Ingest Reality.
Bridge the gap between atoms and bits. The World Capture Engine is a unified ingestion pipeline that aggregates 17 world-class photogrammetry and generative AI providers. Instantly transform drone photography, iPhone LiDAR, and mobile selfies into hyper-optimized Gaussian Splats and 3D Meshes, piped natively into your enterprise workflows.
Cross-Module Spatial Intelligence
Construction BIM Sync
Upload a dataset of 500 drone photos of a job site. The engine routes the job to Pix4D or RealityCapture, generating a massive Gaussian Splat, then automatically mounts the true-to-life 3D scan behind your CAD models in the Construction module for real-time clash detection.
Production Location Scouting
Film crews scan potential shooting locations with Polycam LiDAR. The resulting `.ksplat` file is instantly piped into the Production module. Directors can then overlay real-world sun-path tracking directly onto the 3D terrain to pre-visualize lighting for any date and time.
Biometric Avatar Generation
Employees take a selfie. The engine routes it to Ready Player Me or Avaturn, generating a fully-rigged 3D head mesh. This identity reference feeds directly into the Character Engine, applying procedural clothing and facial-hair to represent the employee accurately in the Spatial MMO or Smart Spaces dashboard.
Spatial Desktop Backdrops
Ditch static 2D wallpapers. Scan your physical office or studio, and the engine transcodes it into an optimized `.ksplat` format and mounts it directly as your Spatial Desktop backdrop, bringing immersive spatial computing right into your browser.
17 Providers, One Unified API
The Modality Router
You provide the input; Aeion OS handles the rest. The engine supports 7 modalities: `text-prompt`, `single-image`, `image-set`, `multi-view-turntable`, `video`, `lidar`, and `selfie`. It automatically routes the job to the optimal provider—from Luma and Trellis for AI generation, to Matterport and DroneDeploy for structural photogrammetry.
Automatic Fallback
Provider outages shouldn't stop your workflow. The router actively monitors provider health and cost, and if Polycam goes down, the engine silently falls back to Scaniverse or Rodin, ensuring your pipeline never stalls.
The 17-Provider Catalog
Matterport
Real-estate + interior walkthroughs (image-set, multi-view) — the industry standard for property tours.
Pix4D
Survey-grade drone photogrammetry (image-set, drone) — engineering-quality mesh + orthomosaic output.
RealityCapture
Large-scale aerial photogrammetry (image-set) — construction + cinema-grade structural scans.
DroneDeploy
Drone fleet workflows (image-set, drone) — tied to drone-flight planning + flight logs.
Polycam
Mobile LiDAR + photogrammetry (lidar, image-set) — iPhone Pro / iPad Pro LiDAR-first.
Scaniverse
Mobile photogrammetry (image-set, video) — free tier, consumer-grade fallback.
Luma
NeRF + Gaussian Splat generative capture (video) — high-quality view-synthesis from video.
Trellis
Single-image to 3D (single-image) — AI-generated mesh from one photo.
Tripo
Text + image to 3D (text-prompt, single-image) — generative AI mesh.
Meshy
Text + image to 3D (text-prompt, single-image) — generative AI with rigged output.
Rodin
Text + image to 3D (text-prompt, single-image) — high-fidelity AI generation.
Stability
Text-to-3D via Stable Diffusion (text-prompt).
CSM
Image-to-3D (single-image, video) — from Common Sense Machines.
Hunyuan
Text + image to 3D (text-prompt, single-image) — Tencent's Hunyuan-3D.
Spline
Designer-friendly 3D (text-prompt, single-image) — Spline.design integration.
Ready Player Me
Selfie → rigged avatar (selfie) — full-body humanoid avatar.
Avaturn
Selfie → rigged avatar (selfie) — photo-realistic head + body.
The Capture Pipeline — End to End
When a user submits a scan job (drone photos, LiDAR scan, video, selfie, text prompt):
- Asset upload. Multi-part chunked upload to your own S3 bucket, with resumability, virus scanning, and format validation. A 500-photo drone dataset (~4 GB) uploads in ~6 min on a 100 Mbps link, and chunked retry survives connection drops.
