Why Story Structure Still Decides Whether AI Video Works
Every few years a new tool promises to remove the hardest part of filmmaking. It never does. What changes is where the difficulty moves. With generative video, the difficulty shifts from can we afford to shoot this to have we decided precisely what this shot is for. That is a structural question, not a technical one, and it is still answered with a script, a beat sheet, and a shot list.
For independent filmmakers working with small crews, short schedules, and modest budgets, the temptation is to start generating immediately. A handful of striking clips feel like progress. Then the edit begins and the clips refuse to sit next to each other: eyelines do not match, the light changes direction between shots, a character wears a different shade of blue in every scene. The footage is beautiful and unusable.
The fix is unglamorous. Lock the story first. Design shots that are both narratively necessary and technically achievable. Then choose the generation method that matches the intent of each individual shot, rather than treating every shot as the same problem. Directors who work this way spend less time re-rolling prompts and more time solving the actual creative problem: what does the audience need to feel, and in what order.
This guide is written for filmmakers who already think in scenes and coverage, but who are new to producing images with generative tools. It covers story structure as it applies to machine-assisted production, shot design vocabulary that translates into prompts, continuity strategy, pacing, tool selection, an end-to-end workflow, and the mistakes that derail most first projects.
The Story Spine: Beats That Translate Cleanly Into Prompts
A beat sheet is useful for any film, but it becomes essential when part of your image-making is delegated to a model. The reason is simple: models cannot render subtext. They render surfaces. A beat that lives only in a character's inner life gives you nothing to describe, and a shot you cannot describe is a shot you cannot generate.
The practical move is to externalize every beat into visible behaviour. Take each line of your outline and ask: if the sound were off, what would the audience see?
| Internal beat | Weak, unrenderable version | Strong, renderable version |
|---|---|---|
| Betrayal lands | She feels betrayed | Close-up: her hand tightens on a coffee cup, liquid ripples at the rim |
| A decision is made | He decides to leave | Wide: he zips a duffel bag in a stripped room, cold morning light through blinds |
| Reconciliation | They forgive each other | Two-shot on a bridge: one figure extends a hand, the other hesitates, then takes it |
| Grief | She cannot let go | Insert: a hand smoothing a photograph on a kitchen table, thumb pausing on a face |
| Threat | Danger is approaching | Static wide of an empty street, a single car headlight entering frame right |
Notice how each strong version contains a subject, an action, a place, and a light condition. That is the minimum information any generation system needs, and it is also the minimum information a camera crew needs. The constraint improves the script.
For shorter work, a five or six beat spine is usually enough: ordinary world, disruption, escalation, reversal, resolution, and an aftertaste. Longer pieces can follow a fuller structure, but keep the same discipline. Every beat should have at least one visual anchor that can be filmed or generated without dialogue.
From Script to Shot List
A shot list is the bridge between story and production. In AI-assisted work it does double duty, because it also becomes your generation brief.
Describing shots in a form a model can use
Use a consistent six-part description for every shot. It reads naturally to a human and works as a prompt structure:
- Subject — who or what is in frame, including wardrobe and age cues.
- Action — one clear, single motion or state, not a sequence of events.
- Environment — location, weather, background elements that matter.
- Camera — framing, angle, lens feel, movement.
- Light — direction, quality, time of day.
- Look — palette, contrast, film texture, aspect ratio.
Example: A woman in her thirties in a grey wool coat, standing still on a rainy platform, medium shot at eye level with slight handheld drift, overcast daylight from the left, desaturated teal and concrete palette, 2.39:1, subtle grain.
That single line is immediately producible. Multiply it by sixty and you have a film.
Deciding what to generate versus what to shoot
Not every shot should come out of a model. Build a simple triage rule:
- Generate the shots that would be expensive, dangerous, or impossible practically: period establishing shots, weather, crowds, distant landscapes, abstract sequences.
- Shoot the shots that depend on performance, precise eyelines, or dialogue: close-ups of actors speaking, reaction beats, anything where micro-expression carries meaning.
- Reuse archival or stock material for texture: maps, screens, television inserts, photographs.
A hybrid project is almost always stronger than an all-generated one, and it is far less likely to fall apart in the edit.
