Capture to Delivery Workflow
This page walks the end-to-end Fluxx workflow as a single chain of operations: pre-production viz cascade, shoot day capture against an LED wall (or green screen, or location), Vault ingest, split compute render, Near-Time PostViz return to set, overnight high-res final pixel, and per-shot routing into one of two finishing tracks. Production crews, vendors, studio leadership, and investors read this page to see how the five Fluxx-native primitives stop being separate components and start being one continuous workflow that runs from camera sensor to final composite.
The workflow at a glance
- Pre-production viz cascade builds the picture: TechViz aligns geometry, LightViz validates lighting against the on-set rig, CompViz pre-builds the Nuke auto-assembly script.
- Shoot day: Warp drives the LED wall, runs the real-time renderer, and handles ComposiTrack sub-frame processing in-line.
- Capture: three ComposiTrack sub-frames are captured per take and land in the Vault as first-class Representations.
- Fluxx Automate: absorbs the capture, attaches camera, lens, tracking, and pipeline provenance, content-addresses the sub-frames into the Vault.
- Fabrixx: renders at quality against the LightViz lighting recipe, applies the pre-comp recipe in Nuke via the CompViz script.
- Near-Time PostViz returns to set in 30 to 60 minutes at quarter resolution while the unit is still on stage.
- Overnight high-res render produces final pixel; per-shot finish-track routing decides Near-Time or Near-Time VFX.
Phase 1, Pre-production viz cascade
Most of the creative VFX work happens upstream of the camera. By the time the shoot day arrives, the picture has largely been built. The viz cascade runs in three bounded sequential phases.
TechViz aligns geometry. The virtual asset development (VAD) and the practical set are reconciled to one coordinate system from the camera’s point of view. By the end of TechViz, what the camera sees in the digital build matches what it will see when the practical pieces arrive on stage. LightViz validates lighting. Virtual lights established during the lighting build are matched against the physical lights of the on-set rig until the digital and physical environments cast the same shadows on the same surfaces. CompViz pre-builds compositing automation. The Nuke script that will assemble the render passes during the shoot is constructed and tested against representative inputs: lens corrections, color management, plate integration, shot-family-specific comp operations.
What pre-production produces for the rest of the workflow is two artifacts: the lighting recipe (every virtual light’s position, color, intensity, and falloff, locked in LightViz) and the Nuke script that auto-assembles render passes at runtime (the CompViz output). Both are inputs Fabrixx will consume the moment a shot is cut. See Viz Cascade for the per-phase deep dive.
Phase 2, Shoot day capture
The shoot day runs against the picture pre-production already built. The crew lights the immediate set and the actors. The camera frames against the LED wall, the green screen, or the location, as the day calls for. Warp drives the on-set compute layer: 10x NVIDIA RTX Pro 6000 GPUs in 5 flight cases, SMPTE 2110 transport, mobile and liquid-cooled. Warp is sized for crew-flow only. It runs the LED wall real-time renderer, handles ComposiTrack sub-frame processing in-line, and absorbs camera-layer feeds. It does not do the render-quality work; that lives in Fabrixx.
In the microsecond between the display frames the human eye sees, the wall flashes a chroma-shifted version of the same background image. The shift is to the average color of the plate background, not a saturated green or blue. The camera, synchronized to the wall’s refresh, captures both the lit scene as displayed and the chroma-shifted reference. The difference between the two captured frames is a mathematically perfect matte. No green spill, no blue spill, no compromise in the cinematographer’s lighting.
Three ComposiTrack sub-frames land in the Vault as first-class capture artifacts per take: ChromaTrack (final-image feed), ContentTrack (derived matte), and CompensationTrack (compensation pass). See On-Set Operations for the per-take detail. Executive Director Ari Jampolsky owns the on-set workflow.
Phase 3, Fluxx Automate plus Vault commit
The director calls cut. Fluxx Automate fires the cut-to-render chain. The captured plate is transcoded to a quarter-resolution proxy on a dedicated transcode flight case. Lens metadata (focal length, distortion, aperture) and 3D camera tracking metadata from Luster are attached. The three ComposiTrack sub-frames plus the proxy plus the metadata bundle move into the Vault.
The Vault is the substrate. It is content-addressed at the storage tier: asset identity is a hash, not a file system path. Storage backends are abstracted (S3, GCS, Azure Blob). Provenance is a first-class field at every Version: camera origin (sensor, lens, takes, timecode), pipeline origin (which render pass, which agent, which artist), and AI origin where applicable (prompt, seed, model, model_version, generation_tool, parent_asset_id). Multi-tenant scoping isolates the production’s perimeter; the studio owns the data inside that perimeter.
