Your GPU has never drawn a quad. Every quad you have ever modelled was split into two triangles somewhere between your viewport and the screen, and the same is true of every n-gon. Quads are an authoring convenience, not a rendering feature. Once that is clear, the whole triangles-versus-quads argument stops being religious and becomes a set of specific, answerable questions about what you plan to do with the mesh.
The short version: if the asset is static and going straight into an engine, triangles are fine and worrying about quads is wasted time. If it deforms, subdivides, or will be edited by hand later, quads pay for themselves several times over. Everything else is detail.
What quads actually buy you
Predictable subdivision. Catmull-Clark subdivision turns every face into quads regardless of input, but it does so cleanly only when the input is already quads. Feed it triangles and each one becomes three quads meeting at a valence-3 vertex — a pole — which produces a visible pinch on any smooth surface. Subdivide an 8,000-quad base mesh twice and you get 128,000 quads; do the same to a triangulated version and you get the same face count with a scattering of shading artifacts you did not ask for.
Deformation that behaves. When a joint bends, edges perpendicular to the bend axis carry the deformation. Quad topology lets you place edge loops deliberately: three loops through an elbow, four or five across a shoulder, a radial arrangement around the mouth. Triangulated topology has no loops to select, so you get whatever collapse the mesh happens to produce. This is the single biggest reason character artists refuse to hand off triangles.
Editability. Loop select, loop cut, edge slide, and bevel all depend on quad flow. On a triangulated mesh these tools either fail or wander. If a human will open this file again in six months, quads are a courtesy to that human.
Cleaner bakes and cleaner UVs. Quad edges follow features, so UV seams land where you meant them to rather than zig-zagging across a panel.
The triangulation trap
Here is the bug that costs people a full day, usually twice, before they learn it. A quad can be split along either diagonal. Those two splits produce different surfaces whenever the quad is non-planar, and almost every quad on a curved surface is non-planar.
Now consider the normal-bake workflow. You bake a normal map in Blender, which triangulates the low-poly one way for the bake. You export FBX, and the exporter — or Unity’s importer, or Unreal’s — triangulates it the other way. The baked normals now describe a surface that does not match the rendered surface. The symptom is a faint diagonal gradient across large flat-ish faces, worst on a 45-degree angle, and it survives every attempt to fix it in the texture.
Triangulate the low-poly mesh before you bake, and export that exact triangulated mesh. Not a copy. Not a modifier evaluated at export time by a different code path. The same mesh.
In Blender this means applying the Triangulate modifier (Beauty method for both quads and n-gons) before baking. Keep an untriangulated version in a separate collection if you need to edit later, and re-bake if you do. glTF avoids the ambiguity entirely because the format only stores triangles, which is one of several quiet reasons GLB round-trips more reliably than FBX.
Triangulation also changes your vertex count, which is what actually costs memory. A quad shares four vertices; splitting it adds none. But hard edges and UV seams do split vertices, because a vertex can only carry one normal and one UV coordinate. A 30,000-triangle prop with generous smoothing groups might weld down to 16,000 vertices, while the same prop with hard edges on every panel line runs to 26,000. That is a 60 percent jump in vertex buffer size for a mesh whose triangle count never moved. Check both numbers before you congratulate yourself on hitting a budget.
N-gons: almost always wrong
An n-gon is any face with five or more sides. Renderers triangulate them with a fan or ear-clipping algorithm whose output you cannot predict, and different applications choose differently. On a perfectly flat surface an n-gon is harmless — a flat hexagonal panel triangulates identically no matter how you slice it. On anything curved, or anything that deforms, it is a defect waiting to surface at the worst moment.
The compromise most studios settle on: n-gons allowed on flat, non-deforming, non-subdivided surfaces during blockout; zero n-gons in anything delivered. Triangles allowed anywhere they do not deform and do not subdivide.
Poles, and where to hide them
A pole is a vertex where a number of edges other than four meet. A valence-3 pole (N-pole) and a valence-5 pole (E-pole) are unavoidable — you cannot cover a sphere in quads without them, and you cannot merge a dense area into a sparse one without them either. The rule is not “no poles”, it is “no poles on a smooth curved surface that catches a highlight”.
Put them in flat regions, inside concave corners, under a strap, behind a panel line. On a face, poles belong at the wing of the nose and the corner of the mouth, where they are hidden by anatomy and by the deformation itself. A valence-6 or higher pole on a car hood will show as a dimple in the reflection under every environment map you try.
Pick by use case
| Asset | Topology | Typical budget | Why |
|---|---|---|---|
| Background prop (crate, barrel) | Triangles | 300–2,000 tris | Never deforms, never subdivides, never edited |
| Hero static prop | Triangles, quad-ish flow | 8,000–30,000 tris | Bake quality benefits from deliberate edges |
| Hard-surface with bevels | Quads + supporting loops | 15,000–60,000 tris | Bevels need controlled loop spacing |
| Deforming character | Quads, mandatory | 10,000–35,000 quads | Edge loops drive skinning |
| Cloth or sim mesh | Quads, even density | 5,000–20,000 quads | Solvers assume regular spacing |
| Subdivision base mesh | Quads, all-quad | 2,000–10,000 quads | Catmull-Clark needs clean input |
| Scan or photogrammetry | Triangles | 50,000–500,000 tris | No flow to preserve; decimate and bake |
| 3D-printed part | Triangles, watertight | Whatever holds tolerance | Slicers only read triangles |
Remember the conversion when comparing numbers: one quad is two triangles. A “12k character” quoted in quads is a 24,000-triangle character in your engine’s stats panel, and the two figures get mixed up constantly in asset store listings.
What generated meshes give you
Meshes produced by text-to-3D or image-to-3D come out of an isosurface extraction step — marching cubes, dual contouring, or a variant — followed by decimation. The result is uniformly sized triangles with no edge flow whatsoever. It is dense where the shape is flat and dense where the shape is curved, because the algorithm has no concept of feature importance.
For a static prop this is a non-issue. Drop it in the scene, bake if you need to, ship. For a character it is unusable as-is, and no amount of decimation will fix it, because decimation removes triangles without ever creating the loops you need. The honest workflow is to treat generated output as a sculpt: it is your high-poly reference, and you retopologise over it.
MeshyFlix’s remesh option will convert generated output to a quad-dominant mesh at a target polycount, which handles organic shapes and simple props well. It is not a substitute for hand retopology on a hero character that needs facial deformation, and anyone who tells you otherwise has not tried to weight-paint the result.
Do and don’t
- Do triangulate before baking, and ship the triangulated mesh.
- Do put three edge loops minimum through every joint that bends past 90 degrees.
- Do check topology with a matcap or a clay shader, not a textured view — texture hides shading errors that a flat grey surface exposes immediately.
- Do keep quad flow on anything a human will edit later, even static props, if the project runs longer than a few months.
- Don’t spend a day retopologising a background crate. Decimate it and move on.
- Don’t ship n-gons on curved or deforming surfaces, ever.
- Don’t put poles on smooth reflective surfaces such as vehicle panels or helmet shells.
- Don’t compare quad counts to triangle counts without converting; the factor of two has embarrassed better artists than us.
Where to go from here
Open the last asset you delivered and turn on face-type highlighting — in Blender that is Select Similar by face sides, or the Statistics overlay with the mesh in edit mode. Count your n-gons. If any of them sit on a curved surface, that is your first fix.
Then take a normal-baked asset you were never quite happy with and check whether the mesh you baked from was triangulated identically to the mesh you exported. If it was not, re-bake with a triangulated low-poly and compare the two normal maps in difference mode. The diagonal artifact you have been blaming on your baker will be sitting right there.