---
title: What Happens When AI Makes a Short Film Alone? — Shooting a 5-Minute Film on One MacBook, From Script to Edit
url: https://oosioo.com/en/p/ai%EA%B0%80-%ED%98%BC%EC%9E%90-%EB%8B%A8%ED%8E%B8-%EC%98%81%ED%99%94%EB%A5%BC-%EC%B0%8D%EC%9C%BC%EB%A9%B4
date: 2026-11-10T14:06:12+00:00
author: SYSOP
summary: We gathered where AI has actually been used in film and fiction, then handed a 5-minute short over to AI entirely. Netflix said it used generative AI in about 300 titles this year, but most of that was post-production filling in crowd and background shots that were never in the budget. Two films billed as "Korea's first AI features" each drew audiences in the 600-700 range. In fiction, studies show AI catching up to humans on sentence quality, but readers still price in who wrote it. In our MacBook experiment, the script came out in 54 seconds. Once we told the AI director the equipment's limits, it rewrote all 26 lines of dialogue to play off-screen in a 59-shot shooting script. Faces were cast by rule, voices by description, and the three music tracks took 26 minutes. Instead of the full film, four scenes are shown as GIFs. The script works in one click — the screen doesn't, yet.
---
# What Happens When AI Makes a Short Film Alone? — Shooting a 5-Minute Film on One MacBook, From Script to Edit

![A grid of scenes from the AI-drawn short "Tuesday's Stop": a wide shot of a green city bus parked with its headlights on in a dark bus depot, a woman driver in her fifties wiping the rearview mirror with a cloth in the driver's seat, and more](/uploads/82064e2b933790fd.webp)

Film and fiction are often named as places "AI will never fully break into." And yet Netflix said it used generative AI in roughly 300 titles this year, an "AI feature film" has played in Korean theaters, and a sentence from ChatGPT made it into an Akutagawa Prize-winning novel.

So we did two things. First, we gathered cases of where AI has actually entered film and fiction, and where it stalled. Then we handed a 5-minute short film over to AI entirely — screenwriter, director, casting, cinematography, sound recording, composing, and editing, all of it.

## What we set out to do

1. Gather cases of where AI has actually been used in fields where "human creativity" is supposedly the core — film and fiction. Look at both the numbers companies disclose and the actual audience numbers.
2. Map out, stage by stage in filmmaking (film studies, screenwriting, planning, directing, shooting, editing, sound, distribution), how far AI has come.
3. Try it ourselves. Have AI alone make a 5-minute short, from script to edit. A human only turns the equipment on and doesn't touch anything else. The equipment is a single MacBook (M4 Max, 48 GB memory) and nothing but open models.

## Film: where has it gotten in

### Big studios: filling in "shots that were outside the budget"

At its July 2025 earnings call, Netflix said the building-collapse scene in the Argentine drama "El Eternauta" was the company's first generative-AI-made final footage to air. Co-CEO Ted Sarandos said it was ten times faster than conventional VFX and that the scene simply couldn't have been shot on that budget by the original method. A year later, in July 2026, Netflix said about 300 titles this year used generative AI, mostly in post-production. The U.S. production "The American Experiment" contained 17 minutes of AI-enhanced footage, reportedly "twice as fast and half the cost." Both are company-reported figures. At the same event, Sarandos drew a line: "Great artists make great things, and AI doesn't change that."

Amazon's biblical drama "House of David" made 73 of 850 VFX shots with generative AI in season 1, and used it for more than 350 shots in season 2. Producer Jon Erwin said, "this was the only way we could shoot it on this budget."

What these cases share is that **what to shoot had already been decided.** Crowds, battles, collapsing buildings, wide backgrounds — shots that were in the script but nearly cut for lack of money — were the ones AI filled in.

