VRM

VRM avatars, including everything VRoid Studio exports, carry their physics as spring bones and their faces as expressions. Both formats are read: spring bones become Basis jiggle physics, and expressions become HVR Vixxy controls.

UniVRM is needed to import a .vrm file, not to convert one. A VRM avatar handed over as a prefab arrives with its components as missing scripts, and the data is read from the file either way.

Bringing in a .vrm file

Install UniVRM and drag the file into the project. Use 0.131.2 or newer: 0.131.0 does not compile on the Unity version Basis targets.

Drag the imported avatar into a scene and convert it. It does not need unpacking, and its assets do not need extracting. A prefab saved from it without unpacking converts too, and the report names the file it was read from as source.modelRead. A .vrm is binary, so there is no text to read; its spring bones, constraints, expressions, licence and eye offset are read from the components instead.

UniVRM's data components stay on the converted avatar. They are not on the Basis allow-list, so the Basis build strips them.

The two formats

VRM 0.x puts one component on the avatar carrying a group of chains and a single set of parameters for all of them. Each root bone it names becomes a jiggle rig.

VRM 1.0 puts a joint on each bone and lists which joints make up which chain on the avatar's own component. Each chain becomes a jiggle rig rooted at its first joint.

UniVRM 1.0 converts a 0.x file to 1.0 components as it imports, so a .vrm brought in that way reads through the 1.0 path whichever version it was written as. UniVRM 0.x instead writes a prefab beside the .vrm as it imports; convert that prefab. The 0.x path also covers prefabs made with an older UniVRM.

UniVRM's runtime is removed

A converted VRM keeps UniVRM's data components and loses its runtime drivers: Vrm10Instance for 1.0; VRMSpringBone, VRMBlendShapeProxy and the look-at components for 0.x. They would rewrite every blendshape and run their own physics each frame over the conversion. Basis strips them at build anyway. Undo restores them: vrm.runtimeRemoved.

What carries across

VRM	Basis jiggle
Drag force	Drag, on the same 0 to 1 scale
Joint radius / hit radius	Collision radius, both in metres
Gravity power and direction	Gravity multiplier, the vertical part, as a fit: VRM adds a per-step force
Collider groups	The rig's colliders
Stiffness force	Stiffness, as a fit rather than a conversion
Centre transform	Ignore root motion, fully, measured at the rig's root bone
Cone angle limit (1.0)	Angle limit, capped at 90 degrees. Hinge and spherical limits are dropped

Parameters that differ along a chain become a curve. VRM 1.0 carries a value per joint, and Basis evaluates its parameters over distance from the chain root, which is the same axis. A chain that stiffens toward its tip keeps that shape instead of being averaged.

Bones the chain does not name are excluded. A VRM spring names the bones it moves; a jiggle rig simulates everything under its root. An accessory hanging off a hair bone would start swinging, so it is excluded to leave it as still as VRM left it.

Expressions

A VRM expression is a set of blendshape weights. The avatar's expressions become one Vixxy selector named Expression: Neutral first, then one choice per expression. Every shape any expression touches is set at every choice, at the expression's weight or zero, as the VRM specification applies them. Neutral is every shape at zero; the avatar's own neutral preset is not applied, since VRM applications do not apply it either.

Material colour and texture offset changes carry over as Vixxy material properties under MToon's names, on every renderer that uses the material alone. A renderer that also carries other materials is left alone: vrm.expression.materialShared. A material no renderer uses is left out: vrm.expression.materials.

Two things the spec allows are reported rather than converted. An expression that is not isBinary can be worn at any strength; a choice is all or nothing: vrm.expression.continuous. An expression's overrides, which block or attenuate blink, gaze or lip sync while it is worn, have no counterpart, since Basis keeps those running: vrm.expression.override.

The lip sync shapes, blinking and looking around do not become controls. Basis drives those itself, and a menu item would fight it.

The five vowels and blink go on the BasisAvatar, so a converted VRM talks and blinks. Blink keeps every shape the expression moves, on one mesh. Basis has fifteen visemes and VRM five: aa, E, ih, oh and ou are filled, the consonants stay unset. A viseme holds one blendshape, so an expression that moves several is reported.

VRM refers to a blendshape by its position in the mesh rather than by name, so each one is looked up on the renderer it names. A shape that is no longer there, usually because the mesh changed after the expression was authored, is reported rather than guessed at.

The avatar's licence

Every VRM states who may wear it and what may be done with it. The window shows all of it before you convert: the title, the author, who may wear it, whether it may be changed or passed on, whether commercial, violent, sexual, political or hateful use is allowed, whether credit is required, and the licence it names.

A permission the format has no field for is left out rather than guessed at. VRM 0.x has no political or antisocial fields, and no redistribution or credit fields; VRM 1.0 has all of them.

The licence is a warning when it forbids changing the avatar or limits who may wear it. Nothing is blocked. Converting changes an avatar, and using it on Basis is a use; check the licence allows both.

Constraints

VRM 1.0 has three node constraints, and each becomes a Basis constraint on the object it sits on.

• Rotation becomes a Basis rotation constraint. VRM copies how far the source has turned from its rest pose; a Basis constraint takes the source's rotation itself, with a rotation offset so the authored pose holds while the source rests. The two differ if the source's rest pose changes.
• Aim becomes a Basis aim constraint, pointing the same local axis at the same source. Basis also holds an up direction, which VRM does not state, so the scene's up is used and the roll around the aim may differ.
• Roll copies the source's rotation about one axis alone, which nothing in Basis or in Unity's own set does. It becomes a rotation constraint limited to that axis, which follows the source's rotation rather than its roll.

All three are reported, since none of them is exact.

Where the eyes sit

Both formats record the point the camera sits at as an offset from the head bone. Basis stores the same point as the avatar's eye position, so it carries across: a converted VRM avatar has a correct eye height instead of one Basis has to guess.

A VRM also marks renderers to hide from the wearer or show only to them. Basis hides the head bone and everything under it in first person, which covers a face skinned to the head and hats parented to it. The flags are reported rather than converted. If something still blocks the camera, add a Basis Head Chop naming it.

What does not

• The centre transform. VRM simulates relative to it so hair does not lag behind a moving avatar. Basis has no equivalent, though its own root motion handling covers some of the same ground.
• A plane collider's normal, when it is not its transform's Y axis. Basis planes face that axis, so the plane is written facing Y and reported as vrm.collider.planeNormal.
• Inside colliders, which hold bones within a shape. Basis only pushes out, so they are written as ordinary colliders and reported. Remove them if the result looks wrong.
• Where the avatar looks. VRM aims the eyes with curves or an expression; Basis drives gaze from the eye bones. Only the eye offset carries over. VRM 0.x look-at components: vrm.lookAt. An expression-type look-at has no eye bones, so the eyes stay still: vrm.lookAt.expression.
• How far the eyes may turn. VRM 1.0 states a limit, 10 degrees on a VRoid export. Basis turns eyes up to 25 degrees with no setting on the avatar, so a large eye shows white past the model's limit: vrm.lookAt.range.
• The avatar's metadata: its title, author and permissions.