Cinematic camera physics married to dynamic lighting.
Seven rig profiles. Eight starter scenarios. Target-linked lights that hold exposure as the camera moves. An AI Inverse Cinematography loop that matches reference frames in 30–90 seconds. Bakes to keyframes in Blender, Maya, and Unreal — same math, same output, every host.
STUDIO · TIER III
The 6-axis arm. The LED volume. Unreal driving the wall in real time. A techno-dolly that hits the same mark to a tenth of a frame, take after take. Form Studios, Pixomondo PXO AKIRA, MRMC have spent the last five years productionising it.
What you didn't buy is the layer above it — the one that turns a reference frame, a storyboard panel, or one line of prose into the camera move, the key-fill ratio, and the volumetric fall-off that match it. Today a DP and a tech-vis lead sit in a room and translate intent into fcurves by hand. That's the layer Scene Stager replaces.
A reference still goes in. The Inverse Cinematography loop recovers the camera move, the
key-fill ratio, and the volumetric fall-off that match it, then bakes deterministic fcurves on
whichever host you point it at: Blender, Maya, or Unreal.
The intermediate representation and the
inversion method are proprietary — withheld pending filing.
The techno-dolly path, the
UE VCam
animation, and the Maya post-shoot replicate are the same keyframes, by construction.
No per-host drift. No re-keying between previz and final.
Eval(plan, manifest) → USD, pure.
Five tools, one deterministic math substrate, one USD output. They sit on top of your robotic stage the way Sequencer sits on top of Unreal — as the layer that turns intent into the exact motion the hardware will execute.
These are the renderers and the rigs. ZForge is the agentic layer that decides what they execute.
The same deterministic compiler keeps shipping into new surfaces. Some are already green; some are staged for the next cycle. None of them require rewriting the core — every new adapter just exposes another DCC or another output modality to the SceneScript that already runs.
1360 / 1360 tests green. Blender ↔ Maya ↔ Unreal bake parity already proven across the line.
Mock-studio loop: UE5 → 3D print → LED dome → recorded plate → corpus. The staged walkthrough is live; authentic captures drop in act by act.
▸ Watch the demo →Fourth DCC adapter on the same compiler. Same SceneScript, same bake guarantees.
Fifth output modality joining USD / 3D print / HDR dome. Diffusion-extended infinity backdrops trained on the scene's motif corpus.
DPs ship their own LUTs straight into the SceneScript vocabulary. Custom stocks alongside Kodachrome 25 / Vision3 / Tri-X.
Updated against the live .planning/STATE.md on each milestone close.
What ships next is decided by which adapter the queued shots demand — not by a marketing roadmap.
Originally shipped as two separate products — Light Stager and Camera Director —
they collapsed into Scene Stager the moment we realised they were two halves of the same idea.
The lights and the camera obey the same physics. The bake is one operator.
Pick a rig. Pick a scenario. Bake. Done.
Lights parented to the focal target stay perfectly exposed as the camera moves. A single slider drives wind drift, fog scatter, and light flicker together.
Hand it a reference still. It analyses, fits a matched look, renders a preview, critiques, calibrates, iterates — to convergence.
30–90 seconds per match. Batch operation.
Each heuristic in the loop has a theorem-grade counterpart with named, bounded failure modes — closed-form constructions from discrete topology and Riemannian geometry.
The specific constructions — patent-candidate methods — withheld pending filing.
Scene Stager is a typed compiler with an AI calibration loop bolted onto its boundary. The geometric tranche replaces every estimator inside that loop with a closed-form construction from discrete topology, persistent homology, and Riemannian geometry — theorem-grade where the heuristic baseline was tuned.
The specific constructions that make the loop theorem-grade — each mapping a named closed-form geometric result onto an estimator the loop would otherwise eyeball — are patent-candidate methods. The constructions, the math-to-effect mappings, and the equations are proprietary and withheld pending filing. The public claim stands: closed-form, host-portable, fewer iterations to converge, robust to clutter.