### Fixing performances already shot

- "Everything Everywhere All at Once" (2022) had five artists handle 80% of its VFX. For the rock scenes, Runway's rotoscoping tool cut background separation that would have taken days down to minutes.
- "Fall" (2022) used Flawless's TrueSync to alter the mouth shapes on roughly 30 profanities, lowering the rating from R to PG-13. Instead of reshoots, it took about two weeks.
- The Swedish film "UFO Sweden" used the same company's "visual dubbing" to create a version where the original actors appear to speak English, and it played in about 100 U.S. theaters in 2025. Reviews noted the seams were "occasionally visible."
- "Here" (2024) showed de-aged versions of Tom Hanks and Robin Wright's faces live on set monitors (Metaphysic Live). The final 4K delivery was further refined by artists in post.
- The audio-separation technology (MAL) Peter Jackson's team built for the documentary "Get Back" pulled John Lennon's voice away from the piano for the Beatles' "Now and Then" (2023).

The work in this category **has an original.** The actor already performed it, and AI transfers or refines that performance.

### The same technology becomes a problem when disclosed late

In January 2025, in the middle of Oscar season, "The Brutalist"'s editor revealed that Respeecher had been used to clean up actors' Hungarian dialogue and consonants. Director Brady Corbet had to clarify that the English dialogue was untouched and that the performances belonged to the actors. Around the same time, it resurfaced that "Emilia Pérez" had used the same company's technology to extend a lead actor's singing range. The controversy centered less on the technology itself than on **when, and to whom, it was disclosed.**

### Films where AI is the lead

From here, things are different.

- shy kids' Sora short "Air Head" (2024) reportedly had a roughly **300-to-1** ratio of clips generated to clips actually used. The protagonist's hair color kept changing between shots, so it had to be repainted in After Effects.
- The 8-minute short "ANCESTRA" (2025), made by Google DeepMind and Darren Aronofsky's Primordial Soup, mixed live-action footage with Veo-generated scenes; Google said the crew numbered over 200.
- The OpenAI-backed animated feature "Critterz," budgeted under $30 million with roughly nine months of production, had aimed for a 2026 Cannes premiere. But OpenAI announced Sora's shutdown in March 2026 (app in April, API in September), and the official Cannes premiere didn't happen. What came after Cannes is unconfirmed.
- In Korea, director Kang Yoon-sung's "Middle Realm" (October 2025, CGV) used AI for 18 creature designs and a city-collapse scene. Post-production that would normally take a year reportedly took about a month and a half. Two films that billed themselves as "Korea's first AI features" in May 2026, "I Am Poppo" and "The Man in Hanbok," closed after about two weeks, with cumulative audiences in the 600-700 range each. The robot sci-fi "Steel Ray," introduced in September, had a budget under 1 billion won, of which **AI token costs alone exceeded 500 million won**, and rising tool prices and shifting licenses reportedly threw off the schedule.

The number of people making these films has grown fast. Entries in the Bucheon International Fantastic Film Festival's AI film category went from 114 in 2024 to 350 in 2025 to 576 in 2026. Runway's AI Film Festival went from about 300 entries in its first year to about 6,000 in 2025. The Korean Film Council opened new AI film production support in 2026 (around 8 features, 30 shorts). Still, audience numbers in the 600s show that **making something easier doesn't mean people come to see it.** "Critterz" and "Steel Ray" show something else: if you build on someone else's tools, your project stalls when those tools disappear or get more expensive.

## Fiction: sentences caught up, but

- Rie Kudan, who won the Akutagawa Prize in January 2024 for "Tokyo-to Dojo-to," said at the award press conference that about 5% of the text was ChatGPT sentences used as-is. The following year, at an ad agency's suggestion, she published a short story that was 95% AI-written in a literary magazine.
- Tsinghua University professor Shen Yang wrote a science fiction story in about three hours through 66 rounds of conversation with AI, and it won second prize at the 2023 Jiangsu Youth Science Fiction Competition. Three of six judges recommended it, not knowing it was AI-written.
- The U.S. SF magazine Clarkesworld closed its submissions window for a time in February 2023 after fielding 500 AI-generated submissions in 20 days. Amazon's Kindle Direct Publishing made AI-use disclosure mandatory that year and capped uploads at three books per person per day.

