Convert GLB to VRM — a rigged .glb in, a VRM 1.0 out

VRM is glTF 2.0 with a set of extensions on top. The mesh, the textures and the skeleton carry over unchanged. What the conversion has to add is meaning: a humanoid bone map (this joint is hips, that one is leftUpperArm), an expression set (which morph target is aa, which is blink), look-at settings for the eyes, and the avatar's metadata.

The bone map is mechanical once the skeleton is sane. The expression set is the hard part: a GLB from an AI generator or an auto-rigger almost never contains a single morph target for the face, so there is nothing to bind. Someone has to build the mouth shapes - which is why the usual answer is Unity, UniVRM and an afternoon of BlendShape clips.

• It has a skeleton. Open it in any glTF viewer and look for bones, or check that the exporter's skeleton or skin option was on. A GLB with no skin is a statue - Riggle refuses it at upload and tells you so.
• It is humanoid. Two arms, two legs, a head on a neck. Stylised proportions are fine; quadrupeds are not, because VRM's bone map has nowhere to put them.
• It is not meshopt-compressed. Files run through gltfpack or an "optimise for web" step often are; re-export without mesh compression.
• The mouth is open, ideally. Riggle builds its best visemes from a mouth that was modelled open, with teeth and depth behind them - see why an open mouth decides your visemes.

Units, facing direction, A-pose versus T-pose and a leftover animation clip do not matter: Riggle returns the model to its bind pose, measures and rescales it, turns it to face forward and bakes an A-pose to a T-pose before anything else happens.

Upload the .glb. Riggle maps the skeleton onto the VRM humanoid bones - it recognises the Mixamo, Tripo, VRoid and Character Creator naming conventions and falls back to matching by name pattern and hierarchy - and renders a straight-on view of the face. Then it asks you for five things:

1. Trace the lip boundary - four or more clicks around the outer edge of the lips, with a point at each corner and the bottom edge on the crease above the chin.
2. Click inside the mouth opening.
3. Tag the upper lip, then the lower lip - one click each.
4. Click the centre of each eye - this pins the blink.

Press Rig this model and wait about a minute. If the result is not right, mark it again - re-marking is free and unlimited.

The download is a VRM 1.0 with the humanoid bones mapped and the expression set bound: the five visemes (aa, ih, ee, oh, ou), blink with independent left and right, happy, angry and surprised, and eye look-at when the rig has eye bones. Before you pay anything, Riggle loads the file in a real VRM renderer, drives every expression, and shows you a video of the whole set - an expression that exists but does not visibly move is flagged rather than shipped.

What it does not do: it does not rig an unrigged mesh (the skeleton has to be there already), it does not build the 52-shape ARKit "perfect sync" set - see why perfect sync is a different job - and it cannot invent teeth and a mouth cavity that were never modelled. A character whose mouth was generated closed still converts, but its visemes are a softer procedural approximation.

The steps above are the same for every source. The differences are all upstream, in how you get a rigged GLB out of the tool you used: Meshy and Tripo (the AI image to VRM guide) export one directly; Mixamo and most marketplaces hand you an FBX, which needs one conversion to GLB first. If you already have a VRM and only want the face fixed, upload the .vrm itself - see adding visemes and blinks to a VRM.