In short: generate an original character in Blender, finish the mesh and humanoid rig, refine the deformation and create the expression controls. Export through a current VRM Add-on for Blender release, then test the VRM in a viewer before calling it finished. Blender AI compresses the modelling and body-rigging work. It does not turn incomplete Blender data into a valid VRM file by itself.
That distinction matters because VRM is both a 3D file and a runtime contract. The mesh first has to survive lighting, deformation, texture packing and scale checks. The avatar then needs a correct humanoid bone map, readable expressions, material names, metadata and optional spring motion. A beautiful character that fails one of those contracts is still an unfinished avatar.
What is a VRM avatar in Blender?
A VRM avatar is a glTF-based 3D character with added humanoid, expression and secondary-motion metadata for avatar applications. VRM is designed for portability. The same file can contain a character body, skeleton, face expressions, look-at configuration, spring bones, thumbnail, license metadata and material information.
In Blender, those elements are separate pieces of ordinary scene data. A body is a mesh. A skeleton is an armature. A smile is usually a shape key or bone-driven expression. Hair motion may be a SpringBone chain. The VRM specification defines how those pieces are written into a portable avatar file.
| VRM element | What it controls | Blender source |
|---|---|---|
| Humanoid rig | Body movement and retargeting across applications | Armature with a valid humanoid bone mapping |
| Expressions | Blink, mouth shapes, emotions and look-at response | Shape keys or mapped facial bones |
| Look-at | Eye direction toward a runtime target | Configured eye bones and a look-at mapping |
| SpringBone | Hair, tails, skirts and accessory follow-through | Spring joint chains with colliders and stiffness settings |
| Metadata | Avatar identity, thumbnail, permissions and attribution | VRM Add-on metadata fields |
Can Blender AI make a VRM avatar?
Yes, when “make” means create the character and prepare it for export. Describe an original character in Blender AI using an A-pose and a clean silhouette. Generate the body first, then separate hair, clothing and accessories when they need independent weighting or materials. Check the source at thumbnail size before investing in rigging.
Use a prompt that names the constraints that affect export:
| Weak prompt | Production prompt |
|---|---|
| “Anime girl avatar” | Original stylised humanoid avatar, neutral A-pose, separate hair shells, simple layered clothing, readable eyes, hands and shoes, under 45,000 triangles |
| “Make it ready for VRM” | Humanoid character at one metre scale, clean silhouette, applied transforms, mouth and eye geometry suitable for shape keys |
| “Add a flowing coat” | Separate coat panels with deformable cloth near the body and optional rigid accessories, keeping shoulders and hips accessible for weight painting |
The output remains an editable Blender asset. If the eyes, mouth or eyelids do not have room for expression shapes, fix that geometry before the rig is final. If the clothing is fused into the body, decide which parts should deform and which should be rigid before painting weights.
How does the VRM avatar pipeline work in Blender?
The useful pipeline has six stages: generate, clean, rig, deform, express and export. Each stage creates a gate for the next. Skipping the cleanup gate creates bad weights. Skipping the expression gate creates a body-only avatar. Skipping the export test creates a file that looks finished in Blender but fails in the application that matters.
01 - Generate the character. Start with an original design and a neutral pose. Keep the face, hands and clothing readable enough to survive close inspection.
02 - Clean the scene. Apply scale and rotation, merge hidden duplicate geometry, inspect normals and consolidate materials. Give the body a predictable origin and real-world scale.
03 - Build the body rig. The Blender AI rigging workflow can create a humanoid armature and initial weights. Check joint placement before touching the vertex groups. Moving a bone after extensive weight painting wastes the work around that joint.
04 - Refine deformation. Test both arms overhead, a full elbow bend, a deep squat and a twisted torso. Follow the weight-painting correction guide when shoulders collapse, hips pull or rigid armour bends like rubber.
05 - Create expression controls. Build the face and mouth controls the destination app expects. Test them at the final face topology, not on a temporary head.
06 - Export and verify. Map humanoid bones in the VRM panel, configure expressions and SpringBone, export, reimport, then open the file in a compatible viewer.
Which humanoid and weight checks must pass before VRM export?
The rig must have one clear bone for every required humanoid role and stable deformation across the body. VRM applications rely on semantic roles, not informal bone names. A bone called “upper arm” can be mapped correctly, but two bones competing for the same humanoid slot cannot.
| Check | Pass condition | Failure caused by |
|---|---|---|
| Humanoid map | Every required slot points to the intended unique bone | Missing bones, duplicate roles or mirrored naming errors |
| Rest pose | Arms, legs and spine map consistently in a neutral pose | Pose stored as the rest pose or an uncleared animation offset |
| Scale | Character height and object transforms are predictable | Unapplied scale, mixed unit assumptions or oversized props |
| Weights | Shoulders, hips and knees hold volume in extreme tests | Unnormalized groups, overlapping shells or missing influences |
| Materials | Textures are packed or referenced correctly and survive export | Missing image paths, unsupported shader nodes or excessive slots |
Do not judge the rig in one polite A-pose. The avatar should survive the motion your application is likely to produce: dancing, reaching, bowing, crouching and turning the head. A calm test pose can hide a collapsed armpit or a hip that pulls half the torso with it.
