Orelon logoOrelon
料金

Curating Cinematic AI Video Projects: Build a Story Vault

2026年11月17日 · Orelon Team 著

AI動画テンプレートを見る

着想のためにコミュニティ作品をいくつか閲覧し、任意のテンプレートを開いて Orelon で作成を続けましょう。

A practical system for curating cinematic AI video projects: story spines, model selection, look development, consistency, naming, and edit assembly.

A Storyteller's Vault is not a folder full of clips. It is the system that lets a cinematic idea survive contact with production: the treatment, the look bible, the reference frames, the prompt history, the shot ledger, and the edit decisions that turned a vague ambition into something watchable. Most AI video projects fail for a boring reason — nothing was curated. Every generation was a fresh gamble, and every new take quietly contradicted the last one.

Curation is the difference between generating footage and directing a film. This guide lays out a working method: how to structure a vault of cinematic AI video projects, how to pick between models shot by shot, how to lock a visual identity before you burn hours on generation, and how to keep a project legible when it grows from three clips to sixty.

What a Storyteller's Vault Really Is

Strip away the romance and a vault is four things stacked together.

  • Intent — the treatment, the logline, the emotional target of each scene. This is what you check against when a beautiful shot is wrong for the film.
  • Look — reference frames, color scripts, lens choices, grain and contrast notes. The visual contract every shot must honor.
  • Assets — character sheets, location plates, props, reusable style prompts, and the seed or reference images that produced them.
  • Provenance — which model, which prompt, which settings, which version produced each approved shot, so a rewrite six weeks later is not archaeology.

A vault without provenance becomes a shoebox. You find a shot you love, you cannot reproduce it, and the project stalls while you reverse-engineer your own work.

The vault is also a decision filter. When you have twenty candidate takes for a two-second insert, the vault tells you which one belongs: the one that matches the look bible and serves the intent. Without that filter, taste becomes mood, and mood changes with every scroll.

Start With the Story Spine, Not the Model

The most common failure mode in AI filmmaking is starting in the generator. A prompt feels exciting, a clip looks expensive, and suddenly you are building a story around whatever the model happened to render well that afternoon.

Reverse it. Before a single generation, write three artifacts.

The one-page treatment

Six paragraphs, no more: who wants what, what blocks them, what it costs them, how it ends, and what the audience should feel in the final ten seconds. If you cannot fill the emotional target line, you are not ready to generate.

The shot ledger

A plain table with one row per shot: shot ID, scene, duration, story purpose, subject, camera behavior, location, and status. Twenty to forty rows is typical for a two-to-three-minute piece. The ledger is the project's spine — everything else hangs off it, and it prevents the classic AI short film problem of twelve gorgeous establishing shots and no coverage.

The constraint list

Decide up front what must stay stable: a face, a coat, a kitchen, a color temperature, a time of day. Each constraint becomes a reference asset you build before production rather than a problem you solve on shot thirty.

Writing these three artifacts takes an afternoon. It saves weeks.

Build a Model Selection Matrix That Survives Change

No single video model is best at everything, and the gap between them is largest exactly where cinematic work lives: faces in motion, hands, camera moves that respect physics, text, crowds, and the specific texture of light. Rather than crowning a winner, build a matrix that maps shot types to the tool most likely to deliver them.

Shot type Priority What to evaluate
Dialogue close-up Facial stability Micro-expression, lip sync, identity drift over 4+ seconds
Establishing wide Composition control Aspect ratio options, horizon logic, atmospheric depth
Action insert Motion fidelity Object permanence, motion blur, frame-to-frame coherence
Product macro Texture realism Material response, reflections, fine detail retention
Stylized montage Aesthetic range How far the model bends toward illustration without breaking

Fill the matrix by running the same five-second test prompt through each candidate: one face, one camera move, one texture. Score them 1–5 on identity hold, motion smoothness, and prompt adherence. Keep the test clips in the vault under tests/ so future-you can see why the choice was made.

