Hollywood's VFX Workforce Could Shrink 80%. Promise Studios Has Google's Models Powering the Cut.

MUSE OS and the Google DeepMind Stack

A studio founded by YouTube veterans and an AI filmmaker just got Google's most advanced video and reasoning models embedded directly into its production pipeline, and the first feature is shooting this year.

Promise Advanced Imagination launched in November 2024 on a thesis: generative AI belongs at the center of filmmaking, not the periphery. Founders George Strompolos (Fullscreen), Jamie Byrne (YouTube), and AI filmmaker Dave Clark built a proprietary operating system called MUSE — Promise's production pipeline and workflow software — to run the entire production lifecycle on one AI-augmented platform. In May 2025, Google's AI Futures Fund led a strategic investment that embedded DeepMind engineers inside Promise's development loop, giving MUSE native access to Gemini 2.5 Pro Preview and Veo 2.

This is not a plugin arrangement. DeepMind researchers collaborate with Promise's teams to adapt Gemini and Veo for cinematic storytelling. Inside MUSE, Gemini performs deep script analysis (mapping character arcs, identifying thematic through-lines, and suggesting visual enhancements from a human-authored screenplay). Directors upload hand-drawn character art and scene descriptions; Gemini generates dynamic storyboards that feed directly into Veo for pre-visualization, VFX concept reels, and production-quality mood boards.

Veo 3.1's "Ingredients to Video" feature, updated in January 2026 with stronger prompt adherence, improved audiovisual fidelity, vertical format support, and 4K upscaling, lets filmmakers anchor generated footage to specific visual references, reducing the slot-machine randomness that plagued earlier video models.

MUSE's architecture spans pre-visualization, scriptwriting, character design, budget visualization, and post-production in one environment. Its modular design ingests best-in-class GenAI tools as they emerge, but the Google partnership locks in a privileged access tier: early testing, technical co-development, and co-marketing. Jonathan Silber, co-founder and director of Google's AI Futures Fund, called Promise "one of the most forward-thinking studios we've encountered — operating on the leading edge of what's creatively possible with generative AI."

The stack's first stress test is a feature film slated for production this year. Details remain under wraps, but the project is designed as a proof-of-concept for AI-assisted workflows that compress months of location scouting, asset creation, and pre-vis into days. Promise's acquisition of Curious Refuge, the world's largest GenAI film school, supplies a talent pipeline trained on these tools, reducing dependence on traditional VFX specialists. The investor roster reads like a cross-section of media-tech power: Andreessen Horowitz and Peter Chernin's North Road Company co-led the seed round; the May 2025 round added Crossbeam Venture Partners (Michael Ovitz), Kivu Ventures, Saga Ventures, Offline Ventures, plus angels Emery Wells (Frame.io) and Aaron Levie (Box).

What happens when a studio's core infrastructure is built around models that improve weekly? The answer will shape whether MUSE becomes a proprietary advantage or a licensable layer for the rest of Hollywood.

Where the Money Goes: VFX Budgets Meet AI

Venture capital investment in AI surged to $202.3 billion in 2025, up 75% year over year. Hollywood studios themselves spent $450 million on AI technology in 2023, a 60 percent jump. Capital is betting that generative AI will rewrite the cost structure of filmed entertainment.

The VFX line item — historically 15 to 50 percent of a production budget — is the primary target. The Independent Producers Guild projects 80 to 90 percent efficiency gains in specific VFX and 3D asset tasks. Early pilots bear that out across roto, layout, editing, and complex sequences. The data is summarized below.

The economics compound. When cost-per-shot drops, teams don't simply bank the savings. They generate more shots, more versions, more iterations. The LBB analysis of generative AI cost compression notes this dynamic: cheaper shots spur more iterations, not just savings. "More shots, more options, more versions, all of which need more reviews in less time." The friction relocates from creation to collaboration.

The strategic decision to use AI is based on creative requirements — such as risk mitigation, flexibility, or IP preservation, rather than simple labor cost avoidance. (Independent Producers Guild, AI in Professional Film Production (2025))

Most post facilities have folded AI into workflows; surveyed VFX houses show majority adoption in a $30 billion sector where AI claims a fifth of the market. Enterprise suites run millions annually; studio subscriptions average a quarter-million. The infrastructure spend is real and recurring.

