be doing something right. Avatars liek that are not supposed to work at all, not even on the strongest gming computers yet somehow they do.

[Gabriele Graves]

1 minute ago, ChinRey said:

Oh, it may well. There's one thing so many people get wrong about optimisation, it's not about simplifying for the sake of simplifying, it's all about making the most out of the resources we have. Sometimes that means investing a little bit extra in one particular field.

Understood.

1 minute ago, ChinRey said:

 

So, with tattoo layer and two clothes layers we're talking 700 ktris for the body alone, not including head, hair, clothes and accessories? There are top notch computer games with state-of-the-art graphics using fewer tris than that.

I know we keep complaining about how outdated the SL software is and how LL neglects client side performance. We have good reasons to but to b e fair, they must be doing something right. Avatars liek that are not supposed to work at all, not even on the strongest gming computers yet somehow they do.

Agreed, it sounds like the SL codebase isn't as bad as everyone says after all.

What I would like to know is that if a like-for-like body was made, still had the same alpha cuts and was 100% indistinguishable to the human eye, what should we be expecting to get in terms of triangles if an optimisation expert were doing it?  Is it possible for an expert to examine the body and see what the problem areas are using wireframe and make an estimate on the possible savings that do not affect how the body looks?  It would be good to get some actual numbers on this.

[FinnfinnLost]

8 minutes ago, Gabriele Graves said:

Understood.

Agreed, it sounds like the SL codebase isn't as bad as everyone says after all.

What I would like to know is that if a like-for-like body was made, still had the same alpha cuts and was 100% indistinguishable to the human eye, what should we be expecting to get in terms of triangles if an optimisation expert were doing it?  Is it possible for an expert to examine the body and see what the problem areas are using wireframe and make an estimate on the possible savings that do not affect how the body looks?  It would be good to get some actual numbers on this.

That would be very interesting indeed, but for it to work we'd require the actual source file without any changes that may have been made by the SL uploader. Ideally, it would be pre-triangulation as well, makes examining the topology easier. At least for me.

I highly doubt the more popular body creators would agree anyway, what with compromising a large part of their work.

[ChinRey]

4 hours ago, Gabriele Graves said:

Is it possible for an expert to examine the body and see what the problem areas are using wireframe and make an estimate on the possible savings that do not affect how the body looks?  It would be good to get some actual numbers on this.

I'm not sure if I count as an expert here since I've never really worked on avatars. But looking at the picture Aquila kindly provided and viewing it just as another mesh, I can't see many superfluous tris or vertices. I wouldn't expect to either; the avatar was made by Eric Call, the ultimate master of early SL optimisation, after all.

As for improving the look beyond what was possible in 2003, there are some curves that should be smoothed out, especially in the chest and bum area, some joints that might need some extra vertices to reduce animation distortion (although a more advanced rigging may be a better solution to this), a couple of places where more vertices could potentially open up for a wider range of body shapes (that would require an upgrade to the shape modifiers too) and of course, the fingers and especially toes could do with a lot more tris (with good LoD handling those tris wouldn't even increase the render load very much since they would mostly be gone at mid LoD).

As a good builder's estimate with all that in mind I'd say it would probably take about 20,000-30,000 tris and certainly no more than 50,000 to bring it up to the same level of visual quality as the average modern mesh body. But as I said, I don't have much experience with avatars and I may have missed some points. There are some parts of the head/face that might need the same treatment as the toes and I'm not sure if a good Bento facial animation would require a lot more tris. I too would really like to hear what a real expert has to say and there's one name that stands out. @Medhue Simoni, are you there?

Edited April 3, 2021 by ChinRey

[Janet Voxel]

2 hours ago, FinnfinnLost said:

That would be very interesting indeed, but for it to work we'd require the actual source file without any changes that may have been made by the SL uploader. Ideally, it would be pre-triangulation as well, makes examining the topology easier.

I feel like I read somewhere that the original Ruth body was a modified open source body. Finding that would probably be as close as you can get.

[Wulfie Reanimator]

6 hours ago, Gabriele Graves said:

I expect one of the reasons for the high poly count is the alpha cuts.  There are 208 cuts on the ML body which at eight faces per mesh object makes 26 objects that make up the body.  Obviously the head isn't included in this.  The main thing to understand about this though is that for a mainstream female body in SL, these kinds of cuts are essential for a large amount of users.

It's worth noting that triangles aren't the end-all-be-all. It's the overall trend that's being decried. There are other aspects of a mesh object that is arguably worse than having too many tris, but reducing tris is one of the simpler things if you've got the skills to make the model to begin with. When it comes to alpha cuts, the triangle count isn't increased by edge-splits but there are other (sometimes more severe) side-effects such as increased draw-calls.

6 hours ago, Gabriele Graves said:

It is easy to decry that this is because of the insane amount of triangles going into mesh avatars and that this shouldn't be happening.  I can accept that but I am also told that this can all be optimized without noticeable quality loss too so I am not sure what to believe there.

Did you misspeak here? Those two statements seem to be in line with each other.