World Capture Architecture & Specs

A unified, fault-tolerant orchestration pipeline for photogrammetry and generative 3D, featuring deterministic state machines, Gaussian Splat compression, and cross-module event hooks.

System & Integration Metrics

Supported Modalities

7 (Text, Image, Video, LiDAR, Selfie, etc.)

Provider Adapters

17 Standardized Integrations

Output Formats

`.glb`, `.obj`, `.ksplat`, `.spz`, `.ply`

State Machine

15-State Deterministic Lifecycle

1. Provider Abstraction & The State Machine

Aeion OS does not execute GPU-heavy reconstruction natively. Instead, it acts as a hyper-intelligent orchestrator. - A Standardized Provider Contract: All 17 external providers (from Luma to Matterport) are wrapped in a strict, unified integration layer. This normalizes the chaotic variance in third-party APIs — webhooks vs. polling, OAuth vs. API keys — into a single, predictable schema, so every provider behaves the same way from Aeion's side. - Deterministic 15-State Lifecycle: Every scan job moves through a rigorous state machine (pending -> submitting -> submitted -> processing -> downloading -> uploading_to_assets -> completed). Because each transition is explicit and persisted, a job is never in an ambiguous "somewhere between submitted and done" state — an operator (or the recovery worker) can always read exactly which phase a job stalled in and what the safe next action is. If a provider's own photogrammetry timeout fires (which can take 4-6 hours for dense point clouds), Aeion gracefully catches the failure, queues it for recovery, and lets the operator replay the job on a different provider without re-uploading the multi-gigabyte source images — the raw input assets already live in your bucket, so a replay re-submits the cached inputs rather than starting the upload over. For example, when a location-scouting scan completes, Aeion automatically links the resulting Gaussian Splat to the matching record in the Production module — no manual download-and-attach step required.

2. Output Optimization & Security

The pipeline normalizes the chaotic outputs of 17 different providers into standard, enterprise-ready formats. - Format Normalization for Streaming: Provider outputs land in whatever format the provider natively produces — from heavier raw .ply point clouds to already-compressed splat formats like .ksplat and .spz. Aeion normalizes asset metadata and content-type handling across all supported formats on ingest, so the Asset Library can serve any of them for streaming rendering in the browser without a per-provider special case. Where a raw research-grade splat (.ply / .npz) can weigh 100-500 MB, the canonical .ksplat variant quantizes position, scale, rotation, color, and opacity to 16-bit and groups splats into spatial chunks — roughly halving on-disk size while letting the renderer fetch only the chunks inside the camera frustum. A 5-tier LOD chain (200K / 100K / 50K / 15K / 1K splats) is auto-generated so distant geometry renders with fewer splats, and progressive streaming shows a coarse pass within about a second on a phone instead of a blank canvas. - Biometric Provenance & Compliance: When using the selfie modality (e.g., via Ready Player Me or Avaturn), the engine automatically flags the generated asset's provenance as biometric input. This natively triggers EU AI Act disclosures and feeds World Capture's own legal-hold and retention workflows — soft-deleting the held scan job and its outputs when a hold is placed or a right-to-erasure claim is upheld.

Frequently Asked Questions

They stay put. Raw inputs are uploaded to your own bucket before submission, so a failure just moves the job into the recovery queue. Replaying against a different provider re-submits the cached inputs — you never re-upload a multi-gigabyte drone dataset.

The selector ranks all providers compatible with the input modality by health (30-second polling, `operational > unknown > degraded > outage`), then by real-time cost from the provider cost ledger, with a deterministic tie-break for consistent routing. Outage providers are stripped from the primary queue, and operators can override the pick per job.

No — some (Luma, Polycam, Pix4D, RealityCapture) support signed webhooks, while others (like Trellis) are polling-only. The provider integration layer hides that: it exposes a signed webhook endpoint where available and runs a background polling watchdog where not, so downstream modules always receive the same standard completion event.

`.glb` and `.obj` meshes, plus `.ksplat`, `.spz`, and `.ply` for Gaussian Splats. The `.ksplat` variant is the streaming-optimized default — quantized attributes, spatial chunking, and a 5-tier LOD chain for smooth browser and WebXR rendering.

When the `selfie` modality is used (e.g., Ready Player Me or Avaturn), the engine flags the asset's provenance as biometric input, which triggers EU AI Act disclosures and feeds the engine's own legal-hold and retention workflows — soft-deleting the held scan job and its outputs when a hold is placed or a right-to-erasure claim is upheld.

Composes with the rest of the OS

See the Capture Pipeline in Action