Scene Stager is a layered system — a typed substrate and compiler, host adapters for Blender · Maya · Unreal, an image-metrics front-end, a fuser + calibration loop, a persistence layer, an AI inspection/promotion layer, and the geometric topology core. The specific module decomposition and the algorithms inside each layer are proprietary patent-candidate methods and are withheld pending filing.
What's public is the envelope: ~1,000+ tests green cross-DCC parity by construction pure-math IR boundary.
Four reference plates, in the same Nostromo Deep treatment as the math page. The schemas behind them — the typed scene representation, the rig taxonomy, the closed inverse-cinematography loop, and the atmospheric phase broker — are withheld pending filing; the plates below show the envelope only.
Scene Stager compiles free prose into a compact, typed scene representation: a small set of controlled-vocabulary slots that together pin shot, lens, detail, cinema, style, light, and atmosphere. The exact slot set, vocabularies, and target bindings are a patent-candidate method and are withheld pending filing.
The Inverse Cinematography loop matches reference stills like these — drop a frame in, get a matched 3D scene back, and ship it as a compact preset any host can re-bake. Method withheld pending filing.
The geometric tranche is fully realised. The remaining items are integration edges between what shipped and the real world — each one is small, isolated, and individually mappable.
Create the draft in the Gumroad UI · assign gumroad_product_id to the manifest · run publish.py --no-dry-run with GUMROAD_API_TOKEN. Cover, listing, dry-run all green.
Final shakedown of the matched-preview pass inside live Blender. The harness is staged and tested headless; remaining work is the live-render confirmation. Method withheld pending filing.
MISSING_TARGET_* spawn
Blender & Maya adapters honour the universal ProxyNullSpec contract by spawning Empty / Locator on unresolved selectors. Universal-layer contract shipped (Option 3); host integration is a small adapter PR.
Drop 3–5 Eevee, Cycles, and Lumen reference renders into assets/calibration_refs/. The per-renderer attached fields (qualitative baselines shipped) re-fit to ground truth via the calibration loop itself.
An adjacent research implementation converges onto the shared topology classifier extracted earlier. Cross-citation in place; conversion is incremental. Method withheld pending filing.
Scene Stager is one node in a larger workflow: a hybrid 3D × AI film pipeline where image-driven generation, virtual cinematography, and physical 3D-printed mini-bigatures all share one USD substrate. Scene Stager drives lighting and camera physics from reference images. Kitbash Conductor browses asset libraries by shape. Topo Doctor enforces interchange-grade meshes between every stage. HolyGrail produces clean quad topology from triangulated booleans. Scene Composer orchestrates them all from prose prompts.
Three tested producers converge into one deterministic bake. The plumbing is already built — what stays public is the promise, not the recipe. The convergence method is a patent-candidate invention and is withheld pending filing.
The typed scene representation, the convergence socket, the additive merge, the cross-DCC parity guarantee, and the function that lands one producer's output into the other are proprietary and withheld pending filing. What stays public: reference image + silhouette / intent → staged, lit, filmed 3D scene → identical bake across Blender · Maya · Unreal.
Sequencer · EXR · ProRes. UE5 Symmetry/PCG widgets. Tracked camera + 3D-stage match. RED-VRAPTOR 8K VV virtual cinematography.
Same USD asset, printed at scale (1 uu = 1 cm). HDR dome / LED stage backdrop. Camera match. In-camera weight.
Diffusion-extended infinity backdrops trained on the scene's motif corpus. Non-linear feedback loop into ideation.
Sketch → photoreal → curate → custom LDM → image→3D → ZBrush sculpt → Maya retopo → USD interchange → UE5 assembly → variant generation → branch to digital / physical / latent. The eleven-step pipeline that ties every ZForge AI-tier tool together. Scene Stager is the lighting/camera node.
Read the workflow PDF →
