← Voxel World Model

End-to-End Pipeline

The complete intended path from real-world input to display — with the Adaptive Voxel World as the single source of truth, and AI only improving observations and caches.

Goal: one coherent pipeline, not a pile of tools

Reality becomes evidence in an Adaptive Voxel World. Simulation changes that world. The World Compiler emits derived artifacts (including Gaussians). Heavy AI may improve the Gaussian cache when compute is free. A rasterizer produces a frame packet every frame. Light local AI and optional full-frame AI enhance the image. Display never owns authority.

Authoritative Data Flow

Reality ↓ Ingestion & Normalization ↓ Registration / Evidence ↓ Adaptive Voxel World ←── SINGLE SOURCE OF TRUTH ↓ Simulation ↓ World Compiler ↓ Gaussian Representation ↓ Background Gaussian Optimization (heavy AI, cached, not every frame) ↓ Enhanced Gaussian Cache ↓ Gaussian Rasterizer → GaussianImagePacket ↓ Real-Time Local Enhancement (light AI, every frame, optional) ↓ Optional Full-Frame Post Process ↓ Display (Desktop · Browser · VR · AR · Mobile · Cloud)

Background improvement can update the Gaussian cache without rebuilding the voxel world, re-running simulation, re-importing data, or interrupting rendering.

Real World Input

┌──────────────────────┐ │ REAL WORLD INPUT │ └──────────────────────┘ │ ┌───────────────────────┼────────────────────────┐ │ │ │ Images Video Streams Live Sensors │ RGB + IMU + GPS LiDAR / Radar └──────────┬────────────┴──────────┬─────────────┘ │ │ 3D Scans / Point Clouds Existing Assets (LiDAR, Meshes, CAD…) (glTF, FBX, USD…) │ │ └───────────┬───────────┘ ▼ Ingestion & Normalization
Shipped: V20 Reality Capture · V10–V14 foundation

Ingestion & Normalization

Decode Frame Extraction Camera Calibration Pose Estimation SLAM / SfM Depth Estimation Sensor Fusion Coordinate Normalization Timestamp Alignment Quality Validation

Raw captures become time-aligned, calibrated ObservationPackets ready for registration — never silent merges of incompatible frames. See V20.

Shipped: V11 evidence · V14 pose graph

Registration / Evidence

Device Registration Observation Registration Evidence Graph Confidence Tracking Source Attribution Temporal History Conflict Resolution

Observations accumulate into evidence. Agreement raises confidence; contradiction lowers it without overwriting history. Pose graphs align multi-capture sessions into one world frame.

Authority

Adaptive Voxel World

Authoritative world state — the single source of truth.

Spatial Hierarchy Adaptive Voxels Semantic Labels Materials Physics Object Identity Evidence Confidence Relationships Behavior Attachments Simulation State

Everything below this line is derived or observational. Gaussians, meshes, neural frames, and stylizations never replace this layer.

Shipped: V17

World Simulation

Behaviors Triggers Scheduler Timeline Runtime Interactions Weather Robots Vehicles Doors Lights Trees Water Physics Time

Simulation advances world state over time through the Runtime. Compiler and visualization respond automatically (e.g. door open → dirty chunk → gaussian recompile) with no special orchestration in the sim engine.

Shipped: V15 Part 1

World Compiler

Visible Region Selection LOD Selection Chunk Compilation Visibility Materials Lighting Metadata Generation

Outputs (artifacts): Gaussian Scene · Mesh Scene · Voxel Scene · Point Cloud · future render formats. Compilers never write the authoritative world — they publish committed artifacts only.

Foundation: V12 · V16 Goal: V13

Gaussian Representation

Positions Covariance Rotation Opacity Color SH Coefficients Material Features Object IDs Semantic IDs

A derived, chunked representation of the world for real-time observation — not authority.

Shipped: V19 classical splat opt · AI stubs guarded

Background Gaussian Optimization

Runs whenever compute is available — not every frame. Improves the Gaussian representation itself; results are cached as versioned revisions. See V19.