- Modality + provider selection. The router examines input type, quality preset, cost ceiling, and provider health, then picks the best provider for the job. Operators can override.
- Provider submission. The integration for the chosen provider (e.g. Pix4D, Matterport) formats the request and submits it, handling that provider's quirks — auth flow, format requirements, billing API, webhook registration — automatically.
- Progress tracking. Each job moves through a per-job pipeline (submitted → processing → reconstructing → optimizing → completed, or failed), with progress polling for providers without webhooks and webhook receivers for those that have them.
- Webhook reception. The provider's completion event triggers a status update, asset download, and automatic replay support for any missed deliveries.
- Output transcoding. Raw provider output (PLY, OBJ, FBX, ZIP) is transcoded to Aeion's canonical formats:
.ksplatfor Gaussian Splats (web-optimized),.glbfor meshes, with embedded metadata. - LOD generation. A 5-tier LOD chain is auto-generated: high-detail (200K+ splats), high (100K), medium (50K), low (15K), placeholder (1K). LOD switching is distance-based on the client renderer.
- Provenance log. An append-only audit row tracks who submitted the job, which provider processed it, the billed amount, an output checksum, and timestamps for each pipeline phase — auditor-friendly and Aegis-backed.
- Event notification. Completion is broadcast automatically, so the Construction, Production, Spatial Desktop, and Sentinel modules can pick up and use the new asset the moment it's ready.
- Retention enforcement. A per-tenant retention policy purges raw inputs after a configurable window (default 90 days); the optimized output is retained indefinitely.
The .ksplat Format — Web-Optimized Gaussian Splats
Raw Gaussian Splat files from research-grade reconstruction (.ply / .npz) are typically 100-500 MB. That doesn't ship to a phone. The Aeion .ksplat format is a streaming-optimized variant:
- Quantized attributes. Position + scale + rotation + color + opacity packed to 16-bit (vs 32-bit floats), ~50% size reduction with imperceptible visual loss.
- Spatial chunking. Splats grouped into spatial chunks; the renderer fetches only chunks in the camera frustum + within draw distance.
- 5-tier LOD chain. Same scene at 5 detail levels (200K / 100K / 50K / 15K / 1K splats). LOD switching is per-chunk based on screen-space size — distant chunks render with fewer splats.
- Progressive streaming. Coarse LOD loads first; finer detail streams in as bandwidth permits. User sees something within ~1 sec on a phone instead of a blank canvas for 30 sec.
- Cross-platform decoder. Decoding runs in a Web Worker on the client, so it never blocks the UI thread. Native iOS / Android decoders are coming for Mobile Builder integration.
- Renderer. Three.js + WebGPU. 60 fps on mid-range mobile for 100K-splat scenes; 30 fps on entry phones via LOD downgrades.
Result: a Matterport-style space tour that's 80 MB and renders smoothly on the user's phone — instead of "download our 500 MB native app, then wait 30 seconds."
Cross-Module Spatial Workflows
Construction
Drone-scan a job site weekly, and the engine auto-overlays the scan against your BIM model (see the Universal AST Engine architecture). Differences get flagged automatically — missing rebar, deviated framing, schedule slippage — and RFIs auto-attach the scan as evidence.
Production / VFX
Location scouts capture sites via phone LiDAR, and pre-vis tools overlay sun-path, camera-placement, and crew-blocking directly on the real-world scan. Directors see the final shot composition before the crew ever arrives, and the output ships straight to the dailies pipeline.
Sentinel
Drone-scan a facility quarterly and let anomaly detection flag new structures, removed assets, or unauthorized changes, tying straight into the physical-security incident workflow.
Real Estate / Hospitality
Self-serve 360° walkthroughs for property listings, captured via Matterport or Polycam. The same asset powers the web-listing-page hero, the in-person agent's tablet, and the WebXR remote tour.