Studies measuring quality are mixed.

| Study | What was compared | Result |
|---|---|---|
| Pron vs Prompt (2024) | GPT-4 vs. Argentine writer Patricio Pron, each writing short stories from the same 30 titles | 5,400 critic evaluations; **Pron won on every measure.** GPT-4 was more creative when writing from titles Pron supplied |
| Porter & Machery (2024) | AI poems vs. famous poets including Shakespeare | 1,634 general readers correctly identified authorship 46.6% of the time (below chance). They rated AI poems higher, but scores dropped once told "AI wrote this" |
| Chakrabarty et al. (2025) | Mimicking the style of 50 award-winning authors | AI using only prompts fell well short with MFA-trained readers. Fine-tuned on **that author's own body of work**, MFA readers chose the AI version. Cost about $81 per author; AI detectors caught only 3% |
| Villanova University (2026) | ChatGPT short stories vs. human short stories | Readers rated the AI stories higher — but scores were highest when readers believed "a human wrote this" |

In short: sentence by sentence, AI now matches or exceeds humans. But **when a human supplies what to write (the title), AI writes better**, and readers still price in who wrote it.

## How far has AI gotten at each stage of filmmaking

| Stage | What AI does | Example | What humans hold |
|---|---|---|---|
| Film studies research | Measuring shot length, color, composition across thousands of films | Cinemetrics (shot-length database), the Cutting team's analysis of 150 Hollywood films | What to ask |
| Planning/financing | Predicting box office from a script | Warner Bros. × Cinelytic (2020), Largo.ai (claims 86% accuracy) | Whether to make it |
| Screenwriting | Drafts, variations, summaries, coverage | Large language models broadly | What to write and why; credit (WGA: AI output is not an "original work") |
| Previsualization/storyboarding | Scene images, moving storyboards | Runway, Wonder Studio (acquired by Autodesk, auto-replaces actors with CG characters) | Selection |
| Direction/acting | De-aging, mouth replacement, digital doubles | "Here," "Fall," "UFO Sweden" | Original performance and consent |
| Cinematography/lighting | Relighting after the shoot, filling missing shots | Beeble SwitchLight, InterPositive (acquired by Netflix in March 2026, trained on that production's own raw footage) | What goes in frame |
| Editing | Extending cuts, finding dialogue, rough cuts | Premiere Pro Generative Extend (extends video 2 seconds, ambient sound 10 seconds; doesn't extend dialogue) | Rhythm and sequence |
| VFX | Rotoscoping, crowd/background extension | "Everything...," "El Eternauta," "House of David" | Picking 1 out of 300 |
| Sound | Dialogue separation, voice correction, dubbing | "Get Back" MAL, Respeecher | Disclosure |
| Distribution/marketing | Per-viewer thumbnails | Netflix artwork personalization | — |

Reading the table down the column, most of what AI handles is **work where the right answer already exists off-screen.** A scene that had to be shot, a profanity someone wants erased, a voice someone wants isolated — there's a human-set goal, and AI fills it in cheaply and fast.

## Rules are nailing down the human's place

- **Writers**: After a 148-day strike in 2023, the Writers Guild of America (WGA) wrote into its contract that "AI is not a writer, and AI output is not an original work." The 2026 new agreement (ratified 90.38%) kept that clause, but didn't manage to secure payment for scripts used to train AI.
- **Actors**: When the "AI actor" Tilly Norwood became a talking point in September 2025, SAG-AFTRA issued a statement saying, "It's not an actor, it's a character trained on performers' work without consent or compensation." The 2026 new agreement (ratified 91.42%) requires productions using synthetic performers to demonstrate "substantial additional value." In the interim, a feature-film lead role was announced for Tilly Norwood.
- **Awards**: In April 2025 the Academy added a rule stating that "use of generative AI neither helps nor harms a film's eligibility — what matters is whether a human was central to the creative authorship."
- **Copyright**: In January 2025 the U.S. Copyright Office concluded that "prompting alone does not constitute enough control to claim authorship." In March 2026 the U.S. Supreme Court let stand the Thaler ruling that "the author must be human." Korea's Copyright Commission guidance likewise states that only the human's creative contribution gets registered. On the training-data side, Anthropic settled with authors for $1.5 billion over downloading pirated books, with final approval in July 2026 (about $3,000 per work).

What the law and contracts protect is **the part a human selected and arranged.** Which also means that nobody protects the "click" output this series produces, by itself.

