Uploader-generated LODs are the Devil's vomit

[Rick Nightingale]

...and I should know, being the Devil.

Just wanted to say this to any prospective meshers who might be reading here (and some old ones who should know better!).

Yesterday I saw something I wanted to buy at someone's place... started searching the MP for it, then checked out the LODs... nope. Stopped searching the MP for it and felt miffed because I really liked it at first. It was made by a well known name that I won't say, but probably most here will have heard of them.

The thing was lovely, at High LOD. All four LOD models were present at what looked like reasonable tri counts which got my hopes up at first, but when I viewed them they were obviously uploader generated and just looked rubbish. I (or anyone who cares) could have made the levels with that many tris and had the thing look indistiguishable from the High LOD at a reasonable LOD factor (like 2, not 4) and distance. Maybe even saved some tris and dropped the LI from 3 to 2.

So please... newbie (and not so newbie) mesh makers... take the time to make your own LOD models. It matters. It'll look better,  shouldn't increase LI if you do it right (maybe even lower it), and if I like it is far more likely to get you a sale.

Uploader-generated LODs are the Devil's vomit!

[Candide LeMay]

That ship has long sailed. LL made two mistakes here 1) they included the LOD generator in the uploader instead of forcing people to provide lods 2) implemented adjustable LOD factor in the client (how are you supposed to fine tune the lod switching if the distance at which it happens is different for different people??). Each of them is bad but the combination of both at the same time is lethal. We'll swim in vertex vomit until the end of times.

(for extra amusement, try to run around with LOD factor set to 0 for a while and weep. there's basically no preserved object silhouette in sight, unless it's old skool prims)

Edited December 16, 2022 by Candide LeMay

[ChinRey]

Maybe we should illustrate this. Here are two copies of the same tree. The one to the left is optimized with manually created LOD models and it's 4 LI. The one to the right is uploaded with default auto generated LOD models and its land impact is 6:

Here's how they look at mid LOD:

And at low and lowest LOD (those two are identical for both copies in this case):

Does anybbody notice the difference? And keep in mind that the tree to the right actually has a 50% higher land impact than the one to the left.

This isn't a particularly bad case. I've seen better results from auto generated LOD models but I've also seen much worse. Then again, to be fair making a tree like the one to the left would be well beyond the skills of even most seasoned true professional.

Even so:

3 hours ago, Rick Daylight said:

Uploader-generated LODs are the Devil's vomit!

I wouldn't go that far. Second Life is supposed to be a platform for merry hobbyists. We have to respect and accept that they don't have pro level skills. Also, there are some occasions when auto generated LOD models give an acceptable, although never great, result.

Here is my advice to all SL mesh makers:

• If you don't know how to make LOD models manually (or can't be bothered to), do not pretend you are a pro or even an advanced amateur because you are not. Don't get me wrong, it's perfectly fine if you aren't, just do not pretend you are.
• Err on the right side. It's better to sacrifice a little bit of LI and performance by over-strenghtening the LOD than to have your meshes collapse. Look at Mole meshes. The Moles have no clue whatsoever how to handle LOD but their meshes still hold up well. Crazy high land impact, yes, rather laggy. yes, but they hardly ever collapse and that's the most important thing.
• At least try to make the LOD models yourself. You'll be amazed how easy it actually is to outperform the uploader's LOD generator.
• Don't expect perfection. I'm probably one of the top ten, possibly top three, experts SL optimization but even I'm still learning and my meshes are still far from perfect.

Edited December 16, 2022 by ChinRey

[animats]

LOD generation needs to smartened up. Everybody else has that automated now.

Technical notes:

• LOD generation needs what Unreal Engine's LOD generator calls "silhouette protection". The reduced version will have roughly the same outline as the original. No holes.
• We need simple impostors, and a generator for them. SL doesn't have this because, traditionally, textures and meshes are handled separately, and an impostor is a texture picture of a textured mesh. The plan for converting to gLTF uploads should include this for new content.
• A tempting thought: an LOD checker for content. Renders each uploaded object from 8 directions plus straight down, for each LOD. Compares each LOD with a reduced-size version of the main image. Marketplace upload should check this.

[ChinRey]

On 12/16/2022 at 9:39 PM, animats said:

LOD generation needs to smartened up. Everybody else has that automated now.

