What AI creation lacks isn't generation quality — it's the film industry's script supervisor. Nobody owns continuity. AI is the strongest crew you'll ever hire, but it is not the director. Models are rented lenses; your character assets and your process are the negatives — the only things that are truly yours.
What Character Drift Really Is: A Crew With No Script Supervisor
Q: You came from the film industry. What's your first reaction to today's AI-generated shorts?
A: The single-shot quality startles me, honestly — the lighting in a lot of these shots would have taken a whole team a full day, ten years ago. But watch three shots in a row and the trouble starts: the lead's face has changed, the buttons on the jacket have moved, the window has migrated from the left wall to the right.
The industry has a name for this: continuity errors. On a traditional set there's a role whose whole job is to catch them — the script supervisor. She logs every take: the camera position, the lighting setup, where the actor's hands were, how many cigarette butts are in the ashtray. The audience never knows the job exists. But without her, the cut you get is full of drifting faces.
AI creation right now is a crew of staggering talent with no script supervisor. Every shot is imagined from scratch; nobody remembers what happened in the last one. So when I see "character consistency" debated as a model problem, it always feels aimed at the wrong target — this is a process problem first. Models will keep getting stronger, generation after generation, but as long as nobody owns continuity, the strongest model in the world will still drift.
Q: So MajoFlow's asset workshop and generation records — you're essentially building a script supervisor?
A: Yes — that's the most precise way to put it. A character pack is a character-sheet archive. The shot library is the continuity log. Generation records are "this take used this camera, this lighting." On a film set, these things are common sense refined over a century. Ported into AI creation, they somehow count as differentiating features — which tells you how young this industry still is.
AI Is the Strongest Crew, Not the Director
Q: "Don't let AI create for you" — plenty of people find that line contradictory coming from an AI tool.
A: Not contradictory at all — the line comes straight from the set. A director doesn't carry the camera, doesn't rig the lights, doesn't draw the storyboards himself. Every one of those jobs, someone does better than he could. But what each shot needs to be — that's his call, and his alone. When have you ever seen a cinematographer decide the emotional temperature of a scene for the director?
What AI hands every creator is a crew that only big-name directors used to be able to afford: top-tier cinematography, top-tier production design, top-tier sound, on call around the clock, never once complaining about overtime. That's a remarkable thing. But no matter how strong the crew is, it doesn't know what you're trying to shoot. The moment you hand the "what" over to AI, you stop being the creator. You become the audience — just the first one in the room.
So every design decision in MajoFlow defends that boundary: models show up as creative actions — generate a character, generate a shot, plan a storyboard — and initiating the action, keeping or killing the result, stays in human hands. We don't build a "one-click complete short film" feature. Not because we can't. Because it hands the director's chair over to gacha pulls.
You Test-Shoot a Rented Lens. Why Would a Model Get a Pass?
Q: MajoFlow ships models slower than many competitors. Is that deliberate?
A: Deliberate. There's a rule on set: untested gear doesn't go into principal photography. You rent a lens, you shoot a test roll first and check the edges — and that's just a lens, not the generative model that decides what your entire film looks like.
Before any model ships, we co-test it with experienced creators on real tasks. Not benchmarks — real work, under pressure: can the same character's character sheets reliably hold up in a profile shot; does packing two actions into ten seconds fall apart; if you change one word in the prompt, does it actually listen. Three metrics: consistency, instruction following, controllability. The findings get written up as that model's "dialect file" — what sentence shapes it eats, where its word-count sweet spot sits, how to handle negations. (We wrote the whole method up separately: How We Decide to Ship a Model.)
The fashionable move is to list a model the day it launches. I understand the logic — "day-one integration" looks great in marketing. But when a creator runs an untested model on a real project, the cost of the crash is theirs. That's math we do on our users' behalf.
Process Isn't Bureaucracy — It's a Promise to Shot Thirty
Q: The film industry's process is famously heavy. Do independent creators really need it?
A: What they need isn't the bureaucracy. It's the promise underneath it: shot thirty is as good as shot one. Strip it down, and the core asset the film industry has banked over a hundred years is continuity management — how a few hundred people, over months, across thousands of shots, produce one coherent thing.
An independent creator doesn't need a production manager and three tiers of sign-off. But they do need a lightweight version of the same machinery: character sheets standing in for wardrobe and makeup tests, shot cards standing in for the shot list, version records standing in for the slate. A crew of one is still a crew. The orchestration canvas is that machinery, compressed to a weight one person can carry.
Q: And to creators just walking in the door — what would you tell them?
A: Don't wait for "the stronger model." That wait has no end — there's a new generation every six months. Put your effort into the things a model swap can't take from you: your character assets, your process, your judgment about shots. Models are rented lenses — this one today, that one tomorrow. Your assets are the negatives. Hold the negatives, and you can print your film on any machine.
About This Interview
Was MajoFlow built by a team with a film industry background?
The founder has more than a decade in the film industry. The product's asset workshop, orchestration canvas, shot cards, and review flow all come directly from the continuity-management methods of film production.
Is "don't let AI create for you" a stance against AI generation?
No. What it opposes is outsourcing creative decisions to re-rolling. MajoFlow uses generative models heavily — but insists they serve human judgment in the form of creative actions. AI is the crew; you are the director.
What is the model co-testing mentioned in the interview?
Before any model ships in MajoFlow, it's co-tested with experienced creators on real creative tasks for consistency, instruction following, and controllability. The findings become that model's prompt dialect file. See How We Decide to Ship a Model.