Metaverse avatar development: formats, rigging, tracking and cost

A metaverse avatar is a rigged 3D character that stands for a person in a shared virtual space. The person drives it in real time with a headset, a camera, a keyboard or a controller. This guide covers the main decisions in avatar development: style, file format, pipeline, tracking, performance, customization, NFTs, privacy and cost.

Users judge a virtual space by its avatars. They see at once when the lips move after the voice, when a hand passes through the body, or when the character does not look the way they want to look. Avatars also cost the most rendering time in a social scene. One detailed avatar renders easily, but fifty avatars on a standalone headset can drop the frame rate.

Choosing a metaverse avatar style

The style you pick decides how much tracking error users notice and how much each avatar costs to render.

Apple's Personas on Vision Pro show where photoreal avatars are going. In visionOS 26, released in September 2025, Apple rebuilt Personas with volumetric rendering and machine learning, according to its visionOS 26 announcement. Apple creates these avatars on the device with its own capture hardware, and a typical third-party app cannot copy this capture process. Most apps still use a stylized look.

File formats and interoperability

Open formats let an avatar work in several apps and survive a change of game engine.

• glTF 2.0 and GLB. The Khronos Group maintains this real-time 3D format, and every major engine and web 3D library can load it.
• VRM. VRM is built on glTF and adds what a humanoid avatar needs. The VRM 1.0 specification, released in September 2022, defines a standard humanoid bone map, expressions, look-at settings, spring-bone physics for hair and clothing, the MToon toon material and node constraints. VRM files also carry license metadata that says how others may use the avatar. The VRM Consortium maintains the format, and VTuber and social VR communities use it widely.
• OpenUSD. Film and design pipelines and Apple platforms use OpenUSD, but few apps use it to exchange user avatars.
• FBX. Authoring tools still export FBX. Autodesk owns the format, and it is a poor choice for delivery at runtime.

Hosted avatar services supply cross-app avatars through an SDK and can save months of work, but they can also disappear. Netflix bought Ready Player Me in December 2025, according to Road to VR, and Ready Player Me then closed its public services, including the PlayerZero avatar tool, on January 31, 2026. Apps that used Ready Player Me had to replace their avatar systems. If you depend on a hosted service, keep the right to export each avatar as glTF or VRM, and store a copy yourself.

The avatar development pipeline

1. Base mesh. Define body proportions, a style guide and a polygon budget. A modular body with separate head, torso, legs and hands makes customization and levels of detail easier.
2. Topology. Place edge loops around the joints, mouth and eyes so the mesh can bend cleanly during animation.
3. Rigging. Build a humanoid skeleton that your engine's retargeting system accepts, such as Unity Humanoid, the Unreal skeleton or the VRM bone map. Tune the skin weights at the shoulders, hips and fingers.
4. Facial setup. Add blendshapes, also called morph targets, for expressions and for visemes. Visemes are the mouth shapes used for lip sync. Apple's ARKit defines 52 face blendshapes, and many face-tracking tools use this set as a reference, so support for it makes tracking integration easier.
5. Materials and textures. Use physically based materials for realistic styles or toon shaders for stylized ones. Pack textures into atlases to reduce draw calls, and compress them with a format such as KTX2.
6. Levels of detail (LODs). Make two to four LODs. The lowest LOD should be cheap enough for crowds at a distance. Very distant avatars can use impostors, which are pre-rendered sprites.
7. Customization system. Define swappable meshes, color parameters and attachment points as data, so your team can add new items without an app update.
8. Validation. Check bone names, triangle count, material count, texture size and scale automatically before any avatar or item reaches users.

Driving the avatar in real time

Body tracking and inverse kinematics

A VR headset usually tracks the head and the two hands or controllers, so the app must estimate the rest of the body. Inverse kinematics (IK) calculates the elbows, shoulders and spine from those three points, and procedural animation or a machine-learning model estimates the legs.

Meta's Movement SDK offers two features for Quest. Inside-out body tracking uses the Quest 3 cameras to track the upper body. Generative Legs estimates leg movement on all Quest headsets, as Meta explains in its developer blog. Generative Legs does not follow every real movement, such as a raised knee. On other platforms, developers use the engine's IK or a third-party solver. Crouching, reaching behind the body and sitting need extra tuning on any platform.

Face, eyes and lip sync

Some headsets track the face and eyes, and that data can drive the blendshapes directly, which makes the avatar feel much more present. Without face tracking, audio-driven lip sync maps speech to visemes. Procedural blinking and small eye movements keep the avatar from staring blankly. Let users turn face and eye tracking off, and treat that data as sensitive.

Desktop and mobile users

Users on flat screens walk, emote and gesture with a keyboard or a touch screen. Avatars should look natural for both headset and flat-screen users. Many platforms offer webcam face tracking as an optional upgrade.

Performance budgets

Each avatar is skinned, animated and unique, so avatars usually take the largest share of rendering cost in a social scene. Many teams start from these budgets for standalone headsets:

• A triangle budget in the low tens of thousands per avatar at the highest LOD, and much less at a distance.
• One or two materials per avatar, with textures packed into atlases.
• A limit on bones per avatar and on spring-bone chains for hair and clothes.
• Lower animation and IK update rates for distant avatars.

Confirm these numbers by profiling on your weakest target device with the largest number of avatars you expect on screen. If users can upload their own avatars, check each upload against the budget and replace one that exceeds it with a simplified version.

Customization and user-created items

Customization keeps users active, and many social platforms earn most of their revenue from it. A customization system needs answers to three questions:

• Who creates items? Your own artists, approved creators or any user. User-generated items need validation, moderation of offensive designs and a takedown process for copyright claims.
• How do items fit different bodies? Shared body templates or automatic fitting stop a jacket from clipping through an unusual body shape.
• How are items sold? Through in-app purchases, a marketplace or limited drops.

NFT avatars and wearables

In 2021 and 2022, many profile-picture NFT collections promised 3D avatars that would work "across the metaverse." Few collections delivered avatars that worked in several apps. Each app has its own style, rig and performance limits, and some collections were abandoned. If you want avatars or wearables as NFTs, plan for these points:

• Use the NFT as a record of ownership and license. The token metadata points to a glTF or VRM file on durable storage, such as IPFS with pinning or Arweave.
• State the license clearly. Holders need to know whether they may use the avatar commercially, in any app, or only in yours.
• Other apps must support your format and style before the avatar works there. The token alone does not make the avatar portable.
• Use ERC-1155 for wearable editions and ERC-721 for unique characters. The guides to BAYC-style collections and CryptoPunks-style projects cover the collection side.

Most products use an off-chain avatar system and add NFT wearables as an option.

Privacy, safety and inclusion

• Face, eye and body tracking data can reveal health conditions and emotions. Collect as little as possible, process it on the device where you can, and tell users what you collect.
• Scanned or AI-generated likenesses need the person's consent. Block avatars that impersonate real people.
• Offer many body types, skin tones, hair textures, assistive devices and cultural clothing, so users can represent themselves.
• Let users control their personal space and who can see their avatar up close.

Avatar development cost drivers

A rough example: a stylized avatar system with a few body types, a facial rig and a few dozen launch items might need one technical artist, two 3D artists and one engineer for three to five months. A photoreal pipeline or a full marketplace for user-made items is a much larger project. The guide to metaverse app development covers the application around the avatars. The social metaverse guide covers the spaces where avatars matter most.

Frequently asked questions