Yes but there are several factors that makes this much harder for SL than for other game/VW engines. For a start Second Life scenes tend to have far more objects than we usually find in other environments so optimizing geometry (and textures/surface maps) is far more important here. Also, we can't adjust swap distances, draw distance and the number of LOD levels for individual objects and that makes it much harder to make adequate LOD models. Then there is the high land impact penalty for LOD model geometry. I do not know how well this reflects the actual load higher poly LOD models cause but it certainly entices content creators to decimate far more than what you usually see elsewhere.

Generally a mesh with LOD models that is perfectly good for, say Unity or UE, will not perform very well in SL and that means the automatic LOD generation algorithms used there won't either.

On 12/16/2022 at 9:39 PM, animats said:

We need simple impostors, and a generator for them.

Oh yes, that would be very useful. Is there a chance we'll ever get it?

[ChinRey]

I thought I'd revisit this thread with a new example since I didn't take into account that Fs at least (don't know about the LL viewer) offer several alternative decimation algorithms now. I do not know what they are but it seems they are just alternative ways to use GLOD rather than completely different algorithms. Maybe @Beq Janus can answer that? (Edit: Aquila andsered it, they aren't.)

Anyway, here are five copies of the same marble column. The one to the left is uploaded with proper LOD models, the other with default "Auto", "Precise", "Sloppy" and "Reliable" respectively.

The manual, Auto and Reliable copies all have a nice and low 0.1 downlaod weight which means they are well and truly below the limit for 1 LI. The Precise and Sloppy ones on the other hand, both have 3.8 download weight which means they end up at a whopping 4 LI - waaaaay too high for such a simple object.

(The reason why the seem to be different size is partly that you see more of the sides of the columns toward the ends of the row, partly that Linden Lab still hasn't noticed it's the 21st Century so Second Life's rendering engine only works properly with old 20th C. 5:4 ratio monitors. On modern 4:3 screens the picture is noticeably stretched sideways once you move off the very center of the picture.)

The size of the column is 0.2x0.2x1 m which means it's the medium and to some degree the low LOD models that matter the most. You only see the high model when you bump into the column and you only see the lowest when they are way off in the distance like this:

 

So here are the main - that is medium LOD - models:

As you see, "Auto" and "Reliable" do a lousy job, "Sloppy" and "Reliable" are downright distastrous.

Moving on to low LOD:

This is where the good old Auto really fails and the swap distance is only 9.7 m with LOD factor 1.25 or 17.3 m with LOD factor 2 so that model is still important. But the "Precise" and "Sloppy" algorithms grossly overdo this LOD level. You do not need all those details and this is of course why they give such huge land impacts. The "Sloppy" algorithm really lives up to its name too, adding far more tris to the low (and lowest) models than it thought the medium one was worth. The "Reliable" algorithm still fails miserably to live up to its name.

Finally lowest LOD:

You do not need much detail at this level, just enough to smoothen the transition as you move towards the object. The proper LOD model only has eight tris and that's enough. The "Auto" model has 12 tris, eight that actually show and four wasted ones, apparently hidden away in some alternative pocket universe. 12 is a good tri count for this but the algorithm doesn't exactly put them to good use, in fact the model it produces is more of a nuisance. The "Reliable" algorithm just gives up at this stage and that's probably just as well.

But look at the "Precise" and "Sloppy" variants. They are going all in here keeping 68 tris out of the original 120. That's total overkill of course and look at the second picture in this post again. Those extra tris don't help preserve the looks of the models, they actually distort it! This is another reason why over-strengthening the LOD models is a bad thing.

Edited December 23, 2022 by ChinRey

[Aquila Kytori]

4 hours ago, ChinRey said:

I didn't take into account that Fs at least (don't know about the LL viewer) offer several alternative decimation algorithms now. I do not know what they are but it seems they are just alternative ways to sue GLOD rather than completely different algorithms.

GLOD replaced by MeshOptimizer  (one word).

 

From Inara Pey's Living in a modern world,  https://modemworld.me/2022/09/09/firestorm-6-6-3-performance-improvements/#lab

"  The Performance Improvements viewer also includes the MeshOptimizer Project .

