Overview

LARGE is the top tier of the Lenscowboy SaaS hierarchy (Creator → Influencer → Studio → Enterprise → LARGE). Where the lower tiers focus on a particular surface — script breakdown (LCBE), previz and layout (PAL), client review (SPECTACLE), editor handoff (Resolver) — LARGE is the layer that conducts all of them, plus external DCCs and generative vendors, under one data agreement.

The technical centrepiece is a G-buffer broker that preserves spatial channels (depth, normals, position, motion vectors, ID mattes, segmentation) through generative passes. The result is the round-trip commitment: when a director gives a note ("hero's jacket two shades darker"), only the noted thing changes. Spatial layout, character identity, lighting intent — everything else holds.

The Channel Contract

The Channel Contract is the data agreement every component in LARGE participates in. Each generation vendor declares which datasets it consumes and which it produces; the broker routes bundles to a vendor only if the capability surfaces match. Each Shot Bundle the broker dispatches is a folded set of dataset instances with stable IDs, hashes, and rights tiers.

The Contract's vocabulary is VES dataset IDs. We don't invent terminology. We point at the canonical reference the traditional VFX community already uses, and we point back.

VES alignment

The VES On-Set VFX Data Collection & Usage Guide v1.1.0 (published 2026-05-12 by the VES Technology Committee, Apache 2.0 licensed) catalogues 68 datasets across 17 categories. Each dataset is tagged with:

• Description & Usage
• Scope — one of 21 terms (Take, Shot, Sequence, Set, Lens, Camera, Actor, Production Location, Virtual Stage, …)
• Creator — who makes it on set
• Consumer — which departments and vendors use it downstream
• VFX Types — complexity filter (Basic 2D, Character VFX, Complex VFX, Motion Capture, Virtual Production, Real Time, Digital Matte Painting)

LARGE vendors the canonical JSON at large/ves_taxonomy/data.json under Apache 2.0 with full attribution, and exposes it to the rest of the stack through large/broker/ves_taxonomy.py — a typed Python loader with lookups by creator, consumer, scope, VFX type, and category. Slug IDs are stable across runs:

from large.broker.ves_taxonomy import load_taxonomy

t = load_taxonomy()
face_scans = t.get("ves.5.5.face_scans_body_scans")
camera_dept_datasets = t.by_creator("Camera Department")
previz_inputs = t.by_consumer("Previz Vendor")

Integration across the stack

VES adoption is not LARGE-tier exclusive. The taxonomy threads through every module the Lenscowboy platform already ships.

LCBE — Breakdown Engine

LCBE's existing entity model already speaks VES vocabulary: Shot / Scene / Sequence / Character / Location / Prop / Actor are all VES scope terms. Adoption is mostly a metadata enrichment:

• Each shot record gains ves_datasets_consumed[] — populated by the parser when reference uploads happen, and by the operator when explicit references are declared in the script bible.
• Reference uploads are tagged with their VES ID at write time. Face scans → ves.5.5. Photogrammetry → ves.5.3. Prop scans → ves.5.2. Set reference photography → ves.5.6.
• The VFX detector's complexity scoring maps onto the 7 VES VFX Types, so the breakdown speaks to traditional supervisors in their own filter language.
• The Score Generator and cost engine can speak cost-per-VES-dataset alongside cost-per-shot — useful for producers reconciling against traditional-VFX budget templates.

PAL — Previz & Layout

PAL outputs are literally VES ves.7.1.previz_techviz dataset instances. Imports map cleanly:

The PAL freeze-to-GCS layer preserves VES tags onto frozen GLBs, so the external renderer service knows it's loading a prop scan, not just an opaque mesh. PAL exports carry VES IDs to downstream generative consumers without translation.

Pipeline — Generation

The Daily / Runs / 2D generation pipelines and the LSV provenance ledger gain VES vocabulary at the recording layer:

• Each call into compliance/pipeline_hook.py records ves_datasets_consumed[] (reference images, source media, prompts) and ves_datasets_produced[] (the generated artefact).
• Vendor adapters in src/vendors_new/ declare their VES capability surface — a routable contract, not just an endpoint URL.
• The Extracted Prompt log column gains a paired ves_input_datasets column when references are typed.

Comfy PAL Node — the open edge

The Comfy PAL node is Lenscowboy's integration into the ComfyUI ecosystem. Two VES-native nodes are planned alongside the existing PAL integration:

• VES Dataset Loader — reads a VES bundle from disk and routes the right asset (face scan, LIDAR, previz video, camera track) to downstream nodes via typed sockets.
• VES Dataset Emitter — wraps a Comfy workflow output as a VES-typed dataset entry with provenance metadata for the ledger.

The result: ComfyUI users get a path to compliance-grade output without leaving their workflow editor. The open generative ecosystem and the broadcasters consuming its output speak the same vocabulary.

DCC adapter roadmap

Nine production tools are first-class peers under the Channel Contract. They span three relationships: adapter targets (DCCs where LARGE installs a plugin that reads and writes VES-typed bundles), finishing tools (compositors and editors where LARGE exports cleanly without trying to replicate their work), and open ecosystems (engines like ComfyUI where LARGE builds on top and contributes back to mainline). We ship adapters one at a time so each one is real before the next is announced.

Complement, don't compete

The relationship column above is load-bearing. LARGE does not aim to replace any of these tools. Nuke and Fusion have decades of refinement in production compositing that we will not try to match — we hand off to them cleanly via export. ComfyUI is the standard for AI workflow definition and execution; we build on top of it rather than forking it, and we contribute back so the ecosystem benefits from Lenscowboy's production-grade additions. Adapter targets get plugins; finishing tools get exports; open ecosystems get contributions. LARGE's role is the layer above — the production-aware conductor that none of those tools occupy — and that is where the moat lives.

Provenance & rights

Every dispatched bundle carries a rights manifest that gates which vendor tier the broker is allowed to send it to:

The provenance side lives in the LSV (Lenscowboy Studio Vault) module: hash-chained, tamper-evident ledger; per-step VES dataset attribution; C2PA Content Credentials integration in progress. The Channel Contract and the ledger speak the same vocabulary, so every recorded step is legible to compliance reviewers without translation.

Giving back to the VES community

Four contribution-back commitments, all Apache 2.0:

1. Empirical schema feedback — observations from running a real broker, filed as GitHub issues / PRs on richardssam/on-set.
2. Category 17: Generative Production Data — proposed extension covering model identity, prompt records, training-data attestations, LoRA fingerprints, dispatch logs, C2PA manifests, and retention / deletion certificates. Drafted now; contributed upstream once stable.
3. Open Python reference loader — the typed loader in large/broker/ves_taxonomy.py extracted as a small open-source package.
4. C2PA × VES assertion mapping — published once stable; benefits both the VES TC and the C2PA TC.

Credit