Skip to main content

Lightfielder Live Grade Video Flowchart

Prepared by: Andrew Hazelden andrew@andrewhazelden.com
Date Created: 2026-04-02

System Overview​

This document describes the video flow for a 50-camera array system that uses SDI output at 6K video resolution from RED Digital Cinema camera bodies, ultimately producing 3DGS (Gaussian Splat) trained 3D model assets.

Flowchart​

Figure 1: Complete video flow from camera array to 3D model output

Detailed Video Flow Summary​

A. Camera Array & Source​

  • 50x RED Digital Cinema cameras, each set to output high-resolution video.
  • Each camera utilizes a dedicated SDI output connection to transmit its video signal.

B. Signal Routing and Switching​

  • All 50x SDI video feeds are physically connected to the inputs of a Blackmagic Smart Videohub 72x72 router.
  • 40x SDI outputs are patched from the Videohub to the 40 SDI inputs of a Blackmagic ATEM switcher, such as an ATEM Constellation 8K.

C. Capture and Host System​

  • A single SDI video output, the program output, is routed from the ATEM switcher.
  • This signal is captured by a Blackmagic SDI video capture card, capable of 8K resolution, installed within a Lenovo Workstation host computer.

D. Color Grading and Multi-View Application​

  • The workstation runs a DaVinci Resolve Studio “Live Grade” session that monitors the incoming signal.
  • The Live Grade process creates a DRX/LUT file.
  • This DRX/LUT is applied to enable multi-view color grading on the entire array’s footage.
  • This is possible using either the DRX file with the DaVinci Resolve Open FX Renderer plugin in any OpenFX-compatible host, or a standard LUT-based color grade.

E. 3D Model Asset Generation​

  • The multi-view media is pulled as a set of timecode-synchronized stills at a predefined interval from the camera array.
  • This footage is graded and then routed for processing into a Docker container.
  • The container can be run on an existing Rocky Linux 10-based grading suite with an NVIDIA GPU, or pushed as a background task to a cloud-hosted GPU compute instance on a platform such as Amazon AWS or Google GCP Cloud.
  • Inside the container:
    • COLMAP runs Structure-from-Motion (SfM) to solve for camera positions and point clouds.
    • The NVIDIA 3DGRUT library executes 3D Gaussian Raytracing for training the final model.
  • The workflow concludes with the generation of a 3DGS (Gaussian Splat) trained 3D model asset.

System Specifications​

ComponentSpecification
CameraRED Digital Cinemas
Camera Count50 cameras
Native Video Resolution6K
RouterBlackmagic Smart Videohub 72x72
SwitcherBlackmagic ATEM, with 40 SDI inputs
Capture CardBlackmagic SDI, 8K capable
WorkstationLenovo Workstation
SoftwareDaVinci Resolve Studio, Live Grade
3D ProcessingDocker + COLMAP + NVIDIA 3DGRUT
Output3DGS (Gaussian Splat) 3D Model