• By default, this replaces obsolete GLOD’s mesh simplification in the mesh uploader with the newer and more powerful MeshOptimizer package to provide better level of detail (LOD) modelling during the upload process.
• Note that as an option, Firestorm also retains the GLOD capability – see below for more. " 

 

The Glod option is the one named Reliable.

 

 

[Charlotte Bartlett]

5 minutes ago, Aquila Kytori said:

GLOD replaced by MeshOptimizer  (one word).

 

From Inara Pey's Living in a modern world,  https://modemworld.me/2022/09/09/firestorm-6-6-3-performance-improvements/#lab

"  The Performance Improvements viewer also includes the MeshOptimizer Project .

• By default, this replaces obsolete GLOD’s mesh simplification in the mesh uploader with the newer and more powerful MeshOptimizer package to provide better level of detail (LOD) modelling during the upload process.
• Note that as an option, Firestorm also retains the GLOD capability – see below for more. " 

 

The Glod option is the one named Reliable.

 

 

Just a note C4d for the rare users still has the issue.

I just did a new house release yesterday for the Sat Sale and have been doing full custom LODs by hand.  I have to “wash” the file from fbx to dae in Blender for SL due to the C4D issue.    The effort creating a full house is intense and doesn’t pay back commercially.    That being said I can at least rest easy knowing I optimized the very best I could.   I do tell customers re the LODs as I think education helps too so people start to adjust their view in what “good looks like”.      But honestly if I was releasing more frequently I am not sure the workload could be managed.    

[ChinRey]

3 hours ago, Aquila Kytori said:

GLOD replaced by MeshOptimizer  (one word).

Ah!

In any case, not much improvement in this case then. It is of course possible MeshOptimizer will handle other models better. This column is open at the top and bottom so it's a shell, not a closed body and I've told that is something MO is particularly bad at.

  

3 hours ago, Aquila Kytori said:

The Glod option is the one named Reliable.

Somebody on the Firestorm dev team has a weird sense of humour.

Edited December 23, 2022 by ChinRey
Typos

[Beq Janus]

On 12/23/2022 at 6:00 PM, ChinRey said:

Somebody on the Firestorm dev team has a weird sense of humour.

/me takes a bow and grins. 

I chose that name, not as a cruel joke, but for a lack of better words to describe it. The reason I settled on reliable is because (whether you like the results or not - I am certainly not suggesting they are good) GLOD reliably produces something that is close to what you told it to aim for, and is what you have used in the past. It will reliably adhere to your request to crate very low poly (awful looking) LODs. MeshOptimiser on the other hand is bloody awful in about 2/3 of all cases that I tested, often resulting in excruciatingly bad LODs that cost the unobservant creator vast amounts of Lindens for unusable uploads. It almost completely ignores the target numbers, taking them as a guide rather than a limit. GLOD is also (typically) far more reliable when it comes to retaining UVs, whereas both auto (which is really just a wrapper around precise and sloppy) and sloppy pretty much destroy your UVs.

I am happy to explore other options that people feel are worth having. They need to be freely distributable of course, and be able to be used from within the viewer.

MeshOptimiser is not a bad library, the way that the uploader is setup to use it though is far from ideal, to the point where I requested it not be released with the performance viewer update, but kept separate so that people could focus on it as a beta and perhaps nudge it in the right direction. I did not feel it was ready, which is reflected also by my decision to reinstate the GLOD library.

 

[ChinRey]

6 hours ago, Beq Janus said:

MeshOptimiser on the other hand is bloody awful in about 2/3 of all cases that I tested, often resulting in excruciatingly bad LODs that cost the unobservant creator vast amounts of Lindens for unusable uploads. It almost completely ignores the target numbers, taking them as a guide rather than a limit

My two tests agree with you there. That is, with "Auto" it did perform marginally better than GLOD for the column. Not anywhere near good enough to keep though and the tree, which was also done with "Auto", was of course a total disaster.

 

6 hours ago, Beq Janus said:

I am happy to explore other options that people feel are worth having. They need to be freely distributable of course, and be able to be used from within the viewer.