Netflix's acquisition of Ben Affleck's InterPositive for up to $600 million (Netflix reported) illustrates the stakes. The startup's patent application projects 50 percent cuts to visual effects costs, 70 percent savings on background actors, 40 percent on additional production units, and 30 percent on art department. Principal photography savings are estimated at 10 to 20 percent of total production costs. Applied across Netflix's data shows a $56–84 billion original content spend, a 20 percent efficiency gain translates to $8–17 billion in potential savings. Ted Sarandos has told Wall Street the bigger business is making content 10 percent better, not 50 percent cheaper, but the cost lever unlocks the quality lever at scale.

Wonder Studios (backed by Atomico and Adobe) and Promise (backed by North Road Company) are building slates with an IP catalog model: a library of films and series generates recurring revenue through licensing, franchising, and distribution deals that appreciate as new platforms emerge. The global generative AI in movies market, valued at $367 million in 2023, is projected to reach $3.9 billion by 2033. The AI-generated interactive movies segment could hit $12 billion by 2032.

The labor implications (displacement, reskilling, hybrid role emergence) are the subject of the next section. But the economic die is cast. Studios that don't integrate AI-native pipelines face a structural cost disadvantage that compounds with every production cycle. The question isn't whether VFX budgets shrink. It's whether the savings fund better stories or simply fatter margins.

The Workforce Fracture

The credits of a modern blockbuster read like a phone directory. House of the Dragon and Marvel's Deadpool & Wolverine each list more than 1,000 visual effects workers, including 231 from Industrial Light & Magic alone on the latter, plus hundreds more from Framestore, Weta FX, Barnstorm, Base FX, Raynault FX, Lola, and Rising Sun Pictures. TheWrap reported that once generative AI automates enough tasks, those lists could shrink by 80 percent or more. The math is brutal: an industry built on armies of specialists faces a future where a fraction of that headcount delivers the same pixels.

The workforce was already reeling before AI arrived. The pandemic sparked a brief animation boom as live-action stalled; then the streaming bubble burst. Pixar, DreamWorks, Netflix, and Paramount all trimmed animation teams. Rooster Teeth, Technicolor, and ON Animation closed outright. New work emerged in anime, adult animation, and markets like China and India, but the overall pipeline contracted. IATSE seized the moment, unionizing VFX coordinators at Marvel Studios, a first, though coordinators represent only a sliver of the artists on a major production. As of December 2025, Glassdoor listed roughly 800 VFX artist openings in the Hollywood area, a number that feels generous against the projected cuts.

Source: Cartoon Brew / Luminate Intelligence 2025 survey of entertainment workers

The fear is not abstract. A fall 2025 industry survey identified animators, VFX artists, and concept/storyboard artists as the roles most susceptible to AI-related change. Analyst Audrey Schomer confirmed these professionals are receiving less work; when they do land positions, employers increasingly require fluency in generative tools. Chuck Sheetz, a UCLA professor and former Simpsons animator, stopped telling students the old story about producers hiring on the spot after a Saturday screening. "The anecdote feels outdated," he said, "in an industry where entry-level positions have evaporated."

That evaporation threatens the craft's future. Junior artists traditionally learned fundamentals on the early-stage tasks (rotoscoping, clean-up, in-betweening) that AI now targets first. If those rungs disappear, the pipeline for training the next generation of senior artists weakens. Hayao Miyazaki has warned of "deskilling" and an overreliance on automated imagery. Nearly half of generative AI users surveyed admitted prompting models to mimic specific creators or studios, from Shakespearean prose to Studio Ghibli aesthetics. For concept artists who design worlds before production begins, that mimicry represents lost work and potential erasure of their craft. Disney and Universal have already sued Midjourney over Ghibli-style AI output; Stability AI faces class actions in the U.S.

