Why not use uploader generated LOD models?

[ChinRey]

I've lost count how often I've got that question, well here is an answer.

Take a look at this stable:



It's my prize for the Medieval Fantasy Hunt this month and the land impact is 4 with proper LOD, proper physics and all that.

When I did the first test import on the beta grid, I didn't bother with all that, I just let the uploader generate the different models using the default settings. Slightly worse LOD than the final version and a land impact of 44!

This is an extreme example of course, for a number of reasons, but it illustrates the point. Even an inexperienced mesh maker can easily outdo the uploader's simplification algorithm and create LOD models of the same quality but with significantly lower LI. With a bit of experience, you can usually reduce the LI to a fraction of what you can get with automatically created LOD models with no quality reduction whatsoever.

[Drongle McMahon]

Here's another example that illustrates two ways custom LODs are worthwhile. This is a girder that uses a lot of extra geometry to get proper highlights on rounded edges. It uses single segment edge bevelling with edited normals (which gives about half the LI of the easier two segment bevel with profile=1). Uploading with auto-LODs gives triangle counts 1249, 313,73,37, which gives a LI of 4 (at 1x1x6 m). The custom LODs are 1249,241,25,13, for an LI of 1. The autoLOD one could be reduced by setting the lowest triangle count to 1, and actually it's so bad it might as well be. So LI alone isn't reason enough on its own. Instead, it's the quality of the LODs.

The first picture shows the two models, custom LODs at the back, auto LODs in front, with the lower LODs at the nearest distances they appear with the LL default RenderVolumeLODFactor (1.125). The picture below shows the three lower LODs at the same scale, so that you can see what's happening. All taken at noon with no extra lighting.



At the first step, the custom LOD simply replaces the smooth shaded bevelled version with an unbevelled version with sharp corners. The auto LOD generator can't handle this at all. It leaves horible angled faces with smooth shading and horribly incorrect highlights. That's the first reason for using custom LODs. If you use geometry to get proper highlighting, it will be badly messed up at the first LOD switch in the auto-LOD.

For the lowest two LODs, the custom LOD model uses alpha masked textures on simple planes instead of geometry. That way it can save huge amounts of geometry, which is why it's LI is lower that the auto LOD model. That's the second reason, but it's not just for lowering LI (for which you can use the triangles=1 trick anyway). It's for appearance. These lower LODs are far better than the disintegrating scraps left by the auto LOD.



PS. I had to copy your pictures and up the gamma by 1.4 to see your them at their best. Maybe you would do better to tweak them a bit before uploading?

[ChinRey]

---

Drongle McMahon wrote:

The autoLOD one could be reduced by setting the lowest triangle count to 1, and actually it's so bad it might as well be. So LI alone isn't reason enough on its own. Instead, it's the quality of the LODs.

---

Yes, and that's an important point. You can always reduce LI by sacrificing LOD. But good mesh making - or at least the technical (as opposed to artistic) aspect of good mesh making - is all about making the most out of the limited computing resources we have. You obviously don't achieve that by sacrificing viewing quality and you definitely don't achieve it by asking the end user to increase their lag by raising their LOD factor.

To a computer the model is just a set of numbers, it has no visual perception at all. It uses an algorithm - a set of generic rules made by somebody who has no idea how your specific model looks - to determine which details to keep and which to omit. Despite what LL claims GLOD, the algorithm used by the uploader, isn't even a very sophisitcated one. More advanced decimation functions, like the one in MeshLab and the two in Blender, merge triangles to keep as much as possible of the original shape. GLOD just deletes triangles until it reaches the preset number, leaving large holes in the model. And since it doesn't really know which triangles are the most important to the model, it will keep lots of unnecessary ones and remove vital ones.

Even the most novice mesh maker can easily outperform any computer program in this.

 

Except..

 

My latest mesh upload turned out to be an oddity. It's a bush that looks like this:



In this case i ended up using auto generated models for the two lowest levels. Not default, I had to crank up the resolutions a lot, but auto generated all the same.

The LI is 0.4 at the default 1.2x1.2x1.2 m size and 0.8 at 2x2x2 m, the largest size I think it should be scaled to. Not bad at all actually and, more to the point, much better than anything I was able to make myself. Oh well, I don't know how many meshes I have uploaded but it's a five digit number. I guess I can live with being beaten by a computer only once.

I actually think I know what's going on here. Nyx Linden once told me that GLOD's main focus was to preserve as much of the overall contours of the model as possible and in this particular case that was exactly what I needed.

---

Drongle McMahon wrote:

It uses single segment edge bevelling with edited normals (which gives about half the LI of the easier two segment bevel with profile=1).

---

I really have to look at the new custom normals function in Blender. Looks very useful and besides, there has to be some compensation for having to look at that horrible startup image every time I launch the 2.75

 

---

Drongle McMahon wrote:

PS. I had to copy your pictures and up the gamma by 1.4 to see your them at their best. Maybe you would do better to tweak them a bit before uploading?

---

Thank you! I tried it to adjust the gamma this time and I think it improved quality. I had to reduce it a bit though, not increase it.