I've mnetioned it already but I think the problem is that we need much heavier decimation in SL than they do in a computer game, not only because of the LI calculation's overemphasis on the lower LOD models' weights but also because we tend to have far more complex scenes so simplification of each objects becomes much more important. The lack of object level swap distance adjustment makes this problem much worse. Apart from the messed up normals (see below) the 60 tris throughout all LOD models the "Precise" option delivered would have been ok in a typical Unity/UE setting but in SL it's both way too heavy and too poor quality at the same time.

There is another serious issue that I haven't even seen mentioned in any discussions about automatic LOD generation: transitions. Making good LOD models is not only about finding the right tris and vertices to keep for different view distances. It's also about how the various levels relate to each other. Normals are particularly troublesome here. When I made my own lowest LOD model for the column, the challenge wasn't to get the tri coutn low enough - that was dead easy. The problem was to avoid the top of the column to suddenly light up like a beacon or suddenly go all dark. Those effects are really, really noticeable even at that small scale and very visually annoying.

As far as I know, all mesh decimators treat each LOD model as a separate entity and do not treat the set as a whole.

[OptimoMaximo]

1 hour ago, ChinRey said:

Normals are particularly troublesome here. When I made my own lowest LOD model for the column, the challenge wasn't to get the tri coutn low enough - that was dead easy. The problem was to avoid the top of the column to suddenly light up like a beacon or suddenly go all dark. Those effects are really, really noticeable even at that small scale and very visually annoying.

That's what a vertex normal transfer is for...

[ChinRey]

1 hour ago, OptimoMaximo said:

That's what a vertex normal transfer is for...

Oh yes, that feature has solved a lot of LOD models challenges for me! It's far from perfect and doesn't always work at all but it still helps a lot. And of course, more to the point here, it's something no decimation algorithm I am aware of supports at all.

Edited December 27, 2022 by ChinRey

[anitabush]

3 hours ago, OptimoMaximo said:

That's what a vertex normal transfer is for...

Stuff like this is why we need some kind of in depth guide. As someone new to second life creation I had never heard of this before. Im desperately trying to learn and make efficient models but there’s so many techniques, tips and tricks spread across dozens of threads and websites that it’s really hard to find them. 

So thank you for this, I’ve done some reading on vertex normals and the data transfer modifier in blender thanks to your comment and I will be doing some experimenting later.

[Fluffy Sharkfin]

1 hour ago, anitabush said:

Stuff like this is why we need some kind of in depth guide. As someone new to second life creation I had never heard of this before. Im desperately trying to learn and make efficient models but there’s so many techniques, tips and tricks spread across dozens of threads and websites that it’s really hard to find them. 

The problem is the sheer amount of relevant information available given all the various commonly used workflows and tools.

Some folks spend 3 or 4 years at university just to gain a solid foundation upon which to build their knowledge and skills and will continue learning and developing those skills throughout their professional career since the software and techniques used in the content creation process are continually evolving at a fairly rapid pace.  Sure the vast majority of SL creators aren't interested in creating content professionally but that doesn't make learning to do it properly any easier in the same way that wanting to be a brain surgeon "just for fun" doesn't makes brain surgery any less complicated (thankfully the stakes in content creation aren't as high, you may burn out a few GPUs but you probably aren't going to kill anyone).

I would like to see LL provide more resources for creators but I think it would be more impactful if creators were encouraged to expand the scope of what they learn beyond how to make stuff for SL.  All the necessary relevant information is already available online but only a small portion of it is covered in tutorials that focus purely on creating content for Second Life, so a lot of creators end up just learning the bare basics of 3D modelling along with a few SL specific skills like exporting and uploading mesh to SL, rigging mesh for specific bodies and how to make a mesh with a land impact of 1 (even when it should be higher 🙄) rather than actually learning all the important do's & don'ts of creating content for platforms like Second Life.

[OptimoMaximo]

3 hours ago, anitabush said:

Stuff like this is why we need some kind of in depth guide. As someone new to second life creation I had never heard of this before. Im desperately trying to learn and make efficient models but there’s so many techniques, tips and tricks spread across dozens of threads and websites that it’s really hard to find them. 

So thank you for this, I’ve done some reading on vertex normals and the data transfer modifier in blender thanks to your comment and I will be doing some experimenting later.

