curiositech/vr-avatar-engineer

VR Avatar Excellence Engineer

Expert in building high-quality avatar systems for VR/metaverse. Deep knowledge of real-time rendering, facial tracking, and cross-platform development for Vision Pro, Quest, and PC VR.

When to Use This Skill

✅ Use for:

• VR avatar systems (Vision Pro, Quest, PCVR)
• Facial tracking integration (ARKit 52 blend shapes, Meta face tracking)
• Avatar generation from photos/scans
• Real-time networking for multiplayer avatars
• Subsurface scattering and skin rendering
• Performance optimization for VR frame rates
• Cross-platform avatar synchronization

❌ Do NOT use for:

• 2D profile pictures → use image generation tools
• Non-VR game characters → use game engine character tools
• Static 3D modeling → use Blender/Maya skills
• Motion capture hardware setup → specialized mocap domain
• Deepfakes/non-consensual likenesses → ethical boundary

MCP Integrations

| MCP | Purpose | |-----|---------| | Stability AI | Generate avatar concept art, texture references | | Firecrawl | Research Meta/Apple SDKs, avatar papers | | WebFetch | Fetch ARKit, Meta SDK documentation |

Expert vs Novice Shibboleths

| Topic | Novice | Expert | |-------|--------|--------| | Blend shapes | "Just use morph targets" | Knows ARKit has 52 specific shapes; Meta has different set; mapping required | | Skin rendering | "Just use PBR" | SSS is essential; different models for different skin tones | | Eye tracking | "Point eyes at target" | Saccades, microsaccades, blink patterns make presence | | Networking | "Send all data every frame" | Delta compression, interpolation, dead reckoning | | Frame rate | "60fps is fine" | Quest: 72/90/120hz modes; Vision Pro: 90hz minimum; dropped frames = nausea | | LOD | "Lower poly for distance" | Foveated rendering integration, dynamic LOD based on gaze |

Common Anti-Patterns

Anti-Pattern: Uncanny Valley Through Over-Realism

What it looks like: Photorealistic face with robotic expressions Why it's wrong: Partial realism triggers uncanny valley; stylization often works better What to do instead: Match rendering fidelity to tracking fidelity; stylized avatars hide tracking limitations Example: Vision Pro Personas work because they're slightly stylized, not photorealistic

Anti-Pattern: Ignoring Platform Differences

What it looks like: Same avatar pipeline for Quest and Vision Pro Why it's wrong:

• Quest: Mobile GPU, 72fps minimum, limited polys
• Vision Pro: Desktop-class GPU, 90fps, Personas API is different What to do instead: Platform-specific LOD targets, shader variants, API abstractions

Anti-Pattern: Synchronous Networking

What it looks like: Blocking on avatar state updates Why it's wrong: Network latency causes frame drops = VR sickness What to do instead: Asynchronous updates with interpolation and prediction

Anti-Pattern: Single Skin Shader

What it looks like: One SSS configuration for all skin tones Why it's wrong: Melanin affects scattering; darker skin needs different SSS parameters What to do instead: Parameterized skin shader with melanin-aware scattering

Evolution Timeline

Pre-2020: Early VR Avatars

• Stylized/cartoon avatars dominant (VRChat, Rec Room)
• Limited tracking (3-point: HMD + controllers)
• No facial expressions in most apps

2020-2022: Quest 2 Era

• Hand tracking mainstream
• Basic lip sync from audio
• Meta Avatars SDK emerges
• 72fps becomes standard

2023-2024: Spatial Computing

• Vision Pro Personas (Feb 2024): ML-generated photorealistic avatars
• Quest 3 with improved face/eye tracking (add-on)
• Codec Avatars research (Meta) shows photorealistic path
• Cross-platform interop becomes critical

2025+: Current Best Practices

• Hybrid approach: Personas for presence, stylized for games
• Neural rendering for hair/fabric
• Real-time relighting from environment
• Privacy-preserving avatar generation (on-device)

Core Implementation Patterns

Facial Tracking (ARKit → Avatar)

// ARKit face tracking to blend shape weights
func mapARKitToAvatar(faceAnchor: ARFaceAnchor) -> [String: Float] {
    let arkit = faceAnchor.blendShapes

    // Direct mappings (ARKit names → avatar shapes)
    var weights: [String: Float] = [:]
    weights["jawOpen"] = arkit[.jawOpen]?.floatValue ?? 0
    weights["mouthSmileLeft"] = arkit[.mouthSmileLeft]?.floatValue ?? 0
    weights["mouthSmileRight"] = arkit[.mouthSmileRight]?.floatValue ?? 0
    weights["eyeBlinkLeft"] = arkit[.eyeBlinkLeft]?.floatValue ?? 0
    weights["eyeBlinkRight"] = arkit[.eyeBlinkRight]?.floatValue ?? 0

    // Derived expressions (combinations)
    let smile = ((weights["mouthSmileLeft"] ?? 0) + (weights["mouthSmileRight"] ?? 0)) / 2
    weights["expression_happy"] = smile

    return weights
}

Network-Optimized Avatar State

// Photon PUN2 - efficient avatar sync
public struct AvatarState : INetworkStruct {
    // Pack position: 3 floats → 6 bytes (half precision)
    public Half3 Position;

    // Pack rotation: quaternion → 4 bytes (compressed)
    public CompressedQuaternion Rotation;

    // Blend shapes: 52 weights → 52 bytes (uint8 each, 0-255 → 0-1)
    public fixed byte BlendShapes[52];

    // Eye gaze: 2 directions → 4 bytes
    public Half2 LeftEyeGaze;
    public Half2 RightEyeGaze;

    // Total: ~70 bytes per update (vs 400+ uncompressed)
}

Subsurface Scattering for Skin

// Simplified SSS for real-time (pre-integrated)
float3 SubsurfaceScattering(float3 normal, float3 light, float3 view,
                            float curvature, float3 skinColor, float melanin) {
    float NdotL = dot(normal, light);

    // Wrap lighting for soft terminator
    float wrap = 0.5;
    float diffuse = saturate((NdotL + wrap) / (1 + wrap));

    // Pre-integrated scattering lookup (curvature-based)
    float2 sssUV = float2(NdotL * 0.5 + 0.5, curvature);
    float3 sss = tex2D(_SSSLookup, sssUV).rgb;

    // Melanin affects scattering color
    float3 scatterColor = lerp(float3(1, 0.4, 0.25), float3(0.8, 0.5, 0.4), melanin);

    return skinColor * diffuse + sss * scatterColor * (1 - diffuse);
}

Performance Targets

| Platform | Frame Rate | Avatar Poly Budget | Texture Budget | |----------|------------|-------------------|----------------| | Quest 2 | 72-90 fps | 10-15k tris | 512×512 | | Quest 3 | 90-120 fps | 20-30k tris | 1024×1024 | | Vision Pro | 90 fps | 50-100k tris | 2048×2048 | | PCVR | 90-144 fps | 100k+ tris | 4096×4096 |

Integrates With

• metal-shader-expert - Custom skin/hair shaders
• physics-rendering-expert - Hair/cloth simulation
• sound-engineer - Spatial audio for voice
• clip-aware-embeddings - Avatar search/matching

---

Remember: VR avatars are how people represent themselves in shared virtual spaces. Focus on presence (feeling "there"), performance (smooth frame rates), and inclusivity (all bodies, all identities). The best avatar is one that disappears—users forget they're looking at pixels and just see the person.