Skip to content
Fluxx

AI Provenance and Human Likeness Policy

Fluxx’s stance on generative AI is hybrid by design and bounded by one hard wall: human performance and likeness (revised S145, 2026-05-30). Generative AI is welcome in the workflow as a source of intermediate elements, and may deliver the final pixel for full-CG shots that carry no recognizable human performance or likeness (establishing shots, cityscapes, space battles, large-scale destruction, environments, spectacle). The governing test is “actor-driven vs. not,” never “action vs. non-action.” Performance-driven and likeness-bearing pixels still pass through traditional rendering and compositing in V-Ray, Nuke, or Houdini; human performance and likeness are never synthesized. The Vault is the system of record for content origin regardless of how an element was created (camera-captured, hand-built, agent-populated, prompt-adjusted, or generative final-pixel), which gives the production a complete provenance chain. The human-likeness policy below is the wall, not the pixel: AI as labor automation is allowed, AI as performance synthesis is not.

Vault provenance schema

The Vault carries provenance as a first-class field at every Version. Schema:

  • prompt (text input that generated the asset)
  • seed (random seed for reproducibility)
  • model (foundation model name)
  • model_version (specific weights version)
  • LoRA (any LoRA adapters applied)
  • generation_timestamp (UTC timestamp at generation)
  • generation_tool (the surface that generated the asset, e.g. Runway, Stable Diffusion, internal)
  • parent_asset_id (provenance chain to predecessor assets)
  • C2PA Content Credentials pass-through (industry standard for provenance metadata)
  • IPTC Digital Source Type (where applicable)

Plus camera origin fields (sensor, lens, takes, timecode) for captured assets, and pipeline origin fields (render pass, artist, agent) for processed assets. The schema is the same shape regardless of asset source; the populated fields differ.

EU AI Act Article 50 compliance by default

The EU AI Act Article 50 disclosure deadline is 2 August 2026. Productions running on Fluxx meet the studio’s disclosure obligations by default because the Vault carries the required schema fields at the substrate tier. No additional vendor product is needed; no per-shot disclosure tracking workflow is needed. Fluxx does not hold the disclosure obligation, the studio does; Fluxx supplies the provenance infrastructure that makes the studio’s disclosures factually reliable, and follows the studio’s lead on what gets disclosed and how.

Open items

  • Consent metadata for performer-scanned crowds. Schema fields not yet locked. Tied to the SAG-AFTRA contractual alignment question (Q-G50 open thread): does the crowd-scanning-and-payment provision require a specific contract structure that the Vault should track?
  • Royalty tracking fields. Not addressed in current schema; surface when a production-side workflow firms up.
  • Performer estate / posthumous use policy. Not addressed; separate lock when it comes up.

Human likeness policy

The policy line is between AI as labor automation (allowed) and AI as performance synthesis (not allowed). Three categories.

Crowds: yes, with scanning and payment

Generative AI is used for crowds: background extras, stadium fills, large-scale crowd simulation. Faces in crowds are scanned from real performers. The performer is paid. AI generates the body and motion within the bounds of the scan. The performer’s likeness rights are honored and the performer participates in the economics; the labor of generating thousands of bodies is automated. The performer-scanning-and-payment provision is canonical, not optional; drop it and the policy reads as exploitative.

Performance: never

Lead performances, named-character performances, supporting performances, any performer the audience is meant to recognize as a character carrying the scene. No generative AI synthesis of a performance. No AI-driven facial replacement on a performance. No generative-AI face replacement on a known face. The line is performance: when a person is delivering a performance through their face, generative AI does not replace them.

Narrow disfigurement exception

If a face is being destroyed in-shot, a blowout explosion, a disintegration effect, a creature transformation, generative AI can be used to fabricate the disfigured face because no performance is being delivered through that face. The face exists for one beat then ceases to be a face. The exception is narrow by design and does not expand.

Why this line

The boundary between labor automation (allowed) and performance synthesis (not allowed) is the boundary between extending what artists can do and replacing what performers do. Fluxx is on the artist-and-performer side of that line. The policy is decision-grade for productions in jurisdictions with performer likeness rights, essentially all union productions. It is also brand-grade for Fluxx: productions that adopt Fluxx do not need to ask each show whether the policy line will be honored.

One-line for press: Generative AI for crowds, yes, with performer scanning and payment. Generative AI for performance, never. Narrow disfigurement exception. The line is performance, not labor.

Final delivered pixels

The final-pixel wall was retired at S145 (2026-05-30); the wall now sits at performance and likeness. Generative AI may contribute intermediate elements (particle simulations, plate enhancements, populated builds derived from a lead artist’s hand-built keyframes via the Agentic Build Assistant), AND it may deliver the final pixel for full-CG shots that carry no recognizable human performance or likeness, the traditionally most expensive shots in a film: establishing shots and cityscapes, space battles, large-scale destruction, vehicles, non-performing creatures, vistas. The reason is cost and quality: these shots are traditionally built fully in CG and gating them behind a final-pixel wall taxed exactly the work where generative is safest and most valuable.

The governing test is actor-driven, not action. A space battle is action-heavy but performance-free, so generative may take it end to end; a fight carried by a recognizable lead is performance-driven, so it stays traditional. If a recognizable performer’s face or performance appears inside an otherwise-generative shot, that element is governed by the human-likeness policy above (not generative); a digital double is allowed only at distant crowd or spectacle scale, never standing in for a recognizable lead.

Performance-driven and likeness-bearing final frames still pass through V-Ray, Nuke, or Houdini, with compositing as the finishing layer. Per Q-G33, the agentic and automation work sits at Layer 3 of the architecture (Automation plus Agentic Pipeline Assist). Provenance is unchanged and more load-bearing: every generative element, including these new final-pixel shots, is tracked in the Vault, which is exactly what a studio needs for its EU AI Act Article 50 disclosures.

Operational enforcement

The policy is enforced operationally through the plugin layer. A provenance plugin interface flags AI-generated content at ingest time. The provenance chain is queryable: the Vault can answer “was any performer-identity AI-generated for this shot?” by traversing the schema. Counsel evaluates per-shot; production has the audit trail.

Cross-references

  • Pipeline Architecture (primitive 3: Vault as substrate)
  • Plugin SDK (provenance plugin interface)
  • Q-G18 AI Provenance Stance at 04_Fluxx_Shared/_Master_Canon/Q-G/Q-G18_AI_Provenance_Stance.md
  • Q-G50 AI Human Likeness Policy at 04_Fluxx_Shared/_Master_Canon/Q-G/Q-G50_AI_Human_Likeness_Policy.md
  • Q-G33 Three Layers Architecture (the performance-and-likeness bounding line; per S145 the bound is performance/likeness, not the final pixel)

Sources

  • 02_Working/Strategy/Company_Overview/AI_Provenance_Source_2026_04_28.md
  • 02_Working/Strategy/Company_Overview/AI_Human_Policy_Source_2026_05_03.md

Audience visibility

industry, sales, director, production, investor, counsel.