Artists are adapting in silence. Four VFX workers declined to speak on the record to TheWrap, citing uncertainty about how to discuss the oncoming wave. Others described quietly learning Runway, Midjourney, and Sora, resenting the technology while fearing it will make a VFX living unsustainable. "My strategy is just to adapt, learn everything I can about Sora and Midjourney and all the other kinds of software," one anonymous VFX coordinator said. "Until the studios start announcing their moves, it's the best I can do at this point." The industry remains largely non-unionized, contending simultaneously with production outsourcing, Los Angeles living costs, strike-related financial strain, and fewer greenlit projects from cost-conscious studios.

A middle path is forming. Assistive AI tools now automate repetitive steps (coloring, in-betweening, previsualization) while hybrid workflows combine human-directed CGI with AI-generated imagery. New roles are emerging: AI workflow designers, data ethicists, "prompt animators" who specialize in directing models. Startup Animaj reported its proprietary models cut production time by 85 percent, with human animators refining the final look. Toonstar operates similarly. Netflix has published generative AI guidelines for partners working on originals. The Epigraph Creator Fellowship, backed by Macro and a16z's Cultural Leadership Fund, placed seven filmmakers directly into frontier tools from Promise Studios, OpenAI, ElevenLabs, and others, an explicit bet that creators who master hybrid workflows will define the next era.

Bryn Mooser's Asteria offers a counter-narrative. The artist-led studio has hired more than 50 animators and VFX veterans from Hanna-Barbera and Pixar, pairing them with Moonvalley's "clean" foundational model, Marey. "You're always going to need experienced VFX artists and coordinators working constantly with the software and with the director," Mooser said. "This isn't going to be a future where they fire all the VFX people and the next Star Wars is done by a couple guys in a basement." Promise Studios echoes the approach, actively recruiting talent that blends traditional filmmaking with generative AI creation ability.

The fracture is here. The question is not whether the workforce shrinks. It is whether the survivors are the ones who learn to drive the new tools, or the ones who get driven by them.

AYON: The Open Pipeline

The studio pipeline has always been the nervous system of visual effects, the layer that turns a folder of assets into a deliverable show. For two decades that layer was proprietary, brittle, and expensive to maintain. AYON changed the economics. What began as Pype, the internal toolset of a boutique Czech studio, evolved through OpenPype into a production-grade platform now running on feature films, episodics, and commercials across Windows, Linux, and macOS under an Apache 2.0 license. The company behind it, Ynput, moved the platform into production beta in August 2023 with "near feature parity" to its predecessor and has since added paid Node Graph and Web Publisher add-ons for visual project organization and remote review. These signal that the open-core model sustains development without locking out smaller facilities.

AYON's architecture is built for the hybrid reality studios now face. It integrates natively with the DCC stack (Maya, Houdini, Blender, Nuke, After Effects, DaVinci Resolve, Substance, Unreal Engine) and ties into production trackers ftrack, ShotGrid, and Kitsu. Its OpenUSD resolver and project-level workflows keep layer and variant consistency across departments. The platform handles publishing, versioning, scene inventory, workfile management, look assignment, and multi-site file distribution out of the box. Configuration requires no scripting; supervisors can override naming conventions, plugin settings, and pipeline rules per project. For a facility without a dedicated TD team, that flexibility is the difference between adopting a pipeline and abandoning one.

The Phoenician Scheme proved the model at scale. Vivid VFX delivered 700–800 shots on Wes Anderson's feature with no weekend work and no overtime, a rarity in VFX. The team credited AYON's automated ingest, unified task view, and OCIO color pipeline for saving "hundreds of hours of extra work." Artists stayed inside AYON for tasks, media, and editorial notes without context-switching. One Reddit practitioner called it "the best out-of-the-box pipeline solution that is also free and open-source," while noting the publishing workflow remains convoluted and documentation thin (gaps the API lets studios bypass with custom tooling).

The bridge to AI-native studios is already under construction. At FMX 2026, Ynput demonstrated ComfyUI integration for generative AI workflows, AWS Deadline Cloud for burst render scaling, an overhauled OpenRV review panel, and Marvelous Designer for garment simulation tracked through the pipeline. The new Workflows Add-on (a node-based automation framework included with AYON Studio) targets the handoff friction between departments: ingest, plate prep, AI-assisted matte generation, review, and delivery. Sunrise Animation presented how it modernized its pipeline for larger-scale production while contributing improvements upstream instead of building isolated workarounds.