Dynamic Splat Stitching Overlap Repair Coverage Completion Opacity Optimization Covariance Refinement Material Refinement Color Consistency Radiance Optimization Temporal Fusion Surface Reconstruction Hole Filling

Optional Heavy AI

FLUX · HiDream · future foundation models — improve the Gaussian representation offline or on spare capacity. They do not run in the real-time display path. Registered as guarded plugins until GPU weights exist.

Enhanced Gaussian Cache (V19) — optimized splat revisions, nothing overwritten, history preserved. Browser / V18 path transparently prefer best_arrays(). Only recompute when necessary.

Foundation: V12 splat · V16 viz

Gaussian Rasterizer

Visible Splats Rasterization Visibility Depth Transparency Lighting

Produces: GaussianImagePacket

RGB Depth Normals Motion Vectors Material Map Semantic Map Object IDs Confidence Metadata Embeddings
Shipped: V18 classical-CV · AI stubs guarded

Real-Time Local Gaussian Enhancement

Runs when enabled. Packet-only — never touches the voxel world. Profiles scale low (passthrough) → medium → high → cloud. Live classical-CV enhancers; AI stages (RealESRGAN / photorealistic / reflection) are registered but guarded until GPU weights exist.

Sharpen Details Better Glass Cleaner Edges Improve Reflections Reduce Splat Artifacts Better Blending Anti-Aliasing Texture Recovery Surface Detail Temporal Stability

Enhanced Frame — a photorealism-ward Gaussian image, still an observation. See V18 report and architecture status.

Optional Full-Frame AI Post Process

Operates on the final image only — entirely optional (still goal):

HDR Tone Mapping Atmospheric Effects Bloom Lens Effects Depth of Field Film Grain Stylization Anime Watercolor Night Vision Thermal Accessibility Filters Upscaling DLSS/FSR-style Reconstruction

See also Frame Enhancement for the local vs global separation and independent local_ai / global_ai toggles.

Display

Desktop Browser VR AR Mobile Cloud Streaming

Continuous Background Improvement

As more compute becomes available… Camera / view prediction · Idle GPU · Cloud workers · Second GPU · Nightly jobs │ ▼ Background Gaussian Optimization │ ▼ Improves Splats → Updates Enhanced Gaussian Cache │ ▼ Future frames become better WITHOUT • Rebuilding the Voxel World • Re-running Simulation • Re-importing Data • Interrupting Rendering

Invariant: heavier models improve the cached Gaussian representation. They never become the simulator, never invent occupancy into authority, and never block the real-time path.

How this maps to shipped milestones

Pipeline stageStatusMilestone
Ingest / reality captureShippedV20 · report
Evidence / confidenceShippedV11 Runtime
Registration / pose graphShippedV14 Mapping
Adaptive Voxel WorldAuthorityCore (capture era → present)
Reactive Runtime (event bus)ShippedV21 · report
SimulationShippedV17
World CompilerShipped (Part 1)V15
Gaussian artifact + rasterFoundationV12 · V16
Background Gaussian opt + cacheShipped (classical)V19 · report
Local Gaussian enhancersShipped (classical-CV)V18 · report
Neural enhancement backendsShipped (RealESRGAN CUDA)V22 · report
Semantic world engineShipped (evidence + hypotheses)V23 · report
Persistent world identityShipped (timeline + replay)V24 · report
World reasoning engineShipped (facts + causal graph)V25 · report
Integrated world kernelShipped (tx · events · snapshot)V26 · report
World services / queryShipped (cross-layer explain)V27 · report
Distributed replication / syncShipped (vector clocks · conflicts)V28 · report
Distributed trust / operation replicationShipped (ops · HMAC · recompute)V29 · report
Kernel finalization / v1.0Shipped (regression · Ed25519 · governance)V30 · report
Optional full-frame / display AIGoalLater · enhancement layers

Architecture status → Multimodal World I/O → Roadmap → Updates → Frame enhancement → Paper →