Generative AI has already reshaped dubbing, pre-visualization, and de-aging in film and video production — but full synthetic performers and one-click final-pixel VFX are still blocked by consistency, rights, and labor agreements, not just model quality.
In late 2024, Robert Zemeckis's film Here used a real-time AI de-aging system built by Metaphysic to shift Tom Hanks and Robin Wright's on-screen ages by decades, frame by frame, viewable by the director during the actual take rather than added months later in a VFX suite. That's a genuine production milestone: de-aging went from an expensive, offline VFX pass (the kind ILM used on Robert De Niro in The Irishman in 2019, at a cost reportedly north of $1 million per major sequence) to something closer to a live monitor overlay. It's also a useful anchor for a broader question, because most coverage of "AI in Hollywood" swings between breathless (synthetic actors are here) and dismissive (it's all slop). Neither is accurate. Generative AI has landed hard in specific slots of the media production pipeline — pre-visualization, dubbing, temp music, rotoscoping-adjacent cleanup — while the slots involving a real actor's likeness, voice, or final theatrical pixel remain gated by consistency limits, copyright exposure, and, increasingly, union contracts that didn't exist three years ago.
The least controversial and most widely adopted use of generative video sits upstream of anything an audience will ever see: pre-visualization. Directors and DPs have used blocky 3D pre-viz for decades to plan shots before committing a crew and budget to them. Tools like Runway's Gen-3 Alpha and Gen-4, along with Luma's Dream Machine and OpenAI's Sora, let a director generate a rough moving version of a shot from a text or image prompt in minutes instead of days. Lionsgate's 2024 deal to train a custom Runway model on its film catalog was explicitly framed around this: faster storyboarding, shot planning, and marketing concepts, not final-pixel replacement of union crews.
This works because pre-viz has a low bar: it needs to communicate blocking, pacing, and camera movement to a room of collaborators, not survive frame-by-frame scrutiny on a 40-foot screen. The same models that produce distracting artifacts, morphing hands, and temporal flicker in a finished shot are perfectly adequate for an animatic. Adobe has built the same logic into Premiere Pro's Generative Extend, which pads out a clip's in and out points with AI-synthesized frames — useful for smoothing an edit, not for generating new performances. The pattern across this layer: AI compresses the time to a rough draft, and a human still owns the shot that actually ships.
Dubbing is the area where AI has moved furthest into the actual delivered product, because the economics are brutal enough to force adoption. Traditional dubbing requires casting a voice actor per language, booking a studio, and manually re-timing lip movement in post — a process that can cost tens of thousands of dollars per hour of content per language and makes same-day global releases nearly impossible for anything but the biggest titles.
AI-dubbing pipelines from companies like ElevenLabs and Deepdub, and Google's Aloud tool for YouTube creators, collapse that into: transcribe the original audio, translate it, synthesize speech in a cloned or licensed voice, then use a separate lip-sync model to warp the mouth region of the video to match the new phonemes. Meta announced a version of this for Reels and Facebook video at Connect 2024, aimed at creators who currently have no dubbing budget at all. The quality bar here is also more forgiving than people assume — dubbed content has always had an uncanny-valley tolerance built into audience expectations (anyone who's watched a dubbed anime or Bollywood film knows lip-sync is rarely perfect even with human actors), which is part of why AI dubbing reached usable quality faster than AI-generated original performance.
The catch is regulatory, not technical. The EU AI Act's transparency provisions require that AI-generated or manipulated audio/video content that could be mistaken for authentic material be disclosed as such — a rule that lands squarely on synthetic dubbing and voice cloning. Rights holders also have to clear voice-likeness separately from dialogue rights in many jurisdictions, which is why most legitimate dubbing deployments (as opposed to fan projects) start with an actor's consent to have their voice cloned at all, not just a translation license.
The Here example sits at the technically advanced end of an application area that is simultaneously the most legally constrained. AI-based de-aging and digital-double work (used for stunt replacement, posthumous performances, or crowd/background duplication) is real and shipping, but it runs directly into the digital-replica provisions SAG-AFTRA won in its 2023 strike settlement: a studio must get a performer's informed consent and pay them for each specific use of an AI-generated or AI-altered version of their likeness, and can't reuse a scanned performance to generate new dialogue or actions the actor didn't actually perform without separate bargaining.
That's a labor and consent gate on top of a technical one, and it's the main reason "fully synthetic actors" remain a demo, not a business. Corridor Digital's 2023 anime-style short, made with Stable Diffusion img2img over real footage, is instructive here in the other direction: technically it worked, but it triggered an immediate artist backlash over training-data provenance that no dubbing tool has faced, because dubbing augments an existing performance rather than stylistically replacing an artist's visible labor. The lesson generalizes: adoption in media production correlates less with model capability and more with whether the output threatens a specific, organized labor category.
Generative music tools — Suno and Udio being the most prominent — produce full songs, vocals included, from a text prompt, and have found real traction in low-stakes use cases: temp scoring during editing, royalty-free background music for social content, and indie game soundtracks where a licensed composer isn't in the budget. Neither tool, however, has meaningful adoption in film or TV underscore for finished projects, because the major labels sued both companies in June 2024 over training-data copyright, and a studio shipping a theatrical release isn't going to build its score on a copyright claim still in litigation. Production-music libraries (the stock-music catalogs studios already license from) are the more likely near-term entry point, since several have started layering generative tools on top of catalogs they already own the rights to, sidestepping the training-data question entirely.
| Application area | Adoption today | Primary blocker |
|---|---|---|
| Pre-viz / storyboarding | High — standard tool in many pipelines | None significant; low quality bar |
| Editing assists (generative extend, gap-fill) | High | Still needs human review per shot |
| Dubbing / localization | Medium-high, growing fast | Consent + disclosure rules (EU AI Act) |
| De-aging / digital doubles | Medium, high-profile cases | SAG-AFTRA consent & compensation terms |
| Full synthetic performers | Low — demos, not deployed | Consistency, consent, no legal path |
| Music/score for finished releases | Low | Active copyright litigation (Suno, Udio) |
| Temp music / social content scoring | Medium-high | Low stakes, licensing not required |
The pattern in that table isn't random: everywhere adoption is high, the AI output is either invisible to the audience (editing assists) or replacing a step that previously had no budget at all (dubbing for creators who couldn't afford human dubbing, temp music that was always throwaway). Everywhere adoption is low, AI is trying to replace a step that has an existing, organized, compensated human doing it — and that's a negotiation, not a model-quality problem.
If you're evaluating an AI media tool for a real production pipeline, the question that predicts whether it'll actually get used isn't "how good is the output" — most of these models cleared a usable quality bar in 2024. It's "whose paid job does this step currently belong to, and does a consent or compensation framework already exist for replacing it with AI." Pre-viz and dubbing had thin or nonexistent human alternatives in the volume needed, so AI walked in. De-aging and synthetic performance ran into a union that had already anticipated the fight. Score composition ran into a copyright lawsuit. The technology roadmap for all of these areas points the same direction — better consistency, better lip-sync, cheaper inference — but the adoption roadmap is being written in contracts and courtrooms, not in model cards.