That pattern (open platform, community extensions, standards-first) mirrors what Promise's MUSE OS attempts from the proprietary side. MUSE embeds Google DeepMind's Gemini and Veo directly into the creative loop: concept, previs, asset generation, shot iteration. AYON provides the versioned, auditable container those outputs need to enter a facility's existing USD/USDZ, EXR, and review pipelines without breaking color space, naming conventions, or vendor handoffs. Studios can run AI generation inside the same pipeline that manages traditional CG, keeping data lineage intact for the digital-replica and copyright regimes the next section examines.

The hybrid stack is no longer theoretical. A facility can run AYON Community for pipeline backbone, license the Node Graph add-on for visual dependency mapping, plug in ComfyUI for diffusion-based concept and matte passes, and route final frames through the same USD review path used for hero assets. MUSE or similar AI-native OSes sit above that layer, orchestrating prompts, model versions, and creative approvals, then writing results into AYON's immutable publish tree. The result is a single source of truth across human and synthetic work. That interoperability, not any single model, is what lets a studio shift capacity between traditional and AI-driven shots without rebuilding its infrastructure every six months.

Digital Replicas and the Law

California moved first. On September 17, 2024, Governor Gavin Newsom signed AB 1836 and AB 2602 into law, creating the nation's most specific guardrails for AI-generated digital replicas of performers. AB 1836, effective January 1, 2025, targets deceased personalities: anyone who produces, distributes, or makes available a "digital replica" — defined as a computer-generated, highly realistic electronic representation readily identifiable as an individual's voice or visual likeness — without consent faces liability of at least $10,000 or actual damages, whichever is greater. The statute reaches back to personalities who died before 1985 and extends protection 70 years post-mortem. AB 2602, also effective January 1, 2025, governs living performers. It voids any contract provision allowing a digital replica to replace work the individual would have performed in person, unless the agreement provides a "reasonably specific" list of all proposed uses and the performer had legal counsel or union representation during negotiation. Only the offending clause dies; the rest of the contract survives.

"Voice and likeness rights, in an age of digital replication, must have strong guardrails around licensing to protect from abuse," SAG-AFTRA executive director Duncan Crabtree-Ireland said in August 2024. The union's 118-day strike — the first since 1960 — secured a tentative agreement through June 2026 that includes consent and compensation requirements for AI-generated video.

The statutes leave room for interpretation. Exemptions cover news, public affairs, sports, comment, criticism, scholarship, satire, parody, documentary or biographical representation, and fleeting or incidental use. But "reasonably specific" has no judicial gloss yet. Fenwick & West warned clients that broad language like "all media known or hereafter devised" will not suffice for digital replicas. Vondran Legal predicts disputes over whether AI generated "background dialogue" counts as replacement versus supplementation, and whether "supplemental, not substitution" arguments survive scrutiny. Courts will likely demand medium, purpose, duration, and scope, not generalities.

California's common law right of publicity runs parallel. The Eastwood v. Superior Court (1983) test requires showing the defendant's use of the plaintiff's identity, appropriation to the defendant's advantage, lack of consent, and resulting injury. Statutory damages under Civil Code 3344 allow the greater of $750 or actual damages plus attributable profits. AB 1836 adds its own $10,000 floor for deceased performers. A two-year statute of limitations runs from the unauthorized use. The interplay between statutory and common law claims remains untested for AI-generated replicas.

Federal action is accelerating. The NO FAKES Act, introduced in Congress, would create a nationwide digital replica right. The U.S. Copyright Office's July 2024 report called for new federal legislation addressing digital replica harms. California layered on more bills days after AB 1836 and AB 2602: SB 942 requires generative AI platforms to embed imperceptible provenance disclosures; SB 926 criminalizes sexually explicit deepfakes of real individuals; SB 981 mandates reporting and removal mechanisms on social platforms. SB 1047 (the Safe and Secure Innovation for Frontier AI Models Act)

Working in AI? Zero G Talent tracks the openings: see every open Databricks role, browse AI jobs, openings at Anthropic, and the people building the field.