Confused by 'auto retopology' results (Quad Remesh, plus others)

So I’m not really at this stage yet, still working on a bunch of hard surface stuff, but just because I had some free time, and some of my models are planned to be used for a basic game project, I wanted to get into the ‘correct workflows’ kind of mindset. So I was thinking about how you make a low-polycount game model, out of a high-poly (mostly due to subsurf to smooth curves, plus some sculpting) original.

So I think I know the manual route to do this. Using something like Retopoflow, you essentially re-draw the model in a more basic form using their tools. Time consuming, but it should work and give the best overall results (but takes a lot lot longer). That’s fine, and that’s on my to-do list.

But my model is already pretty simple (it’s not finished), so something like Quad Remesh, an auto-retopology tool that gets rave reviews and can retopology organic and complex shapes with ease and retain all details… should be a piece of cake. Right? Right?

So my results are below, in a ‘before’ and ‘after’ pic of part of the mesh. Essentially, the program seems to manage large areas of flat or curved mesh pretty well. But as soon as it hits anything with any kind of detail, it just poops itself and throws out random meshes. Some of them are broken geometry with faces folding back on themselves, but mostly it’s just terrible. And this is with me increasing the face-count, and tweaking all the settings. This is the best result I can get.

I -can- improve it, by doing a workflow that seemed to be suggested - using shrinkwrap project and a subsurf, to get the details back. But this results in a mesh that is… about the same number of faces as my original one was, if not more.

Might just be that I’m using it on the wrong kind of model, and this model doesn’t require / handle retopology well. Just confused me, as people seem to be using it on much more detailed and complex shapes without these issues.

Settings I used for the below I think was 10000 faces, 100% adaptive, adapt quad count, detect hard edges (which it doesn’t seem to be doing?).

I mean in fairness, it’s about as good as I would have expected an ‘automatic’ remesher to be, if I had just thought about it on my own, as it’s a complex task. My confusion is that so many people seem to be recommending it, and that it ‘just works perfectly’. Though they use it on models with millions of faces, while this model only has 200k. So maybe that’s the problem?

Edit: Just to add, I’m also totally willing to admit that I might be misunderstanding what the point of retopology is. I thought it was meant to turn a messy high-poly mesh, into a cleaner low-poly mesh, with big faces in the low detail areas and smaller faces in the detailed areas. Or something like that.

I’ve had good luck with the quad remesher plugin especially on hard surface models, the trick to get it to work well is enable autosmooth in blender and instead of using the detect hard edges option use the option called use normals splitting. But before retopoligizing make sure that there is no bevel or subdivision modifiers enabled you will want to add these modifiers to the mesh after it’s retopoligized.

I tried the normal splitting option, but it fails every time. Ill try it again later though, maybe I had some option still selected that was causing a conflict. Thanks for the suggestion!

One thing you could also try is creating and assigning some temporary materials on your mesh and using the Use Materials option. Maybe you could share a blend file with just that object you showed above and people take a closer look at it.

I’ve always associated retopo with sculpting, largely because sculpt mode uses tri’s not quads. The model in the screenshots doesn’t feel like it would need retopo, you could probably make the main panel from a couple of extrusions from a box and a couple of inset faces.
Bevel the edges and add an auto smooth to remove the infinitely sharp edges.

Model the buttons from a box and do the same with the edges.

here is my take if it helps:
since the goal is for low poly for games, all details will be baked in textures.
1 create primary object shape , may it be using bool, or manual mesh editing, in this case I use bool method( bevel, array, bool modifiers ).

2 clean it up to relatively low poly mesh. depends on how close and/or important the object is in final games, it may requires more or less poly counts. a good trick is to select Ngons - triangulate face - tri to quad.

3 use the low poly to remesh, in this case I use zbrush Zremesher. before zremesher it, use polygroup -> Groups by Normals, also set zremesher options keep groups, detectHardEdges on , smooth groups border = 0. there should be equivalent options in quad remesher (I dont have a copy of blender quad remesher, but same author as zremesher, so I assume the case).

4 use modifiers as you see fit, in this case I use bevel and subd for hard surface stuff, and some sculpting of chips if you wish.

5 finally bake high poly details to lowpoly mesh.

I already tried this. I set the top of one circle as one material, the edge of the circle as another material, and the ‘base’ part (the main panel) as a third, to make the edging as obvious as possible.

If anything, that circle ended up worse than the non-material painted ones. The software just didn’t seem capable of seeing and keeping edges.

Interesting, unfortunately I don’t have ZBrush and I won’t be likely to get it considering the cost. This is why I was trying to make it work in quad remesh. Even though it’s the same creator, it doesn’t seem to have that option to ‘keep groups’, or if it does then it’s not labelled to described in that way in the tooltips.

