Retopology for AI 3D Models: Which Method to Use

Four questions that pick the route

Work down the list. The question that stops you is usually your route.

1. Does the engine need fewer triangles at all? For dense static meshes the honest route may be none: do you still need retopology with Nanite and LODs.
2. Will it bend? A rigged character, a creature, cloth with a skeleton: plan on manual topology for those areas. The Blender manual says automatic remesh tools generally do not produce topology that lends itself to deformation, and that “no perfect automatic tools exist” for it right now.
3. Is the surface broken? Intersections, holes and floaters: rebuild or repair before anything else. Blender’s manual is direct on this: the voxel remesher produces no inner self-intersecting geometry, and the quad remesher is not a replacement for it “because it doesn’t clean up intersecting geometry”.
4. Is it only too dense? If the mesh is sound and you just need a lower count, use reduction, not a remesh. The manual recommends Decimate over both voxel and quad remeshing for reducing the face count of a mesh that otherwise has no geometry problems.

The routes

Decimate (Blender). Collapse mode takes a Ratio (1.0 leaves the mesh unchanged, 0.0 removes all faces) and has a Triangulate setting, Symmetry and a Vertex Group that controls what is decimated. With Triangulate off, untouched quads and n-gons can leave the final count higher than expected. Two other modes have narrow jobs: Un-Subdivide is intended for meshes with mainly grid-based topology, and Planar for forms made mainly of flat surfaces. Decimate is the simplest route. Nothing in its documented options repairs a mesh, so expect it to carry existing holes and floaters along with it. It also has no Edit Mode preview, per the manual, so check it in Object Mode. StraySpark’s pipeline post advises Un-Subdivide or Planar rather than Collapse as a pre-retopology step; given what the manual says those modes are for, confirm it on your mesh.

Voxel remesh, then QuadriFlow. Voxel Remesh rebuilds the surface on a voxel grid; QuadriFlow (Properties > Data > Remesh) builds a quad mesh at a target Ratio, Edge Length or Faces. Voxel output has uniform topology, and the manual’s own use list covers cleaning up a mesh for sculpting or 3D printing and generating a simplified stand-in for physics. Final game topology is not on that list. The Voxel Remesh page says it removes all mesh data layers from the original, so UVs are gone and you rebake (its Preserve options include Color Attributes, which reprojects them; check the result). StraySpark, a MeshPrep vendor, adds that a voxel grid cannot represent anything thinner than one voxel, so straps, antennae, cloth edges and fingers can fuse or disappear. It also reports that QuadriFlow rejected both Meshy files it tried (a 271,174-triangle robot and a 354,724-triangle character), and that the voxel path on the robot produced 10,024 quads (20,048 triangles) against a 20,000 budget; for the character it gives 7,931 quads but no budget. Those are two files from one generator, not a failure rate.

Paid auto-retopology add-ons. These automate the remesh step or use their own algorithm. Prices and documented features are on Blender retopology add-ons compared. Quality claims belong to the vendors.

Manual retopology. You draw the low-poly surface over the dense mesh, using Blender’s retopology overlay, Poly Build tool and snapping (the three items the manual’s retopology section lists). It is the route for faces, joints and anything that deforms, and it costs the most time. The Blender manual’s retopology page describes snapping, not the Shrinkwrap modifier; if you use Shrinkwrap to project a hand-made mesh onto the original, treat that as practitioner method, not documentation.

A hybrid is worth considering. Orange Turbine, which sells a manual tool, says auto-remeshers are good for a quick base and that plenty of artists use both: auto for a first pass, then its tool to clean up objects that need focused attention. A Tripo blog post (Tripo sells automatic retopology) was noted as describing the same split, but we could not re-open that page at our final check, so treat that as unconfirmed. Neither supplies a measurement.

What “AI retopology” actually refers to

The phrase covers several different things, so read the feature, not the label:

• Built into a generator. Meshy’s web app has a Remesh step with a Target Polycount; its docs say Remesh bakes textures onto the new mesh, so textures survive.
• Automatic remeshers marketed as AI. The label does not tell you what the algorithm does. Ask what the tool is documented to do with thin parts, sharp edges and existing UVs.
• What to check on any of them: does the output keep or rebake textures, can you set a target count, are the results quads or triangles, and can you get the original back. Meshy’s Unwrap UV docs carry a useful example of a catch: running it converts quad topology to triangles, and it saves a separate model.

We did not come across a published, independent failure rate for any of these on AI meshes.

What each route keeps and loses

KatsBits, a Blender tutorial site, was noted as arguing that AI UV maps are about as dense as the mesh and that Decimate degrades them along with it; we could not re-open that page at our final check, so treat it as unconfirmed. What the mechanics do establish is that a changed mesh needs its texture checked against its UVs. For anything close to camera, our recommendation is new topology, new UVs, and a bake.

Choosing by asset type

This table is judgment from the documentation, not a measured ranking.

Why AI meshes change the choice

StraySpark says QuadriFlow “wants a closed, manifold surface with consistently oriented faces” and that hundreds of open shells are the opposite of that. If you are going straight to QuadriFlow, fix the shell first with repairing holes, floaters and broken pieces. A voxel pass depends on this less, since StraySpark says it rebuilds a closed surface from the outside of the volume instead of repairing the old one. Whichever route you choose, keep the original dense mesh: reduction without a bake discards the surface detail, and the normal map bake needs it as its source. New UVs are on UV and texture fixes.

After you choose

Check the result against your target’s limits, then make LODs from the final topology. We suggest three checks before calling a mesh done, whatever tool you used: bend it with a simple rig, apply a Subdivision Surface modifier and look for pinching (the Blender manual notes poles can pinch under it), and unwrap it to see whether a clean low-stretch layout comes easily. Skip the first if it never moves.

Stop retopologizing when the silhouette holds and the engine target is met; polishing topology on an asset that will be seen small is time with no visible return.

What is not known

How often each route fails on each generator’s output has not been measured here. The planned measurement runs each route on raw outputs from several generators and logs face count, rejection messages, UV and bake loss, and minutes. Until it exists, treat any claimed success rate, including vendor ones, as unverified.

This page is part of the Cleaning Up AI 3D Models section.