Rig Bake
Repair a 3D model
Weld split vertices, drop degenerate faces and stray islands, fill holes, flip inverted normals. GLB, GLTF, FBX, OBJ, DAE, STL, PLY, VOX or VRM — any size, in your browser, nothing uploaded.
Rig Bake
Weld split vertices, drop degenerate faces and stray islands, fill holes, flip inverted normals. GLB, GLTF, FBX, OBJ, DAE, STL, PLY, VOX or VRM — any size, in your browser, nothing uploaded.
Batch processing runs this exact tool over a whole folder and hands back one zip — no file limit, still nothing uploaded.
A model that renders fine can still be broken in ways that only surface later — when you try to unwrap it, boolean it, 3D-print it, or bake light into it. The usual suspects are all here: vertices split into duplicates at every UV seam, triangles with zero area, disconnected specks of geometry floating inside the mesh, holes where surfaces failed to close, and normals pointing inward on a handful of faces so the model looks inside-out from one angle.
AI generators produce all five, routinely, and none of them are visible in the preview render they hand you.
Weld duplicate vertices collapses vertices that sit at the same position, so the mesh becomes connected topology instead of a pile of loose triangles. Nearly every downstream operation depends on this.
Drop degenerate faces removes triangles with no area — they contribute nothing and make solvers divide by zero.
Remove stray islands deletes the small disconnected components AI generators leave behind: the specks, shards and interior fragments.
Fill holes finds open boundary loops and closes them. Essential if the model is headed for a 3D printer.
Flip inverted normals detects meshes whose faces are wound the wrong way and turns them around.
Most repair tools refuse to touch a skinned mesh, and for a good reason: welding vertices that carry different bone weights averages them, and a rig deformed subtly wrong is worse than a rig with a few stray triangles. That matters here more than usual, because a model out of Meshy or Tripo is typically one rigged mesh — skip it and you have skipped the whole file.
So two vertices are merged here only when they are interchangeable in every attribute they carry, bone weights included. And nothing in a repair pass moves or deletes a vertex, which is what leaves skin weights and morph targets exactly where they were.
The five fixes are not independent. Several of them are meaningless — or actively destructive — unless another has already done its work: point a hole filler at a mesh whose vertices were split at every UV seam and it finds an opening along every one of those seams, then stitches thousands of them shut.
A single pass with everything enabled runs them in an order where each one has something meaningful to work on. Running them one at a time, in whatever order occurs to you, does not — so leave all five on for a first pass.
Hole filling closes each boundary loop with new surface. That is what a slicer or a boolean needs, but the patch across a large opening is flat — it does not reconstruct the surface that should have been there. On the small holes that AI generators leave, you will not notice. On a model missing a whole region, you get a lid, not a repair.
It does not fix UVs. Welding is UV-aware, so seams survive rather than being averaged into smears, but a model with overlapping or unusable UVs comes out with the same UVs it went in with — rebuilding a map is a different operation with real trade-offs.
It does not reduce the triangle count, and deliberately so. Repair should leave you with the same model, working; cutting geometry is a decision with a visible cost and belongs on its own page. Clean first, then decimate — a decimator handed a mesh full of duplicate vertices spends its budget collapsing edges that were never really there.
And it does not tell you what was wrong before you commit. If you want the diagnosis rather than the treatment, the model doctor reports on scale, bind pose, normals and UVs as well as geometry, and only applies the fixes that are unambiguous.
No — it means the model was already clean. The tool reports honestly rather than inventing work, and the file you download is unchanged.
It will weld the mesh and close its holes, which is most of what a printer needs. For the print-specific checks (watertightness, wall thickness), open the model in the editor and use the STL export, which reports on those directly.
Yes. Vertices are only welded when they match in every attribute they carry, bone weights included, so two vertices rigged to different bones are never merged. No pass deletes or moves a vertex either, which leaves skin weights and morph targets exactly as they were.
The welding pass fixes the common cases. For the aggressive repair — splitting edges shared by three or more faces — run the model through the UV re-unwrap tool, which rebuilds a manifold topology as part of its job.
Find the open boundary loops and close them. A hole is an edge sequence with a face on one side and nothing on the other, so it can be detected exactly rather than guessed at. The catch is that most apparent holes are not holes: a mesh whose vertices were split at every UV seam registers as open along every seam, and filling those produces thousands of degenerate slivers. A full pass on this page deals with that before it fills anything, so the real openings are closed while the false ones disappear.
Weld the duplicate vertices, remove the interior debris, and close every boundary loop — which is what a full pass on this page does. Watertight means the surface encloses a volume with no gaps and no self-intersections, and a slicer cannot decide what is inside a mesh that fails it. Two caveats: a filled hole is watertight but flat, so a large opening comes out capped rather than reconstructed; and this does not check wall thickness, which is the other thing a printer will reject you for. The STL export reports on printability directly.
An edge shared by more than two faces. Real surfaces do not work that way — every edge of a physical object has exactly two faces meeting at it — so a non-manifold edge means the mesh describes something that could not exist. Anything that has to reason about inside and outside fails on it: booleans, slicers, physics colliders, most unwrappers. They usually arise from geometry that was merged carelessly or from a generator stitching two surfaces along a shared seam. Welding resolves the common cases here; for the rest, a full re-unwrap rebuilds manifold topology from scratch.
Because rendering wants them. A vertex carries a position and also a UV coordinate and a normal, and at a texture seam or a hard edge the same position needs two different UVs or two different normals — so the exporter splits it into two vertices at identical coordinates. That is correct for a GPU and wrong for every geometric operation, which sees a disconnected mesh. Welding merges only vertices that match in every attribute, which restores connectivity without collapsing the seams that had a reason to exist.
Almost never, and where it does, that is the point. Welding, dropping zero-area faces and removing interior debris are all invisible in a render — they remove things that contributed nothing. Hole filling adds surface where there was a gap, which is visible only if you could see the gap. Flipping inverted normals changes the appearance deliberately, turning a hollow-looking region solid. Nothing here moves a vertex or repositions geometry, so the silhouette you had is the silhouette you keep.