Part of my problem might well be that I’m doing too small a detail on the mesh. In my pictures you guys thought it looked like a box with buttons (which it does), but it’s actually a side-hatch on a much larger mesh. It was made by extruding out of that mesh, while perhaps I should have done it (and can still do it by separating it) as a separate object.

This pic is with me not using the ‘edge detection’ but managing to get ‘split normals’ to work. It actually comes out as an ok low-poly model, without any broken geometry. It loses details, which is to be expected, but at least it’s a mesh. When I turn on hard-edge detection and split normals, it causes a whole bunch of broken geometry. (ignore the little objects on the top of it, they’re actually separate objects and not part of the mesh. The door and the little spike and the ridge are the bits that are affected).

Ideal practise might be to make the ‘door’ a separate object. But then if that was true… it’d be like saying on a complex organic shape, each ear and the nose should all be separate objects. But in the examples I see they’re all just one model. So I dunno.

Again, this model probably doesn’t require any retopology at this point, I’m just trying to get a handle on whats possible / whats needed before I have invested hundreds of hours into some stupidly complex model that then can’t be used because retopo doesn’t work! lol

Yeah, you can get fairly similar results in Blender, but using temporary materials instead of polygroups. I suppose it all depends on what the original mesh looks like

The original mesh looks like this. Which is probably usable as-is for low poly (its only 1600 faces) but it’s not very smoothed and the ‘render smooth’ looks terrible. But a single subsurf ups it to 20k faces, which is too many (or at least so it seems from my research, 10k is the upper limit?)

However for this model, the best solution is probably to keep this 1600 face model (maybe add a few more loops to fix some geometry), and then bake the normals from a subsurfed version to make it look better. I was just experimenting really, as I want to learn how things work (preferably before I actually -need- them to work!)

You are going to become frustrated if you are planning to use Subdivision Surface modifiers on that and expect things to look clean and smooth. You’ll need more supporting edges and the ones you added already have made your sphere non-spherical.

For that kind of workflow, you should either model it exactly as you want it low-poly and then it’s done, or model it high-poly and create a low poly version from that later.

If you started with a higher resolution UV Sphere to start with you’d find it easier. Notice how in this the supporting edges for the inset shape don’t interfere with the shape of the rest of the sphere

A good point, originally I had expected to be able to add extra loops where detail was needed, and use subsurf to complete the spherical smoothness. However this did end up causing more issues than I expected.

It’s my first large scale spherical model though, so it’s a learning experience. Plus I was mostly sketching out the shapes, so I could probably remake it with a higher density uv sphere in no time.

Do you have a recommendation for a sphere mesh density that would be suitable polys for game models?

I’m no expert on game models, but it would have a lot to do with the type of game, the platform and the intended use of each asset in the game. In a first person shooter type game the gun models and player models would have a far greater polygon count than scenery props like furniture or things that are only seen in the background or from a distance.

I’m trying to replicate the original model now, but starting from a much higher-density mesh sphere.
Another thing I discovered which gives me much better geometry to work from - not using a UV sphere at all, but using a ‘Quad Sphere’ or cube sphere. The lines line up so much nicer.

Not really related to the original point of this thread, but still, thought I’d add it here aha.

1 Like

Just had a thought while updating some of my addons. Have you considered using something like DecalMACHINE to detail parts of your object? This is just a sphere made from a cube as you mentioned you were using now. The panel lines and whatever that other thing is are just decals. These can all be baked out to textures.

Interesting… this might well be a quicker workflow in theory (obviously I’d have to learn it first haha). I might well try and use this for my next project.

Out of interest (and unrelated to this thread really), I’ve been asking around for feedback on something. I mostly do hard surface work, and am about to start delving into Mixer for texturing, and I am considering picking up a graphics tablet for the texture-painting workflow. Though looking at Mixer, it seems most people just use a mouse for it still.

If I’m not doing organic sculpting, is there much benefit to using a tablet for hard surface and texture painting?

Reason I was considering getting one, is for adding in the fine edge details like the rust and soot marks/streaks that you get ‘blown’ behind panel lines, but I’m now thinking that’s possible to do without hand-painting it… wondering if aiming to hand-paint anything in 3D is just going to slow down my ability to create things. Just curious on your opinion (as you’ve been so helpful so far :slight_smile: haha)

If you plan on doing any serious texture painting then a tablet would be a good investment. It’s amazing the difference made by being able to vary pressure to control opacity or size on the fly. Drawing with a mouse feels like drawing with a bar of soap. Even for very basic things like simple scratches. Imagine how hard it would be to draw a scratch with a mouse. It would likely be a uniform thickness and strength. With a tablet you can paint in natural looking ones which taper off at the end or ends.

You might even find a use for sculpting down the line. Some people block out hard surface models as sculpts and then retopologise them. In the Blender Livestream that was on earlier today you can see that Daniel Bystedt has sculpted that scifi head/helmet/robot thing