In short: AI retopology takes the dense, triangulated output of a generator or scan and rebuilds it as quad-dominant geometry at a density you choose. The pass clears n-gons and non-manifold edges, follows the form with edge flow that deforms and subdivides predictably, and lets you set targets by destination - a few thousand triangles for a game prop, tens of thousands for a hero render, a closed manifold for printing. It is a prompt-driven step, so a repair is a sentence: fix the topology, reduce to 10k vertices, keep the shape. The mesh stays in the scene throughout, and the result is ordinary geometry you can keep editing by hand. Silhouette is preserved; at aggressive budgets the smallest details simplify first, so check the details that matter before export, and keep the original mesh until the rebuilt version has passed review.
Why AI meshes need a rebuild
Generators optimise for silhouette, not for edge flow. The output is typically dense, unevenly triangulated and - depending on the source - dotted with n-gons and non-manifold edges. That is fine on a still image. It becomes a problem the moment the asset has to deform, be simulated, or fit a real-time budget. The same applies to photogrammetry scans and many store-bought assets.
What the repair pass changes
- n-gons and triangles to quads - quad-dominant topology that follows the form, which is what deformation and subdivision expect.
- Density on request - a target vertex count ("reduce to 10k") turns a high-detail mesh into an engine budget.
- Non-manifold edges and flipped faces - repaired as part of the pass, so slicers and renderers stop complaining.
- Materials and UVs stay attached - maps remain on the object; after a big rebuild, re-check seams and re-wrap problem islands.
The four-step repair loop
01 - Inspect. Turn on wireframe and the statistics overlay. Find the dense regions and the edges that cannot survive - before changing anything.
02 - Rebuild with a target. State the destination, not just the command: "game-ready prop, clean quads, around 8k triangles" gives the pass something to aim at.
03 - Verify. Check face orientation, shading and the silhouette from four views. Open the mesh at a joint or an edge and confirm the loops make sense.
04 - Finish. Re-unwrap islands that stretched, re-check materials, then export to the destination format.
The checks a rebuilt mesh has to pass
The pass is done when the checklist passes, not when the wireframe looks tidy. Run these before the asset leaves the scene:
- Budget. Face and vertex counts match the target you set, and the total fits the destination - engine, render or print.
- Shading and normals. No flipped faces or shading seams; smooth and flat regions read the way the reference does.
- Silhouette. Compare four views against the original: the shape should match, with only the finest details simplified at aggressive targets.
- Deformation. Bend the parts that will move - test elbows, knees and hinges on the new topology, not just the rest pose.
- Printing. For slicers, check manifold edges, wall thickness and a closed volume before STL export.
Density targets by destination
| Destination | Topology | Starting range | Notes |
|---|---|---|---|
| Real-time / game engine | Quad-dominant, or decimated to a budget | 5k - 25k polygons is a common starting range | Static props can ship leaner as triangles; characters keep the quads |
| Rigging and animation | Quad-dominant with even loops | 10k - 60k | Edge flow matters more than the count - keep loops around joints |
| Stills and archviz renders | Density where the camera sees it | Match the shot, not a number | Detail on camera-facing forms; decimate what the frame never sees |
| 3D printing | Manifold and watertight | Solidity over count | Verify in a slicer before the final print |
The ranges are starting points, not rules - the right number is the one the destination actually needs.
It works on meshes from anywhere
The pass runs in-scene, so the source does not matter: a mesh generated from text, reconstructed from an image, a scan, or a purchased asset all take the same route - inspect, rebuild, verify, finish. From there they join the same pipeline as everything else: texture, animate, render, export.
What still needs an artist
- Hero-character edge flow. Deliberate topology for facial deformation - loops around eyes and mouth - is an artistic decision, not an automatic one.
- Precise performance budgets. When a project specifies exact triangle counts per LOD, an artist signs off on the final distribution.
- Where the density belongs. Choosing which forms deserve detail and which can be decimated is still a judgement call - the pass executes it quickly.