## Trying it ourselves: a 5-minute short, AI alone

What the cases above share is that AI filled in a goal a human had set. So what happens if we hand the goal itself to AI too? We set these rules:

- A human only turns on the equipment and runs it in order. No picking or re-generating images or voices.
- Wherever a selection is needed, it's made by a rule set in advance (e.g., among face candidates, pick the one where the face occupies the largest area).
- The equipment is a single MacBook (M4 Max, 48 GB) and downloaded open models. Script and direction were handled by Claude.

| Role | Model used |
|---|---|
| Script/direction (shot list) | Claude |
| Cinematography (one image per shot) | Flux.1-dev (8-bit quantized) |
| Face consistency | PuLID-Flux |
| Camera movement | Parallax effect generated from Depth Anything V2 depth maps |
| Movement without people (rain, wipers, lights) | Wan 2.2 TI2V 5B |
| Voice casting/recording | Qwen3-TTS (voice designed with VoiceDesign, cloned with Base) |
| Music | MusicGen-medium |
| Sound effects | Synthesized in code (filtered noise) |
| Editing | Python + ffmpeg, following the shot list |

### 1. Script: 54 seconds

First we gave it no constraints beyond "a Korean-language short script of about 5 minutes, genre and subject free." In 54 seconds, a 2,637-character script came out. The title: "Tuesday's Stop."

> Oh Jeong-sun, a bus driver in her fifties who has driven the No. 772 last bus for 18 years, has the same elderly man board every Tuesday night. He asks to be let off in front of a textile factory that vanished twenty years ago, saying it's about time his wife gets off from overtime. Jeong-sun asks no questions, stops the bus in front of a construction fence, and quietly texts the man's son ahead of time. Rain falls from a clear sky, and the old man opens an umbrella, tilting it toward the empty seat beside him. That night, Min-jae, a trainee driver on his third day, writes a new stop by hand into the route map.

Five scenes, four characters. It has foreshadowing (the long umbrella under a clear sky), a character left unexplained (Min-jae's understanding is conveyed by having him witness Jeong-sun's text rather than through exposition), and a character who changes by the end (Min-jae, who had been memorizing regulations, writes by hand on the route map). It's a textbook structure for a short script. "Textbook" is both a compliment and a limit. It has a familiar warmth, and no surprising moments.

### 2. Production meeting: once we gave it an equipment list, all the dialogue moved off-screen

Next, we had the same model take on the director's role. This time we told it what this MacBook setup could and couldn't actually do — all limits measured directly in earlier installments of this series.

- Only one face can be locked per shot (PuLID).
- Moving a person with the video model breaks down hands and faces. So character shots only move the camera.
- Lip sync doesn't work. A character on screen can't move their mouth to speak.
- The image model can't render Korean text.
- Up to 12 shots can be animated with Wan (each shot takes over 5 minutes).

400 seconds later, a shooting script and shot list came out: 59 shots, 292.5 seconds (4 minutes 52 seconds). The accompanying "production meeting notes" were interesting.

- **All 26 lines of dialogue moved off-screen.** Instead of the speaker, dialogue was laid over the listener's face, the speaker's hands, a back view, or eyes reflected in the rearview mirror (mouth out of frame). "It's Tuesday today, isn't it?" went over a close-up of eyes in the rearview mirror, directly translating the original stage direction "without taking her eyes off the mirror" into an image.
- **The scene where two people make eye contact disappeared.** The shot of the son bowing to the driver put the driver in the foreground with her back turned, locking only the son's face.
- **Movement was converted into sound.** Opening the umbrella became a still image frozen mid-open. For the final side-mirror shot, since the mirror held two small figures, Wan wasn't used; rain and engine sound stood in for the movement instead.
- **All text is added in editing.** Seven spots — the stop name, a phone text, a construction-site sign, handwriting on the route map, and others.
- **It dropped the line "a sky without a single cloud."** Reasoning that a night sky can't be visually distinguished in an image, it changed it to "a windshield with not a drop of moisture on it."
- After the line "...this used to be the front gate," it inserted a 4.5-second shot of the old man's face with no dialogue, noting: "I judged that receiving this in silence would stay with the viewer longer than the original line."