The shot is created automatically in ftrack on-set as part of the same chain. No manual ingest. No after-hours producer triage. The plate is ready to composite the moment the camera cuts because the matte exists at capture and the metadata is wired before the shot reaches Fabrixx.
Phase 4, Split compute render
Both compute layers carry equal weight. Warp runs on-set crew-flow. Fabrixx runs the render economics. The architecture is split deliberately because even at lower resolutions, on-set proxy playblast rendering requires complex rendering in layers that on-set hardware cannot deliver fast enough; crew waits, cost balloons. The render workload is sized for the cloud.
Fabrixx receives the proxy plus lens data plus tracking plus ComposiTrack matte. A virtual machine is provisioned (the pipeline service owns VM lifecycle as a first-class event). The cloud brute-forces the render passes against the LightViz lighting recipe. This is not a real-time render. The cloud has minutes, not 20 milliseconds. It does the actual physics: ray tracing, light scatter, full-quality textures. Then the CompViz Nuke script picks up the rendered passes and auto-assembles the slap comp. No artist intervention per shot. The automation runs.
Near-Time PostViz returns to set 30 to 60 minutes after the camera cuts, at quarter resolution, on a monitor while the creatives are still on stage. The director can make a creative call now, while the change is still cheap, instead of finding out three months later in editorial. See Cloud-Side Operations for the per-step detail.
Phase 5, Overnight high-res final pixel
The high-resolution brute-force render runs overnight against the same lighting recipe, against the same auto-assembly script, against the same ComposiTrack plate. The output lands in the Vault as final pixel. The next morning, every shot has a finish_track field assigned: Near-Time or Near-Time VFX.
The shot opens back up cleanly. The ComposiTrack clean-key plate, the pre-comp recipe, and the USD scene graph all stay live and re-renderable after first delivery. If the director changes their mind on the digital forest three weeks later, the shot does not require roto recovery to peel the actors out of baked-in pixels. The substrate is intact. The reopen is a re-render, not a rescue operation.
Phase 6, Finish-track routing
Every shot lands in one of two tracks. Same render upstream, same metadata, same Vault.
Near-Time track. The overnight high-res render is the deliverable. A compositing artist does a quick standard pass for color balance, plate integration, and polish, and the shot ships. This track carries the bulk of runtime on any show: dialogue scenes in built environments, action coverage with set extensions, anything that does not need a hero pass.
Near-Time VFX track. The same render is picked up for hero finishing. PostViz first look, slap comp at 80 to 90 percent final, director call, polish on keyshots, Agentic Sequence Assist propagation across the sequence, comp review, delivery. Different polish budget, identical pipeline upstream.
Type of show and type of shot decide which track a shot rides on, not the shoot type and not the budget. A tentpole might run Near-Time VFX for hero beats and Near-Time for the rest; a streaming series might run almost entirely Near-Time; a music video might run entirely Near-Time for iteration speed. Shots can flip tracks mid-pipeline without leaving the system. See Two Finishing Tracks.
What makes this workflow possible
The five Fluxx-native primitives are load-bearing as a unit. Capture-time clean key (ComposiTrack), split compute (Warp plus Fabrixx at equal weight), content-addressed Vault, two-track finish through one workflow, pipeline-owned VM lifecycle. Remove any one and the chain breaks. ComposiTrack alone without Fabrixx cannot return Near-Time PostViz; Fabrixx alone without the Vault has no substrate to render against; the Vault alone without the two-track finish has no per-shot routing primitive. The combination is what runs camera sensor to final composite as one continuous workflow. See Pipeline Architecture for the per-primitive deep dive.
Cross-references
- Pipeline Architecture (the five primitives in detail)
- Viz Cascade (Phase 1 deep dive)
- On-Set Operations (Phase 2 deep dive)
- Cloud-Side Operations (Phase 4 deep dive)
- Two Finishing Tracks (Phase 6 deep dive)
- Q-G31 Capture To Delivery Workflow at
04_Fluxx_Shared/_Master_Canon/Q-G/Q-G31_Capture_To_Delivery_Workflow.md - Q-G45 / Q-G47 Two Finishing Tracks (both files exist; same locked canon)
- Q-G49 Compute Architecture
- Q-G88 ComposiTrack Canonical Name
Sources
02_Working/Strategy/Company_Overview/Compute_Architecture_Source_2026_05_03.md02_Working/Strategy/Company_Overview/Finishing_Workflow_Source_2026_05_01.md02_Working/Strategy/Company_Overview/Fluxx_White_Paper_Near_Time_VFX_DRAFT_v0.1.md§402_Working/Pipeline_Research/Fluxx_Pipeline_Architecture_v0.1.md
Audience visibility
industry, vendor, production, sales, director, investor.