Glad to have given you a track to learn something new and possibly useful in the future! 🤗

And I understand your feelings when you say that it would be beneficial to have an in depth guide. But believe me, a guide like that would be an Egyptian pyramid stone block sized book. Unfortunately, 3D modeling and all the related matters (rigging, animation, etc) currently are university courses matters: require study, practice (lots and lots), trials and errors, features digging... And more.

The very moment that LL introduced Mesh content into the platform, they made content creation require a (at least semi-) professional grade skill set.

Edited December 27, 2022 by OptimoMaximo

[OptimoMaximo]

1 hour ago, Fluffy Sharkfin said:

The problem is the sheer amount of relevant information available given all the various commonly used workflows and tools.

Some folks spend 3 or 4 years at university just to gain a solid foundation upon which to build their knowledge and skills and will continue learning and developing those skills throughout their professional career since the software and techniques used in the content creation process are continually evolving at a fairly rapid pace.  Sure the vast majority of SL creators aren't interested in creating content professionally but that doesn't make learning to do it properly any easier in the same way that wanting to be a brain surgeon "just for fun" doesn't makes brain surgery any less complicated (thankfully the stakes in content creation aren't as high, you may burn out a few GPUs but you probably aren't going to kill anyone).

I would like to see LL provide more resources for creators but I think it would be more impactful if creators were encouraged to expand the scope of what they learn beyond how to make stuff for SL.  All the necessary relevant information is already available online but only a small portion of it is covered in tutorials that focus purely on creating content for Second Life, so a lot of creators end up just learning the bare basics of 3D modelling along with a few SL specific skills like exporting and uploading mesh to SL, rigging mesh for specific bodies and how to make a mesh with a land impact of 1 (even when it should be higher 🙄) rather than actually learning all the important do's & don'ts of creating content for platforms like Second Life.

This^^^^^^

[anitabush]

22 hours ago, Fluffy Sharkfin said:

The problem is the sheer amount of relevant information available given all the various commonly used workflows and tools.

Some folks spend 3 or 4 years at university just to gain a solid foundation upon which to build their knowledge and skills and will continue learning and developing those skills throughout their professional career since the software and techniques used in the content creation process are continually evolving at a fairly rapid pace.  Sure the vast majority of SL creators aren't interested in creating content professionally but that doesn't make learning to do it properly any easier in the same way that wanting to be a brain surgeon "just for fun" doesn't makes brain surgery any less complicated (thankfully the stakes in content creation aren't as high, you may burn out a few GPUs but you probably aren't going to kill anyone).

I would like to see LL provide more resources for creators but I think it would be more impactful if creators were encouraged to expand the scope of what they learn beyond how to make stuff for SL.  All the necessary relevant information is already available online but only a small portion of it is covered in tutorials that focus purely on creating content for Second Life, so a lot of creators end up just learning the bare basics of 3D modelling along with a few SL specific skills like exporting and uploading mesh to SL, rigging mesh for specific bodies and how to make a mesh with a land impact of 1 (even when it should be higher 🙄) rather than actually learning all the important do's & don'ts of creating content for platforms like Second Life.

 

20 hours ago, OptimoMaximo said:

Glad to have given you a track to learn something new and possibly useful in the future! 🤗

And I understand your feelings when you say that it would be beneficial to have an in depth guide. But believe me, a guide like that would be an Egyptian pyramid stone block sized book. Unfortunately, 3D modeling and all the related matters (rigging, animation, etc) currently are university courses matters: require study, practice (lots and lots), trials and errors, features digging... And more.

The very moment that LL introduced Mesh content into the platform, they made content creation require a (at least semi-) professional grade skill set.

Thank you for the replies. I do see what you mean and I suppose fully in depth guides would be unrealistic. It would just be lovely to have some tutorials from linden going through their workflow for some different items using different tricks and techniques. I’m a few years into learning blender via the university of youtube, a month into substance painter (adobe make some lovely tutorials for beginners) and a sporadic year or so learning how to get that stuff into SL. I’ve made a couple of reasonable things recently and lots of terrible ones before that but I’m getting there slowly. 
 