Continuity Across Shots: Characters, Locations, Props
Continuity is the single biggest technical risk in generative production. A model may render a face differently in every clip, a room may grow or shrink, and a jacket may change colour between cuts. No tool currently guarantees perfect consistency, so the professional approach is to design around the limitation instead of fighting it.
Build reference bibles before you generate anything
Create four documents and keep them open while you work:
- Character bible — one approved reference image per character, plus a fixed written description of face, hair, build, wardrobe, and accessories.
- Location bible — one approved image per location, plus notes on layout, window direction, and recurring objects.
- Prop bible — the objects that carry plot weight, with their exact colour, material, and wear.
- Palette sheet — a small set of colours used across the whole film, per storyline or reality layer.
Then reuse the same reference images and the same wording every time those elements appear. Variation in your own language is variation in the output.
Structural workarounds that hide low consistency
Directors have always used tricks to avoid revealing that two shots were captured far apart. The same tricks work here:
- Shoot characters from behind, in silhouette, in reflection, or as hands and partial faces.
- Keep characters separate in frame when a scene involves many people.
- Prefer wide shots and inserts over sustained close-ups of unfamiliar faces.
- Hold one lighting direction for a whole scene, so matching is easier.
- Use environmental storytelling — spaces, objects, weather — to carry emotion where faces cannot.
If a sequence needs a repeatable face across many shots, plan to shoot that sequence with a real actor and use generated material for cutaways. That is not a compromise; it is correct shot planning.
Camera Language, Light, and Colour as Storytelling Tools
Generative tools respond best to a small, well-defined vocabulary. Decide on your camera grammar early so the film feels intentional rather than assembled from whatever the software happened to produce.
Framing and lens feel
- Wide (24–28mm feel) — establishes space, exaggerates distances, useful for isolation and geography.
- Normal (40–50mm) — neutral, observational, good for documentary texture.
- Telephoto (85mm and up) — compresses space, flatters faces, isolates a subject from a crowd.
- Over-the-shoulder — implies a relationship or a point of view.
- Insert — a detail that carries plot information: a phone screen, a key, a stain.
Movement
Static frames read as calm or observational. Slow pushes read as dawning realisation. Handheld reads as immediacy or unease. Crane and drone moves read as scale. Whip pans read as urgency. Pick two or three movements for the whole project and use them consistently — a film with a coherent grammar feels far more expensive than a film with maximum variety.
Light as continuity
Within a single scene, keep the key light direction identical in every shot. If light comes from a window on the left in the wide, it must come from the left in the close-up. This one rule prevents most continuity failures, because mismatched light is what audiences notice first, even when they cannot name it.
Use a small number of lighting states across the film: overcast daylight, hard midday sun, warm practical interior at night, cold fluorescent institutional. Then every scene is a variation on four states rather than a new problem each time.
Colour and finishing
Assign each storyline, timeline, or emotional register a palette, and keep saturation and contrast consistent per register. A useful habit is to generate slightly flat, slightly neutral images and apply the final look in the edit with a grade or LUT. Baking a heavy look into every generated clip makes later matching nearly impossible and locks you out of fixing a scene that does not cut well.
Pacing and Shot Duration
Generative clips usually have a narrow usable window. The beginning may contain a settling artefact; the end may drift or morph. Plan for that by generating longer than you need and cutting into the stable middle.
| Scene type | Typical shot length | Notes |
|---|---|---|
| Dialogue | 4–7 seconds | Cut on the line, not after it |
| Suspense build | 6–12 seconds | Let the shot breathe before the cut |
| Action | 0.5–2 seconds | Short shots hide detail problems |
| Montage | 1–2.5 seconds | Rhythm matters more than content |
| Documentary texture | 4–8 seconds | Slower, observational, sound-led |
| Ending | 8–15 seconds | Longer holds create weight |
Two habits improve results immediately. First, cut on motion — edit while something is moving, and the cut disappears. Second, let sound carry the pace. A slow shot with urgent sound design reads as tense; a fast shot with calm sound reads as dreamlike. Pacing is a relationship between image, sound, and duration, not a property of the image alone.