LMS / Training
Capture real-world physical environments for training simulations — medical students walk through scanned hospital wards, firefighters practice in scanned buildings, and oil & gas workers learn refinery layouts before ever stepping on site.
Spatial Desktop Backdrop
Scan your physical office and mount it as the immersive backdrop for your Aeion Spatial Desktop, giving spatial multitasking real-world environmental coherence.
E-commerce
Single-photo product scans via Trellis, Meshy, or Tripo become a 3D product card on the storefront, so customers rotate and zoom a real 3D model instead of staring at a flat photo.
Avatar Generation
A selfie routed through Ready Player Me or Avaturn becomes a fully-rigged avatar, fed into the Character Engine and worn in Spatial scenes.
Provider Economics — Why the Router Saves Money
| Provider | Per-scan cost (typical) | Use when |
|---|---|---|
| Scaniverse | $0 (free tier) | Casual interior captures; consumer-grade output |
| Polycam | $0.50-2.00 | Mobile LiDAR scans; pro-tier mobile workflows |
| Luma | $0-5 | Video-to-NeRF; AI generative content |
| Trellis / Meshy / Tripo / Rodin | $0.10-1.00 | Single-image-to-3D; AI-generative e-commerce |
| Ready Player Me / Avaturn | $0-2 | Selfie-to-avatar |
| DroneDeploy | $5-50 | Drone fleet workflows tied to flight-planning |
| Pix4D | $10-100 | Survey-grade engineering photogrammetry |
| RealityCapture | $20-200 | Cinema / large-scale aerial structural |
| Matterport | $20-150 | Real-estate interior tours (per-space pricing) |
Frequently Asked Questions
Both work. **Free tier:** you pick the provider explicitly per submission. **Pro+:** the Auto Provider Selector picks based on input modality, quality preset, cost budget, and provider health. Override anytime per job.
No. Provider fees pass through transparently — you see the actual provider invoice. The Aeion orchestration engine is included. The value-add is the integration + routing + .ksplat optimization + cross-module event delivery, not a markup on provider compute.
Splats render *photographic* reality with millions of fuzzy 3D points (gaussians) — great for capture fidelity, but no clean topology, no UV maps, no easy editing. Meshes are clean topology + UV-mappable + animatable, but reconstruction loses fine detail. Aeion supports both: capture as splats for accurate visualization, convert to mesh for editing/animation when needed (provider-dependent).
Highly variable. A 30-photo room scan via Polycam → 5-15 MB .ksplat. A 500-photo building exterior via Pix4D → 80-200 MB .ksplat. A cinema-grade RealityCapture scan → 500 MB-2 GB. The 5-tier LOD chain + spatial chunking means clients only download what they need — typical mobile load: 5-15 MB to start displaying, more streamed on demand.
Yes — the splat renderer is built on Three.js + WebGPU and is WebXR-compatible. Captured .ksplat scenes work natively in Quest, Vision Pro, and Aeion's Spatial Desktop. For native AR (iOS ARKit / Android ARCore), the Mobile Builder includes splat-rendering primitives.
Splats: limited editing (mostly cropping + lighting). Meshes: full editing in any standard DCC (Blender, Maya, Cinema 4D) — export as glb/fbx/obj. The Aeion Spatial Editor includes basic splat-cropping + lighting + scene composition. For complex editing, export to your DCC + bring the result back.
The engine includes face-blur post-processing on capture (provider-dependent for raw inputs; always available on the .ksplat output via the Spatial Editor's privacy pass). For Sentinel surveillance workflows, an opt-out flag suppresses face blur with appropriate audit. Configure per scene + per tenant.
Enterprise tier. The "air-gapped pipeline" route uses on-premise open-source reconstruction (COLMAP + Nerfstudio + 3D Gaussian Splatting) without external API calls. Slower than cloud providers and quality varies, but for sensitive industries (defense, healthcare with PHI scans) it's the right model.