This is exactly what working directors do under budget and schedule pressure — finding ways to "shoot as if you'd shot" a scene you can't actually shoot. Except here, the result was a film where every line of dialogue is heard off-screen. That's less a directing choice than a form dictated by the equipment.

### 3. Casting: faces by rule, voices by description

The four characters' faces were each drawn three times from the appearance descriptions in the shot list. Even though it specified "head and shoulders only,� the two older characters — the 79-year-old man and the 48-year-old son — were drawn full-body all three times, likely because the costume description mentioned shoes and pants. We added a face-cropping step afterward to compensate. Face widths ranged from 86-96 pixels for them, versus 168-199 pixels for the driver and the trainee driver. Using face recognition (InsightFace), we picked the candidate with the largest, clearest face, then cropped just the face area to use as the reference. The 12 images took 37 minutes (the GPU was shared with music generation running at the same time).

Voices were made with Qwen3-TTS's VoiceDesign model. Feeding in a director's description like "mid-fifties woman, low and calm, slightly husky" produces one sentence read in that voice. Using that recording as a reference, the 26 lines of dialogue were cloned in the same voice. The four reference recordings took 102 seconds; the 26 lines (92.6 seconds of audio) took 381 seconds.

We gave up one thing here. Emotional direction ("small, whispering") only applies at the voice-design stage; the cloning stage ignores direction. If every line were read through the design model, the emotion would come through, but there'd be no guarantee it's the same voice. We judged that a voice changing mid-film would be the bigger failure, and dropped the emotion.

We ran the 26 lines back through a transcription model (Whisper large-v3). 21 lines matched exactly (excluding punctuation and spacing). Of the remaining five, three were transcription-style discrepancies: "twenty years" was written as "20 years," "this car" as "car 2," and one word was mistranscribed similarly. "Starting next week, you pull over for him" was transcribed with the subject swapped to "I." In contemporary Korean pronunciation, the subject markers for "you" and "I" are nearly indistinguishable. But this is the film's final line, and swapping the subject reverses its meaning — a human actor would have read it with the alternate pronunciation to disambiguate. The one line that sounds simply wrong by ear is a single mistranscribed word.

### 4. Music: 26 minutes

Three tracks (46 seconds for the depot, 60 seconds for the rainy scene by the fence, 38 seconds for the ending) were made with MusicGen. Since it only outputs up to 30 seconds at a time, the last 10 seconds had to be regenerated and spliced on. The three tracks took 1,563 seconds. The first track suddenly jumped 18 dB in volume at the 50-second mark — the spliced-on portion started as if it were a different song. Since this track is cut at 46 seconds for use, it doesn't make it into the film.

### 5. Cinematography: a bit over 2 minutes per shot

All 59 shots are drawn one at a time as 1344×576 (2.33:1) images. The 24 character shots lock the face with PuLID. Each takes about 2 to 2.5 minutes, so drawing all of them takes a bit over 2 hours. After that, the 12 Wan shots take about an hour, with another 30 minutes or so for the parallax effect and editing.

![A grid of AI-drawn shots in sequence: a wide shot of the depot, the driver in the driver's seat, a hand holding a notebook, an exterior shot of the bus, and more](/uploads/5ef5b002d6723992.webp)

Looking at the early shots, the texture feels close to film. The depot under fluorescent light, a memo taped to the bus dashboard, the green bus at a foggy stop — these come out looking like a lighting director paid attention.

The problem is performance. All three of the face-locked Jeong-sun shots show her **facing the camera directly and smiling.** The instructions were, respectively, "wiping the rearview mirror while looking at it," "only the eyes reflected in the rearview mirror, mouth out of frame," and "profile from the driver's seat, mouth not visible." The "eyes in the mirror" shot in particular was exactly where the director had pushed the mouth out of frame to avoid lip sync. But the image that came out showed a full face, mouth and all, looking straight at camera. Playing "It's Tuesday today, isn't it?" over this shot makes it look like someone with a closed mouth is talking. The problem the production meeting had avoided came right back at the shooting stage. Face-locking (PuLID) pulls the reference face toward a passport-photo-style frontal pose, and the image model learned portraits as "photos looking at the camera." That habit is stronger than the directing instruction.

