Retopology without the pain: when to remesh and when to leave it

A studio spends two days unwrapping a 1.4-million-triangle scanned stone lion, then decides the mesh is too heavy and runs an auto-remesh down to 9,000 triangles. The UVs come back as confetti. Both days are gone, and the person who ran the remesh genuinely did not know that would happen, because the tool was sitting in the same menu as “Decimate” and the difference was never explained.

Three distinct operations get called “retopo” in casual conversation, and they behave nothing alike. Getting them straight is most of the battle.

Three operations, three outcomes

Decimation collapses existing edges using a quadric error metric, removing the ones that change the silhouette least. It keeps your triangles, roughly keeps your UVs (interpolated across the collapse, so they degrade but do not shatter), and takes seconds. It produces no new topology and no edge flow.

Remeshing throws away the surface and rebuilds it — voxel remeshing rebuilds it as uniform triangles, quad remeshing (Quadriflow, ZRemesher, QuadRemesher) rebuilds it as a quad grid aligned to curvature. It destroys UVs completely, because the vertices those UVs belonged to no longer exist. It takes seconds to a couple of minutes.

Retopology is a human deciding where every loop goes, usually with a snap-to-surface tool. It produces topology built for a specific purpose: loops through joints, poles hidden in flat regions, density where deformation happens. It takes hours.

The question that decides it

Does the surface deform? That is nearly the whole decision.

If nothing bends — a rock, a crate, a lamp, a scanned statue, a piece of architecture — you do not need retopology. You need a triangle count and a normal map. Decimate, or generate at the target polycount to begin with, bake the high-poly detail, ship it. Someone spending a day building loops on a background barrel has confused craftsmanship with output.

If it bends, auto-tools will not save you. A quad remesher produces even quad flow that looks tidy and sits wrong: no loop concentration at the elbow, no radial arrangement around the mouth, no density falloff toward the fingertips. You can weight-paint your way through a shoulder like that, but the deltoid will collapse the moment the arm goes above horizontal and you will spend longer fixing it than you would have spent doing the retopo properly.

The middle case is subdivision. If the asset will be subdivided for render, you need clean all-quad topology but not necessarily deformation-aware topology, and quad remeshers are genuinely good at that.

Order of operations, which is where most damage happens

The sequence is not negotiable:

1. Get the high-resolution surface — sculpt, scan, or generation output.
2. Build the low-poly: decimate, remesh, or retopologise.
3. UV unwrap the low-poly.
4. Bake from high to low: normal, ambient occlusion, curvature, and a base color transfer.
5. Texture on the low-poly.

Every disaster in this area comes from doing 3 or 5 before 2. Unwrap before you remesh and the unwrap is gone. Texture before you remesh and the texture is gone, along with any hand-painted work. If you must change topology after texturing, you are not remeshing — you are re-baking, and you need the old textured mesh preserved as the bake source.

Baking the detail back

The bake is what makes an aggressive reduction survive. A 6,000-triangle prop carrying a normal map baked from a 400,000-triangle source reads as far more detailed than a 40,000-triangle prop with no bake at all.

Settings that matter, with numbers that work in practice for a prop around 30 cm across:

• Ray distance / cage extrusion: 5 to 20 mm. Too low and you get black patches where rays missed; too high and nearby surfaces bleed into each other.
• Supersampling: 4× minimum, 8× for anything with fine engraving.
• Padding: 16 pixels at 2048², 8 at 1024². Insufficient padding shows up as bright seams once mip level 3 or 4 kicks in, not in your close-up preview.
• Match the triangulation between the mesh you bake on and the mesh you export, or you get a diagonal shading gradient across large faces.

Bake normal and AO always. Bake curvature too, even if you do not think you need it — it drives edge wear masks in Substance and Mixer, and regenerating it later means setting the whole bake up again.

Where auto-remesh actually fails

Hard-surface bevels get rounded. Quad remeshers align to curvature, and a 2 mm chamfer at a target edge length of 8 mm simply cannot be represented. The fix is edge weighting or guide curves, which most remeshers accept — or accepting that hard-surface work wants manual topology.

Voxel remeshing welds nearby surfaces. At a voxel size of 5 mm, any gap under about 5 mm closes. Fingers fuse, chain links merge, a hinge becomes solid. On a 30 cm object, 2 mm voxels give you roughly a 150³ grid and preserve most detail; if you need finer, expect memory to grow with the cube of the resolution.

Thin double-sided geometry disappears. Cloth, leaves, paper — anything without volume — comes back as either nothing or an inflated slab. Separate these before remeshing and handle them by hand.

Separate shells become one. If a sword and its scabbard are intersecting when you remesh, they will be a single object afterwards. Separate first, always.

Unwrapping what comes out the other side

Remeshing hands you a surface with no UVs, and how you unwrap it depends entirely on how the asset will be textured. For a baked workflow — one atlas, unique texels, everything painted in Substance — an automatic unwrap is genuinely fine. Modern auto-unwrappers hit 65 to 78 percent packing efficiency, which is within a few percent of a careful manual pack, and the extra islands cost you nothing once the bake is done.

For a tiling or trim-sheet workflow, automatic unwrapping is useless. Trim sheets need UV shells aligned to the sheet’s bands, straightened, and scaled to a consistent texel density. No automatic tool infers that intent, and this is a large part of why environment artists retopologise modular pieces by hand even when the pieces never deform.

Whichever path you take, check texel density before texturing. A 2048² map on a 2-metre wall gives about 1,024 pixels per metre; the same map on a 30 cm prop gives 6,827. Mixing those two in one scene is what makes some surfaces look soft next to others, and it is invisible in the UV editor and obvious in the game.

Realistic reduction targets

Note the last row. LODs reuse the parent’s textures, so decimation is the correct tool and remeshing would be actively wrong — new topology means new UVs means new bakes for every level.

Do and don’t

• Do keep the high-resolution source file forever. It is your bake source for every future change.
• Do generate at your target polycount when the tool allows it. MeshyFlix’s remesh action lets you set a triangle target and choose quad or triangle output, which skips a separate decimation pass.
• Do separate intersecting objects before any remesh operation.
• Do check the silhouette against the original from four angles after reduction. Silhouette errors are the ones a normal map cannot hide.
• Don’t remesh anything already unwrapped or textured unless you intend to redo both.
• Don’t retopologise static props by hand. It is the most common way junior artists burn a week producing nothing a client can see.
• Don’t trust an auto-quad result on a deforming character without test-posing it to the extremes of its range first.
• Don’t reduce below the point where the silhouette breaks and expect the bake to rescue it. Normal maps fake lighting, not shape.

Where to go from here

Pick the heaviest asset in your current project and answer one question in writing: does any part of this surface deform? If no, decimate it to your budget, bake a normal map from the original, and be done inside an hour. If yes, block out the retopology for the deforming regions only — the joints, the face — and decimate the rigid parts. Most characters are 20 percent deformation and 80 percent boots, belts and armour plates that never bend at all, and treating the whole model as one problem is why retopology has a reputation for being miserable.

Then set up your bake once and save it as a template. Cage distance, supersampling, padding, output resolution. The setup is what makes baking feel expensive; the bake itself takes 30 seconds.