Between the tooling and the film — one prose prompt, "Coalition dry-dock, abyssal dunes, low-angle 75 mm Kodachrome" — walked through every node of the integrated stack, desk to 3D-printed stage to dropped-in plate.
ZForge is the tooling — a 16-tool studio on a math substrate. Gothic Mechanica is the film — Kodachrome winter, Burkean sublime. This page outlines the workflow that joins them: the Hybrid 3D × AI pipeline diagram, and the Scene Stager · Inverse Cinematography Loop that lets the look refine itself. Mechanica Composer generates a reference card; ZForge Scene Composer reads it and dispatches four skills; the orchestrator merges one USD scene; that scene branches three ways — digital, physical, latent — and the recorded plate drops back into the corpus. The path broadens while the look refines.
This page is a staged walkthrough of the designed workflow. Clips marked AUTHENTIC are real captures of tools running today; clips marked SIMULATED stand in for hardware steps (LED dome, 3D print, robotic rig) until the physical stage is built. Each slot is a drop-in — simulated footage is replaced by the real take as it's shot, with no re-edit.
Scene Stager runs cinematography backwards. Give it a reference image and it
extracts the measurable look — key-light azimuth, chiaroscuro ratio, white balance, lens compression — fuses them
into a SceneScript,
renders a preview, then measures the preview the same way. The difference Δ = preview − reference
drives the next adjustment. The loop closes when |Δ| < tol —
typically in three iterations.
The AI proposes slot adjustments only —
light.key_light_azimuth_deg: +12 —
never raw XYZ. TargetLink and the rig profiles own deterministic placement, so the loop can be aggressive about the
look while placement stays exact and reproducible.
When the loop converges, the settled parameters bake into a permanent
FilmStockProfile /
MotionMacro /
LightingPreset.
Classic mode then inherits every AI gain at zero API cost.
The prose prompt is typed into Mechanica Composer. It generates a "DOCKING BAY 001 / NIGHT" overlay-text card with a Kodachrome aesthetic — overlay font and position chosen, an audio graft from the late-Romantic adagio corpus, alt-text describing the frame. The finished reference card saves to the Library.
→ produces reference_card.json · overlay.png · audio_graft.mp3 · alt_text.txt
The reference card is dropped onto ZForge Scene Composer. The VLM director writes a five-step plan in real time —
KitbashConductor.retrieve(freighter) → SceneStager.fuse(reference) → TopoDoctor.repair(retrieved) → HolyGrail.quadify(boolean) → OrchestratedScene.assemble()
— displayed as a dependency DAG with a cost estimate.
Split-screen: each skill runs in its own panel. Kitbash shows the top-K freighter retrievals with symmetry tags. Scene Stager's calibration loop converges on the reference's chiaroscuro and warm white-balance in three iterations. Topo Doctor flashes defect classifications and fixes. HolyGrail rebuilds quads from the boolean intersection. The orchestrator merges the USD prims into one scene graph, live.
→ produces SceneScript.json · calibration_session.json · assembled_scene.usdThe same USD asset, three outputs in a triptych. One scene graph drives a digital Unreal stage, a physical 3D-printed mini-bigature on an LED backdrop, and a latent diffusion dome — all from a single source of truth.
The final recorded plate — the tracked-camera composite of the printed mini-bigature against the LED-stage backdrop — drops back into Mechanica Composer as a new training reference. The corpus expands: 5 frames, 20, 100, recursive.
"The path broadens while the look refines — anchored by a collaborative compositional array."
This walkthrough is staged from the design at
_shared/DEMO_WORKFLOW_DESIGN.md.
Acts I–III capture tools that run today; Acts IV(b/c) and V are simulated until the physical stage —
3D printer, LED dome, tracked camera — is assembled. Every video slot is a named drop-in, so each
simulated clip is replaced by the authentic take as it's shot.