The instruction not to include text was also broken, by a sign visible past the bus's front door — a broken pattern that looked like letters appeared. And the empty bus parked at the depot had passengers sitting in it.

With a human crew on set, this is where you'd reshoot. Under this experiment's rules, nothing gets regenerated. So these shots go into the finished film as-is.

![Bar chart of time spent per stage: script 54 seconds, production meeting 400 seconds, 12 face candidates 37 minutes, voice 8 minutes, music 26 minutes, 59 shot images about 2 hours 20 minutes (estimated), 12 Wan shots about 1 hour (estimated)](/uploads/0c5245d429252f67.webp)

**Rather than embedding the full 5-minute film, here are four scenes as GIFs.** Being GIFs, they have no sound; the dialogue is overlaid as subtitles at the same points as in the film.

![Wan 2.2 animates the AI-drawn bus depot into a 3-second clip: a green bus slowly approaches under fluorescent lights](/uploads/1813aa3fc0d0ef2b.gif)

This is the first shot. The 12 shots without people were animated with Wan 2.2. Each 3-4 second shot took 3-5 minutes (185-299 seconds). The approaching bus and drifting fog look natural.

![The view through a bus windshield on a rainy night. The wipers never move, the bus and the car ahead that should be parked drive forward, and an orange smear spreads on the right](/uploads/427d3ad0fcfc118e.gif)

Wan was not always right. The instruction for this shot was "static camera, rain runs down, the wipers start sweeping left and right." The wipers never move. Instead the bus, which should be standing still, moves forward, and the parked car ahead drives off. An orange smear spreads along the right edge. The model read the rainy road as "the view from a moving car." Even in a shot with no people, the model decides what moves.

![Under the subtitle 'It's Tuesday today, isn't it?', the driver faces the camera with her mouth closed, in a shot meant to show only her eyes in the mirror](/uploads/484814e1b5fab9b1.gif)

This is the shot described above. The director asked for the mouth to be out of frame, but the driver faces the camera, mouth closed, smiling, while "It's Tuesday today, isn't it?" plays over it. Shots with people were moved only by a parallax effect that slowly pushes the camera in. Nobody moves, so nothing collapses, but there is no performance either.

![The last shot: the view in the side mirror recedes in the rain as the subtitles 'From next week, you stop for him' and '…Okay' pass](/uploads/1d8732978b460811.gif)

This is the last shot. Because of the two tiny people in the mirror, the director chose not to animate it with Wan and left the motion to the sound of rain and the engine. As a silent GIF it is close to a still photo. It shows how much work the sound was doing in this film.

## Where it went wrong

- **One installation got blocked.** We first intended to use ACE-Step 1.5 as the music model. But the step of running an external repository's install script tripped this environment's security rules. Rather than work around it, we switched to MusicGen, which could run on an already-installed library (transformers). MusicGen's weights are under a non-commercial license (CC BY-NC 4.0).
- **The face-candidate prompt wasn't followed.** "Head and shoulders only" turned into full-body for the two older characters, likely because of the shoe and pants details in the costume description. We added the face-cropping step belatedly to fix this.
- **The music splice caused an audio jump** (item 4 above).
- **We were too optimistic on timing at first.** We trusted an old note saying a single image takes "tens of seconds to 2 minutes," but adding face-locking and wide frames pushed it past 2 minutes. That's why the finished film is coming out later than this article's publication.

## So can it be done in one click?

**The script clicks into place.** The script that came out in 54 seconds had a solid structure with little to fix. The production meeting works too. Once told the equipment's limits, it changed scenes on its own and kept minutes explaining why. A human director doing the same work would take a full day.

**The screen doesn't click into place yet.** In this experiment, the single biggest decision wasn't made by the director — it was made by the equipment. Because lip sync doesn't work, all the dialogue moved off-screen; because people can't be animated, only the camera moved. As the 300-to-1 ratio in "Air Head" suggests, in professional production a human selects and fixes what falls in that gap. This experiment removed that human, so what gets lost is laid bare.

