The engineering problem
You have 200 photographs of a heritage site, a construction progress scene, or an industrial facility. You need a geometrically accurate 3D model — not a rough approximation, but something precise enough for measurement, visualization, and downstream analysis.
The traditional path: manually configure COLMAP for sparse reconstruction, tune dense matching parameters, run Meshroom with custom pipeline graphs, export to a mesh format, then spend days iterating on NeRF or Gaussian Splatting parameters to get photorealistic rendering. Each tool has its own environment requirements, GPU dependencies, and failure modes.
The result? Days of configuration before any engineering value is produced. No lineage connecting the input photographs to the final model. No guarantee that the same process will produce the same output next month.
Why manual approaches are insufficient
Three problems compound in 3D reconstruction:
Environment complexity. Nerfstudio requires specific CUDA versions, PyTorch builds, and GPU memory configurations. Meshroom needs its own GPU pipeline. COLMAP has separate compilation requirements. Maintaining these environments across a team means every workstation is a snowflake.
No verification. When you deliver a 3D model to a client, can you prove which photographs produced it? Which algorithm settings were used? Whether the model was post-processed? Manual pipelines have no inherent lineage.
Scale brittleness. Processing one scene works. Processing 50 scenes from a drone survey campaign means managing queues, handling failures, and tracking which outputs correspond to which inputs — all manually.
How Studio executes 3D reconstruction
Intent: "Reconstruct a photorealistic 3D model of the Sponza Palace courtyard from this photo set. Output: web-viewable glTF, point cloud (PLY), and quality metrics."
Studio decomposes this into an execution plan:
1. Image ingestion → validate EXIF, check overlap sufficiency
2. Sparse reconstruction → COLMAP feature matching + SfM
3. Dense reconstruction → Meshroom multi-view stereo
4. Neural rendering → Nerfstudio (nerfacto) or 3D Gaussian Splatting
5. Export → PLY point cloud + glTF mesh + Three.js viewer
6. Verify → hash all outputs, record parameters, generate quality report
The Execution Fabric orchestrates:
- Provisioning an A100-class GPU for high-resolution Gaussian Splatting workloads
- Loading the 3D Reconstruction Capability Pack (Nerfstudio, 3DGS toolkit, Meshroom 2023, COLMAP 3.9)
- Scheduling each stage with dependency awareness (step 4 requires step 2's camera poses)
- Checkpointing long-running NeRF training (resume after interruption without re-processing)
The 3D & Immersive Workbench provides:
- Nerfstudio with multi-method support (nerfacto, instant-ngp, tensorf, gaussian-splatting)
- Meshroom 2023 + COLMAP 3.9 for photogrammetry
- Blender 4.0 for mesh post-processing
- Three.js and CesiumJS for web-native delivery
- CloudCompare + Open3D + PDAL for point cloud processing
- Format export: PLY, OBJ, FBX, glTF, 3D Tiles
Verified output: Every artifact — the point cloud, the trained NeRF model, the exported glTF — receives a SHA-256 hash and execution lineage. Six months later, you can prove exactly which photos, which algorithm, and which parameters produced that model.
Live evidence
The Sponza Palace Viewer and Nefertiti 3D Viewer demonstrate browser-native delivery of 3D assets produced for our public showcase. They provide evidence of the downstream visualization layer; the reconstruction workflow described above represents the Studio execution architecture supported by the 3D & Immersive Workbench.
The Demo Catalogue includes additional 3D outputs across heritage, industrial, and urban domains.
Deployment flexibility
The same reconstruction pipeline is designed to run across all seven deployment modes. A heritage documentation team working in a remote site with limited connectivity can run the identical Capability Pack on a Portable Appliance — encrypted, air-gapped, GPU-equipped — and produce the same verified outputs as a cloud deployment.
This is Part 2 of the How Nebula Cloud Executes Work series.
Previous: Part 1 — From Prompt to Verified Artifact Next: Part 3 — Executing Engineering Design · Part 4 — Executing Geospatial Intelligence
Related: 3D & Immersive Workbench · Capability Packs · Execution Fabric · Demo Catalogue