Retopology: From Raw Scan to Real-Time Mesh

A 3D scan produces a gorgeous file that is completely unusable as it comes. A client sends it over, proud, with the mesh of a facade flown by drone or a statue scanned by hand. 4 million triangles, every crack captured, and the assumption we project onto it tomorrow. Except no media server runs that file at 60 frames a second, no real-time engine swallows it without choking, and the first time it opens in a mapping tool the machine crawls or dies.

Retopology is the step that turns that block of triangles into a clean, light, usable mesh. You start from the scanned surface and rebuild a new topology over it, lighter, with edges that actually follow the shape. This guide explains why the raw scan is unusable, and how to produce the low poly that isn't.

Short version:

• a raw scan mesh is "triangle soup": too dense, random density, no clean edge loops
• decimation cuts the triangle count but never fixes the topology; retopology rebuilds it
• automatic retopology (Quad Remesher, Instant Meshes) for static sets, manual for anything that deforms
• detail lost in the process comes back through a normal map baked from the high poly
• and sometimes, for mapping, the right answer is to not retopologize at all

Why a raw scan mesh is unusable

A scan, whether from photogrammetry or a laser scanner, knows nothing about your object. It knows points in space. The algorithm joins those points into triangles as close to the measurement as possible, with no idea what is a wall, a fold of fabric, or a hard edge. The result has a name in the trade: triangle soup.

In practice this mesh stacks four faults that make it dead on arrival for real time:

• The count. An honest field scan lands between 2 and 20 million triangles. A game or mapping set runs two or three orders of magnitude lower.
• Random density. The scan puts as many triangles on a flat wall as on a molding. Wasted where it's flat, never aligned where it curves.
• No edge loops. The triangles scatter in every direction. You can't select an edge and follow it, you can't deform cleanly, and the UV falls apart.
• Holes and normals. A scan always has blind spots: undersides, recesses, surfaces too dark or too shiny. Holes, flipped faces, inconsistent normals. The engine no longer knows which way is out.

While those four are there, you don't have a 3D model. You have a very precise, very heavy measurement. For where that mass of points comes from before meshing, see the point cloud guide.

Decimation is not retopology

First instinct for everyone: "I'll decimate and I'm done." Decimation cuts the triangle count by merging the smallest ones. Going from 4 million to 200,000 in one click works, the counter drops. The mesh is still triangle soup, just coarser.

Retopology does something else. It rebuilds a new surface, in clean quads, over the scanned geometry. The new points snap to the scan, the edges follow the shape, density goes where the detail needs it and nowhere else.

Simple rule: if the mesh will never move and only serves as a distant visual reference, decimation can be enough. The moment you need clean UVs, a bake, deformation, or a poly count you control to the polygon, you retopologize. The Blender manual on remeshing covers both tool families on the open-source side.

Automatic or manual retopology

Two roads, and the choice comes down to one question: will the object deform.

Automatic. You give the high poly, a target count, and the tool rebuilds a quad mesh in seconds. Quad Remesher by Exoside, which plugs into Blender, Maya, 3ds Max, Houdini and Modo, gives the best automatic quality I know. Instant Meshes is free, open source, and does honest work on sets. Blender ships QuadriFlow and Voxel Remesh natively. For a facade, a statue, a rock, any set that stays still, auto handles 80 % of the job. You clean two or three zones by hand and it's done.

Manual. You draw the new topology by hand over the scan, edge by edge. RetopoFlow in Blender, Quad Draw in Maya, or guided ZRemesher in ZBrush. It's slow, it's craft, and it's unavoidable the moment the object has to animate or bend. A face, a joint, cloth: no automatic tool today places the edge loops in the right spots for the deformation to hold. The topology of an animated mesh has to follow the flow of the muscle, not the measurement.

The classic trap is wanting to do everything by hand on principle. On a mapping set that never moves, retopologizing a statue edge by edge for two days when Quad Remesher outputs it clean in thirty seconds is time billed to no one. (I did it. Once. You learn.)

Polygon budget for real time

The real question isn't "as low as possible." It's "how much my engine holds at the actual viewing distance, on the machine running the show that night."

A few field markers, to adjust for your GPU and media server:

• a scan high poly sits around 2 to 20 million triangles
• a clean retopologized mesh for real time lives between 20,000 and 200,000 quads depending on its role
• an animated character usually frames between 80,000 and 120,000 quads
• a static mapping set, seen from far, can drop far lower: a few thousand polygons is enough if the detail comes from the texture

That last point is the most misread. On a projection, the viewer looks at the light and the texture, not the mesh silhouette at three meters. A set seen from across the square doesn't need the density of an object held to camera. Size for the viewing distance, not for a zoom nobody will do.

UV and baking normals

Retopology loses the fine detail: that's the point, you stripped it to go light. You get it back another way.

Once the low poly is clean, you unwrap its UVs, which is easy precisely because the topology is sound (unwrapping triangle soup is the nightmare). Then you bake: you project the scanned high poly detail into a normal map applied to the low poly. The silhouette stays light, the surface carries the relief as fake depth. Cracks, stone grain, seams: all of it returns as texture, without a single extra triangle.

That retopo-plus-bake pair is what makes the scan workflow viable for real time. The scan gives precision, the retopo gives lightness, the bake reconciles both.

When retopology isn't worth it

Let me kill the mood a little. On plenty of mapping projects, clean retopology is work nobody will pay for and nobody will see.

• The scan is just for alignment. If you scan a facade only to place your projectors in a 3D scene and check coverage, you don't need animation topology. A light decimated mesh works as an alignment proxy. Pretty quad flow won't move a single lux on the wall.
• You project onto the real object. When the final surface is physical and the scan only helps you prep, topology quality has zero impact on the projected result. You align to the real thing, not to your mesh.
• The set is far and static. Decimation plus a normal map, move on. Two days of manual retopo on an object forty meters away is budget spent on invisible finesse.

Retopology earns its keep when the object animates, comes to close-up, or has to live in a demanding real-time pipeline. The rest of the time, a well-made decimated proxy takes you further, faster.

Once you have that clean mesh, or even that simple proxy, it imports into Lumeo to place projectors around it, check overlaps and real lux on the surface, and validate the alignment before a single machine is rented. It's the bridge between the scan and the projection plan.

For the rest of the chain, from the clean file to the playback server, start from the scan to media server guide. And for the full picture, from survey to mapping, the 3D scanning for projection pillar puts it all back in order.

If you're staring at a 6-million-triangle scan, a show in three weeks, and a doubt about what actually needs cleaning, write to me. I've retopologized things that didn't deserve it, so you don't have to.