How do expressions and visemes map to VRM?
Expressions map Blender shape keys or facial bone motion to named runtime actions. Blink, smile and mouth shapes usually begin as shape keys because a target avatar often needs additive control. If the add-on and destination application support bone-driven expressions, the source can be a bone instead.
Build the face in layers. Start with eye closure and the mouth shapes needed for speech. Add the emotional presets your application exposes. Then test how the controls combine. A smile plus a wide mouth shape plus a blink can produce a very different result from each expression in isolation.
- Blink: close upper and lower eyelid geometry without tearing the lash line.
- Visemes: test every mouth shape at the final head topology; AI-generated teeth or tongue geometry often needs cleanup.
- Emotions: preview brow, cheek and mouth combinations instead of accepting one exaggerated default.
- Look-at: place eye bones consistently and test the avatar while the head and eyes move together and independently.
What breaks a VRM export in Blender?
Most failures come from incomplete mapping or Blender data the exporter cannot translate cleanly. Check the rig, scale, modifiers, materials and expression source before changing export settings. The add-on can only serialize data that exists.
| Symptom | Likely cause | Correction |
|---|---|---|
| Humanoid mapping is incomplete | Required bones are missing or mapped to the wrong object | Fix the armature and humanoid assignment before export |
| Face expressions do not fire | Shape keys or bone controls are not bound to a VRM expression | Create and test the explicit expression mapping |
| Hair or skirt explodes | Spring chains, colliders, gravity or stiffness are untested | Reduce chain complexity and verify in the target runtime |
| Avatar imports at the wrong scale | Unapplied transforms or mixed object scales | Apply transforms and test a fresh import |
| Textures appear blank | External paths are missing or unsupported shader nodes remain | Pack images and simplify materials before export |
How do you export and test a VRM avatar?
Install a current release from the VRM Add-on for Blender, open its VRM panel, inspect the humanoid mapping and fill the required metadata. Configure expressions, look-at and SpringBone, then export. Exporting is not the test. Reimporting the finished file and opening it in a viewer is the test.
- Confirm the avatar loads at the expected scale and orientation.
- Pose the shoulders, elbows, hips and knees through their extremes.
- Trigger every blink, mouth and emotion preset.
- Move the look-at target while moving the head.
- Shake the WebGL or viewer camera to expose bad spring chains.
- Inspect textures, transparency and material slots from both sides.
Keep the working .blend file. A VRM is a delivery file, not a good source file. If the destination needs a corrected mouth shape, a new outfit or a different skeleton, the editable scene is where the change belongs.
VRM vs VRChat vs FBX: which should you use?
Choose the delivery format from the destination, not from the generator. VRM is strongest for portable avatar applications and expressions. FBX is a dependable interchange format for game engines and DCC tools. VRChat has its own Unity, SDK, performance and platform workflow.
| Target | Best delivery | Why |
|---|---|---|
| VRM-compatible viewer or VTuber app | VRM | Portable humanoid, expression and SpringBone metadata |
| VRChat | FBX into Unity | VRChat SDK, descriptor, performance rank and Unity Humanoid setup |
| Unity, Unreal or Godot game | FBX or glTF/GLB | Engine-specific rig, material and animation import pipelines |
| Editing and future variants | .blend source | Keeps modifiers, shape keys, materials and cleanup history editable |
If the destination is VRChat, use the VRChat avatar workflow instead. If the destination is a face-tracking stream application, the VTuber model guide covers the expression and runtime handoff in more depth.
What should you generate and what needs hand finishing?
Generate the broad character first pass and use Blender for the decisions that determine compatibility. The generated mesh is useful when it is editable. A polished preview is not evidence that the avatar can blink, speak, deform or export.
| Task | AI-assisted first pass | Required hand review |
|---|---|---|
| Body and outfit mesh | Yes | Silhouette, separations, normals and clean face geometry |
| Humanoid body rig | Yes | Joint placement, bone roles and deformation tests |
| Facial expressions | Partial | Shape keys, visemes, eyelid travel and combination tests |
| Hair and cloth motion | No complete solution | Spring chains, colliders, stiffness and runtime stability |
| VRM metadata and license | No | Author, thumbnail, permissions and third-party asset rights |
Key takeaways
- Blender AI is the character-generation and rigging layer. The VRM Add-on is the export layer.
- A valid VRM needs more than a mesh: humanoid mapping, expressions, materials, metadata and tested spring motion.
- Generate an original A-pose character, clean it, then rig the topology you intend to keep.
- Test deformation before face work and test the exported file after face work. Both gates catch different failures.
- Keep the .blend source. Export a VRM only after the avatar survives a real viewer and a reimport.