Two rules keep the matrix healthy. First, evaluate on your own material, not on showcase reels — demo clips are curated survivors. Second, re-run the test every few months, because model behavior shifts and yesterday's weakness may be fixed while a new one appears.

When you settle on a primary tool, use an AI video generator that gives you enough control over framing, motion strength, and reference conditioning to reproduce a look deliberately rather than by luck.

Look Development: Lock the Style Before You Generate

Look development is where amateur projects and cinematic ones diverge. The amateur generates a hundred clips and hopes for cohesion. The professional builds a style bible, proves it on three shots, then scales.

The style bible

One page, written in plain language, containing:

  • Palette — three to five named colors with hex values, plus the rule for when the accent appears.
  • Lighting — hard or soft, motivated source, contrast ratio, whether shadows are crushed or open.
  • Lens language — the focal lengths you allow, and what each means emotionally (35mm for proximity, 85mm for isolation).
  • Texture — grain, halation, chromatic aberration, and the exact amount, described in words you can paste into a prompt.
  • Negative space — how much empty frame you permit, since AI models love to fill it.

Reference frames before video

Generate or select six stills that represent the film's visual range: a wide, a close-up, an interior night, an exterior day, a stylized moment, and a texture detail. These are your anchors. If you need help producing them, an AI image generator is faster and cheaper for look exploration than video, and the approved stills double as reference conditioning later.

Approve the stills before you approve a single video. Changing a palette after forty clips exist means regenerating forty clips.

Consistency Across Characters, Locations, and Props

Identity drift is the quiet killer of AI narrative. A face that shifts between shots reads as a continuity error, and audiences forgive bad effects far more readily than they forgive a protagonist who changes bone structure mid-scene.

Build a character sheet, not a character description

Create front, three-quarter, and profile stills of each recurring character under consistent lighting. Add two expression variants — neutral and one extreme. Store the approved sheet with a short canonical description: age range, build, hair, wardrobe, and two immutable details (a scar, a specific jacket, a hairline). Those immutable details are your continuity handles.

Treat locations like sets

A location is a lighting setup, not just a place. Shoot a wide plate, a medium, and a detail of each recurring space, and record the time of day and weather attached to each. When you return to the kitchen in scene nine, you are matching a plate, not imagining one.

Props and text

Props that appear twice need a reference image. On-screen text — signage, handwriting, labels — should be added in post rather than generated, because generative text remains the least reliable element in most pipelines.

A useful habit: every approved asset gets a two-line note explaining why it was approved. That note is what you show a collaborator when they ask why the second take was rejected.

A Repeatable Shot-by-Shot Workflow

Consistency comes from repetition, not inspiration. Here is a loop that scales from a single shot to a full sequence.

  1. Pull the row from the ledger. Confirm the story purpose in one sentence. If you cannot state it, cut the shot.
  2. Assemble the prompt from modules. Subject + action + camera + lens + light + palette + texture + negative constraints. Keep each module short and reusable.
  3. Attach references. Character sheet, location plate, or style still — no more than the model can weight meaningfully. Too many references blur into mush.
  4. Generate low-cost exploration. Short durations, low resolution, several variations. Judge composition and motion, not detail.
  5. Lock one direction. Post the winning frame next to the style bible and check it honestly.
  6. Generate the hero take at final duration and resolution, with the locked prompt and seed.
  7. Log it. Prompt, references, settings, model version, and a one-line note on what changed.

Where takes go wrong

Three failure patterns account for most reshoots. Overloaded prompts fight themselves — if the subject turns, the camera pushes, and the light changes all at once, the model compromises on all three. Reference conflicts happen when the character sheet and the style still disagree about lighting. Motion greed is asking for a large move in a short clip; generous camera travel needs duration to read as intentional.

Before generating, ask whether the shot would work as a still. If the composition is dull in a freeze frame, motion will not save it.