Overlaying this against the earlier cases, one line becomes visible. AI is faster than humans at **filling in what's already been decided** (script format, shot lists, backgrounds, sound). **Choosing what to show** — especially the moments where a human face and body perform — is still a human's job. And that is exactly where law and contract are nailing down the human's place.

## What it can't do

- **The full 5 minutes is not included.** Only four scenes are shown, as silent GIFs. Rendering had not finished at first publication, so the GIFs were added the next day.
- **No audience evaluation.** We didn't run an experiment showing a human-made film from the same script side by side and asking viewers to choose.
- **We only ran it once.** Asking for the same script again would produce a different story. We don't know whether this script was average or just lucky.
- **No commercial tools were used.** Paid video models like Veo, Kling, and Runway handle human movement and lip sync better than the open models used here. This result is specific to "a single MacBook and open models."
- **Licensing**: Flux.1-dev and MusicGen's weights are under non-commercial licenses, so using this film commercially would require swapping models. All characters in the film are fictional.

### If you enjoyed reading this, please leave your thoughts.

## Sources

- Netflix: Vice (2025-07-16, "El Eternauta" 10x), IBC (2026-07, ~300 titles/17 minutes in 2026), TheWrap (2026-03-05, InterPositive acquisition)
- "House of David": WIRED reporting (2025)
- "Everything Everywhere All at Once": fxguide interview
- "Fall": AV Club/PetaPixel (2022-08) / "UFO Sweden" ("Watch the Skies"): Paste (2025-05-09)
- "Here": The Hollywood Reporter (2024-10) / "The Brutalist": IndieWire (2025-01)
- "Air Head": fxguide interview (summarized via notebookcheck, 2024-04) / "ANCESTRA": Science & Film (2025-06)
- "Critterz": AV Club (2025-09), reporting via Bloomberg (2026-05-22) / Sora shutdown: MacRumors (2026-03-24)
- "Middle Realm": Herald Business, Kyunghyang Shinmun (2025-10) / Korean AI feature audience figures: Tech42 (2026-08-18) / "Steel Ray": Money Today (2026-09-08)
- BIFAN AI category: Hankyung, Hangyo, Daily Ian (2024-2026) / Runway AI Film Festival: aiff.runwayml.com, AP (2025-06) / Korean Film Council AI production support: Herald Business (2026-02)
- Rie Kudan: Smithsonian (2024-01) / Shen Yang: Locus (2024-01) / Clarkesworld: NPR (2023-02-24) / KDP: Authors Guild (2023-09-18)
- Studies: Pron vs Prompt [arXiv 2407.01119](https://arxiv.org/abs/2407.01119), Porter & Machery [Scientific Reports (2024)](https://pmc.ncbi.nlm.nih.gov/articles/PMC11564748), Chakrabarty et al. [arXiv 2510.13939](https://arxiv.org/abs/2510.13939), Villanova University (reported in Judgment and Decision Making, 2026-08)
- Film studies: Cutting, DeLong & Nothelfer (2010, Psychological Science), Cinemetrics / Cinelytic: The Hollywood Reporter (2020-01) / Wonder Dynamics acquisition: CG Channel (2024-05) / Generative Extend: TV Technology (2025-04)
- Rules: WGA/SAG-AFTRA 2026 ratification (Screen Daily), SAG-AFTRA statement on Tilly Norwood (Variety, 2025-09-30), Academy rule (PetaPixel, 2025-04-23), U.S. Copyright Office AI Report Part 2 (2025-01-29), Thaler appeal denied (Mayer Brown, 2026-03), Korea Copyright Commission "Guide to Copyright Registration of Works Using Generative AI" (2025-06), Bartz v. Anthropic final approval (JURIST, 2026-07)
- Experiment models: [Flux.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev), [PuLID-Flux](https://github.com/ToTheBeginning/PuLID), [Wan 2.2](https://github.com/Wan-Video/Wan2.2), [Depth Anything V2](https://huggingface.co/depth-anything/Depth-Anything-V2-Small-hf), [Qwen3-TTS](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign), [MusicGen](https://huggingface.co/facebook/musicgen-medium)
- The script, shot list, images, voices, music, and charts were all produced for this article's experiment