[OptimoMaximo]

5 hours ago, anitabush said:

 

Thank you for the replies. I do see what you mean and I suppose fully in depth guides would be unrealistic. It would just be lovely to have some tutorials from linden going through their workflow for some different items using different tricks and techniques. I’m a few years into learning blender via the university of youtube, a month into substance painter (adobe make some lovely tutorials for beginners) and a sporadic year or so learning how to get that stuff into SL. I’ve made a couple of reasonable things recently and lots of terrible ones before that but I’m getting there slowly. 
 

If you're learning only through YouTube, you're going slower than you could. My evergreen suggestion is to invest on a manual, best if physical print (well that's my taste 😁) , to read, study and try the featured classes. Doesn't matter if things in those classes aren't SL related or relatable, do them anyway. Build your acquaintance and confidence with your software! I guarantee you'll make a quality jump.

[anitabush]

1 hour ago, OptimoMaximo said:

If you're learning only through YouTube, you're going slower than you could. My evergreen suggestion is to invest on a manual, best if physical print (well that's my taste 😁) , to read, study and try the featured classes. Doesn't matter if things in those classes aren't SL related or relatable, do them anyway. Build your acquaintance and confidence with your software! I guarantee you'll make a quality jump.

Do you have any recommendations? 

[Fluffy Sharkfin]

1 hour ago, OptimoMaximo said:

If you're learning only through YouTube, you're going slower than you could. My evergreen suggestion is to invest on a manual, best if physical print (well that's my taste 😁) , to read, study and try the featured classes. Doesn't matter if things in those classes aren't SL related or relatable, do them anyway. Build your acquaintance and confidence with your software! I guarantee you'll make a quality jump.

This is valuable advice!

There is a huge and often overlooked difference between learning from how to videos and learning from a collection of structured tutorials designed to teach you the fundamentals.

Learning from how to videos may seem faster at first since you just have to memorize a specific workflow but the problem is that if you wish to create something outside the scope of that workflow then, unless you can find another tutorial which builds on the limited knowledge you already possess, your only option is to find a new tutorial and learn an entirely new workflow which may or may not be compatible with the workflow you're used to, which in turn can lead to all sorts of headaches when you try to combine what you've just learned with what you already knew if you don't have a solid understanding of the various tools and features you're using.

In comparison, working your way through an entire manual or compilation of tutorials will take you far longer initially however by the time you're finished you'll have a thorough understanding of all the tools and features available and how they can be combined to create any workflow you may need.  This means that not only will you find it a lot easier to understand and retain the information you gain from watching/reading tutorials, you'll also be less reliant on tutorials in general since once you understand the full capabilities of the software you're using it's a lot easier to work out for yourself the necessary steps to complete a task.

Additionally, having a solid foundational knowledge of how a piece of software works and what all the tools and features are capable of means you can more easily transpose tutorials designed for other similar applications and adapt the workflow to suit the software you're using.  For example, a lot of the more advanced sculpting techniques I've learned in 3D Coat have been from Zbrush tutorials (since there are significantly more tutorials available for Zbrush) however, having taken the time to learn the majority of the tools and features available in 3D Coat, when I see someone use a certain tool or feature in a Zbrush tutorial I can usually recognize what that tool is doing and easily identify the equivalent tool in 3D Coat.

In essence, once you learn all the fundamentals of the various tools and features of a piece of software then learning new techniques and workflows becomes a lot easier and far less frustrating and your creativity will be free to flourish far beyond the boundaries of "how to" videos.

[OptimoMaximo]

1 hour ago, anitabush said:

Do you have any recommendations? 

Well, not really first hand recommendations... I don't even touch Blender anymore for the last 10 years or so, and when I was learning it (2004-2005) the manual I bought was called Blender for Dummies, which I believe is no more... But I'm sure that a look up on Amazon would turn something interesting up

This one, for example...

Blender 3D By Example: A project-based guide to learning the latest Blender 3D, EEVEE rendering engine, and Grease Pencil, 2nd Edition https://amzn.eu/d/9c2ddh0

 

Edited December 29, 2022 by OptimoMaximo

[ChinRey]

On 12/28/2022 at 5:19 PM, anitabush said:

It would just be lovely to have some tutorials from linden going through their workflow for some different items using different tricks and techniques.

No! Linden Lab does not have anybody qualified to do this, their developers do not understand the practical side of mesh making and their content creators do not understand the technical side. If they ever made such tutorials, they would end up as examples how not to do it.