Organizing the Vault: Naming, Metadata, and Versioning

A vault is only as useful as it is findable. Adopt a naming convention on day one and never negotiate with yourself about it.

project_scene-shot_take-version_status — for example nightferry_s03-02_t04-v2_approved. The status tag (explore, hold, approved, killed) does more work than any folder hierarchy, because it lets you filter by decision rather than by date.

Keep four top-level folders: treatment/, assets/ (character sheets, plates, style stills), shots/ (generated footage), and assembly/ (edits, sound, exports). Store prompt text as plain files next to the footage; a prompt trapped in a chat window is a prompt you will lose.

Three organization mistakes that cost the most

  • Saving only finals. The rejected take is often the only record of what the prompt actually did. Keep one representative failure per shot.
  • Versioning by filename anarchy. final_final_v3 is not a versioning system. Use two-digit versions and let status carry the meaning.
  • Skipping the ledger update. If the ledger and the folder disagree, the folder wins and the story loses.

A prompt library of proven modules — lighting phrases, lens phrases, texture phrases — turns each new shot into assembly rather than invention. Reusable video templates help the same way for recurring formats like title cards or product beats.

Sound, Edit, and the Assembly Pass

Cinematic is a sound claim as much as a visual one. Silent AI footage reads as a demo; the same footage with room tone, foley, and a score reads as a film.

Work in three passes. The assembly pass cuts for story only, using placeholder audio, and ruthlessly deletes any shot that does not advance the scene regardless of how expensive it looked. The rhythm pass adjusts cut lengths so tension builds — trim two frames off each cut in a chase, let a wide breathe for a full second before a reveal. The sound pass adds ambience first, then foley, then music, then any dialogue treatment.

Two technical habits pay off: generate or source a few seconds of clean room tone for every location so cuts do not pop, and match grain and contrast across clips before adding a global grade. Mixed grain is the fastest way to make a coherent film look stitched together.

Export a review cut with burned-in shot IDs. When you watch it, note the shot ID of every moment that pulls you out — those notes become the next ledger update, not a vague memory of dissatisfaction.

FAQ

How many AI video projects should a vault hold at once?

One in production, one in look development, and an archive of finished work. Running three productions in parallel usually means three sets of half-locked style rules, which is how consistency dies.

Do I need several different video models?

Not necessarily, but you need to know your primary model's weak shot types so you can avoid designing around them or bring in a second tool deliberately for those shots. The matrix matters more than the count.

How do I keep a character's face stable across shots?

Lock a character sheet with three angles and two expressions, write two immutable physical details into every prompt, and keep lighting consistent between the sheet and the scene. Drift usually comes from lighting mismatch, not from the model forgetting a face.

How long should individual AI-generated shots be?

Short. Three to six seconds covers most narrative beats and keeps identity stable. Build longer sequences from multiple cuts rather than one long generation.

What is the right way to start a project?

The treatment, the shot ledger, and the constraint list — in that order, before generating anything. If the emotional target of the ending is undefined, no amount of generation will find it.

Should I generate stills first?

For anything stylized or character-driven, yes. Stills are cheap, fast to compare, and become the reference conditioning that keeps video consistent. Approve the look as images, then spend on motion.

How do I handle dialogue?

Generate the performance as a visual plate with clear mouth movement, then treat audio separately. Trying to nail lip sync and cinematic framing in one generation usually compromises both.

What belongs in a project post-mortem?

Three lines: what the vault made easy, what it made hard, and which prompt module you will reuse. Post-mortems are how a vault compounds instead of resetting with every project.

Build Your Vault in Orelon

Orelon is built for cinematic ideas in motion — the kind that need a locked look, a consistent character, and a shot ledger rather than a lucky prompt. Start with your treatment, prove your style on a handful of stills, then move into the generator with references and prompt modules you already trust. Browse the Orelon blog for workflow breakdowns, or compare approaches in the alternatives hub if you are still assembling your matrix. Curate first. Generate second. The vault is what makes the film.