Matching Narrative Intent to the Right Generation Approach
Different shots need different methods. Treating them all the same way is the most common inefficient habit in generative production.
| Approach | Best for | Control level | Consistency risk |
|---|---|---|---|
| Text to video | Establishing shots, mood, abstract sequences | Low | High |
| Image to video | Any shot needing a specific look or composition | Medium | Medium |
| Reference-driven character work | Repeated characters | Medium-high | Medium |
| Video-to-video restyling | Footage you already own | Medium | Low |
| 3D previsualisation plus render | Complex camera moves, precision blocking | High | Low |
| Practical plates plus generated inserts | Documentary and hybrid work | High | Very low |
Use a simple priority rule: the more a shot carries emotion or plot, the more control it deserves. Hero shots get reference images, careful iteration, and multiple candidate takes. Connective shots — transitions, atmosphere, textures — get fast generation and a quick approval. A film where every shot receives equal effort rarely finishes.
A Practical End-to-End Workflow
- Lock the script. No generation before the story stops changing. Rewriting during production multiplies cost.
- Build the beat sheet. Externalise every beat into visible action.
- Produce the shot list with the six-part description and columns for continuity notes and intended approach.
- Do look development in stills. Approve one image per location, per character, and per lighting state. Stills are fast and cheap to iterate; clips are not.
- Assemble an animatic. Cut the approved stills to the script with temporary sound. Problems in structure surface here, before any video exists.
- Generate hero shots first. These define the visual standard the rest must match.
- Generate coverage and connective shots in batches, grouped by location and lighting state so consistency is easier to hold.
- Assemble and rough-cut. Cut for story; ignore imperfect frames for now.
- Design sound early. Ambience, foley, and music do more narrative work than most directors expect, and they mask small visual inconsistencies.
- Grade, then conform to delivery specs. Check aspect ratio, resolution, loudness, and frame rate against your actual destination before you export the master.
Keep each generation session grouped by scene rather than by shot type. Working through a single location in one sitting produces far more consistent material than jumping between scenes.
Common Mistakes and How to Fix Them
- Generating before locking the script. Fix: finish the beat sheet first, even a rough one.
- Writing prompts that try to do too much. Fix: one action per shot. Split anything with and then.
- Inconsistent wording for the same character. Fix: copy and paste the approved description. Never retype it.
- Mixing lighting directions within a scene. Fix: add a light direction field to every shot in the list.
- No animatic. Fix: cut stills first. Structural mistakes are cheap to fix at that stage.
- Treating every shot as a hero shot. Fix: prioritise. Save iteration for what the audience will remember.
- Baking a heavy look into every clip. Fix: generate near-neutral and grade in the edit.
- Ignoring sound until the end. Fix: build a temp track during the animatic.
- Forgetting eyelines. Fix: note who looks where in every dialogue shot.
- Delivering in the wrong format. Fix: confirm specs — ratio, resolution, loudness, frame rate — before the final render.
Most of these mistakes are pre-production failures that only become visible in post. The earlier you catch them, the cheaper they are.
FAQ
Do I need a full script before designing shots?
For anything longer than a minute, yes. A locked script prevents the most expensive kind of rework: regenerating sequences that no longer belong in the film.
How many shots does a short film need?
A three to five minute piece usually holds together with 40 to 80 shots, depending on pacing. Slower, observational work sits at the low end; montage-driven work sits higher.
Can I keep a character consistent across an entire film?
Not perfectly with generation alone. Use reference images, identical wording, and structural workarounds — silhouettes, back-of-head shots, hands, reflections — and reserve sustained face close-ups for shots you capture practically.
Should I always generate the longest clip possible?
Generate longer than the shot requires, then cut into the stable portion. The usable window is usually much shorter than the full clip length.
What is the fastest way to test whether a scene works?
Cut it as an animatic with stills and sound. If it works with static images, it will work with video. If it does not, no amount of rendering will save it.
Where does AI fit best in a small production?
Previsualisation, establishing shots, weather and period material, scale shots, and abstract sequences. It fits worst in performance-driven dialogue scenes, where a real actor and a real camera remain hard to beat.
The Takeaway
Generative tools change how images are made, not what makes them work. Structure, shot purpose, continuity, pacing, and sound still decide whether an audience stays with the film. The directors who get the best results are not the ones with the most elaborate prompts; they are the ones who decided, early and precisely, what each shot had to do — and then chose the simplest method that could deliver it.
