A practical shot-by-shot workflow for making short films with AI: pre-production, character consistency, cinematography, rendering, post, and FAQs.
You have a logline, a folder of reference images, and a weekend. A decade ago that combination produced a script and a lot of waiting. Today it can produce a finished short film — provided you treat generative tools like a disciplined production crew rather than a slot machine.
The difference between a reel that looks like a demo and a short that holds an audience is almost never the model. It is the pre-production decisions, the continuity system, and the edit. This guide walks the full pipeline from first concept to final export, with the specific habits that keep AI-assisted shorts coherent, cinematic, and finishable.
What actually changed in AI-assisted filmmaking
The mechanics of generation get the headlines, but the real shift is economic. Iteration is now nearly free. A camera test that once cost a day of crew time costs ninety seconds of compute. That changes how you should plan: you can afford to explore, and you cannot afford to plan in your head.
The bottleneck has moved. It is no longer "can we shoot this?" but "can we keep this consistent across forty shots?" Continuity of faces, wardrobe, props, light direction, and screen geography is the craft problem of AI filmmaking. Every section below is really about solving that problem in a different phase of the pipeline.
A second shift: you now direct with constraints rather than adjectives. A prompt like "cinematic, moody, masterpiece" gives the model nothing to hold onto. A reference image of your lead actor, a lighting diagram, and a stated lens give it almost everything.
Pre-production: build a visual bible before you generate a frame
Write the film on paper first. For a short, this means a one-page treatment, a beat sheet of eight to twelve beats, and a locked script. If the film is under four minutes, dialogue should be sparse — AI video still handles performance nuance better in silence and voiceover than in lip-synced argument scenes.
Then build the visual bible. This is a single document (or Figma board, or Notion page) that any collaborator — human or model — can read to reproduce your film's look.
The four reference assets every AI short needs
- Character sheet. Four to six images of each principal: front, three-quarter, profile, full body, plus two wardrobe variations. Generate these with an image tool, then pick the two strongest and delete the rest. Ambiguity at this stage becomes inconsistency later.
- Location sheet. Wide, medium, and detail views of every set, with a note on window direction and practical light sources. If a room has a window on camera left, that fact must survive twenty shots.
- Palette and grade reference. Three to five stills from real films that share your intended look. Be specific about contrast, saturation, and black level.
- Motion reference. Short clips showing camera behavior — a slow dolly, a handheld drift, a locked-off wide. Verbal descriptions of movement are the weakest link in prompting; footage is the strongest.
Turning the script into a shot list an AI can execute
Traditional shot lists assume a crew that understands context. AI generation does not. Your shot list needs to be explicit enough to be executed by something with no memory of the previous shot.
Use a spreadsheet with these columns:
| Field | What goes in it |
|---|---|
| Shot ID | SC02_SH07 |
| Story beat | What the audience must learn |
| Framing | Wide / medium / close / insert |
| Lens | 24mm, 50mm, 85mm |
| Camera move | Locked, push, drift, orbit |
| Characters | Who is on screen, in what wardrobe |
| Light | Source, direction, time of day |
| Duration | Target seconds in the cut |
| Reference | Which bible images apply |
| Notes | Anything the model must not change |
A filled row looks like: SC02_SH07 | She realizes the letter is gone | Close, 85mm | Locked, slight breathing | Mira, grey coat | Single window camera left, overcast | 3.5s | char_mira_03, loc_kitchen_wide | Keep hands out of frame.
That level of specificity feels excessive until shot nineteen renders a different apartment.
Storyboards and animatics: fail cheaply, fail early
Generate your boards as stills, not video. Stills are fast, easy to revise, and force you to solve composition and blocking before motion complicates everything. You can build a full board set in an afternoon using the AI image generator.
Then cut an animatic. Drop the boards into an editor at your intended running time, add temporary voiceover, a temp music track, and rough sound effects. Watch it three times. Almost every structural problem in a short film — a slow opening, an unclear motivation, a flat ending — is visible in the animatic and invisible in the shot list.
Lock the animatic before you generate a single moving frame. Every shot you generate before the cut is locked is a shot you will probably regenerate.
Two rules that save days:
- Cut for coverage, not completeness. If the animatic works with twenty-two shots, do not generate thirty-five.
- Identify your three "hero" shots. These are the images that sell the film — the poster frame, the emotional close-up, the final image. Give them triple the generation budget and the most attention.
Choosing the right generator for each shot type
No single tool wins every category. Treat your available generators as a specialist crew and route each shot accordingly.
| Shot type | What matters most | Practical approach |
|---|---|---|
| Establishing wide | Composition, scale, atmosphere | Image-to-video from a strong generated plate, minimal motion |
| Dialogue close-up | Facial stability, micro-expression | Short durations, locked camera, one subject per generation |
| Action beat | Motion coherence, physics | Shorter clips cut together rather than one long take |
| Insert / detail | Texture, shallow depth of field | Often better as a still with a slow push than a full generation |
| Transition | Camera continuity | Match direction and speed to the outgoing shot |
| Performance montage | Emotional read | Many short takes, edit for rhythm |
The routing decision matters more than the model version. A 3-second locked-off close-up from a good reference plate will outperform a 10-second ambitious camera move in almost every case.
When you are ready to move from boards to motion, generate in the AI video generator using the board image as the first frame — image-driven generation is the most reliable path to the composition you already approved.
Character and object persistence across shots
This is where most AI shorts fall apart. A face drifts over six shots, a coat changes color, a coffee cup moves between hands. Fix it systematically.
The identity block
Write a fixed paragraph describing each principal — age range, build, hair, distinguishing features, wardrobe, posture — and paste it verbatim into every prompt where they appear. Never paraphrase it. Variation in the description produces variation in the output.
Anchor with images, not words
Always attach the same two character reference images. If your tool supports multi-image referencing, use one image for face and one for wardrobe or body. Keep the reference set frozen for the whole production; swapping in a "better" reference mid-film reintroduces drift.
Control what you can hide
When consistency resists, direct around it:
- Cut to over-the-shoulder or reaction shots instead of the problematic face.
- Use silhouette, backlight, or shadow for characters in motion.
- Insert hands, props, and environment detail to carry the scene.
- Keep the most demanding close-ups short and static.
Editing is a continuity tool. A two-frame cutaway can rescue a shot that would otherwise need five regenerations.
Directing cinematography through prompts
Write prompts in a fixed order so you can debug them: subject → action → framing and lens → camera movement → light → palette → duration. When something goes wrong, you know which clause to change.
A weak prompt: A woman walks into a dark house, cinematic, dramatic, 4k.
A directed prompt: Medium-wide, 32mm, a woman in a grey coat steps through a front door into an unlit hallway, she pauses and looks right. Camera holds locked off then drifts forward slowly. Single hard light from the doorway behind her, deep shadow on the far wall, cool desaturated palette, 4 seconds.
The second version gives the model a subject, a beat of behavior, a lens, a camera instruction, a light source with a direction, and a length. That is a shot, not a wish.
Keep one rule in mind: one dominant motion per shot. If the camera pushes in, the subject should be nearly still. If the subject runs, lock the camera. Competing motions produce warping.
Borrow a working vocabulary from your prompt library rather than inventing phrasing each time — reuse is what makes results repeatable. A curated set of prompt templates will keep your language consistent across the whole film.
Rendering discipline: takes, naming, and versioning
Generous iteration only helps if you can find things again.
Generate four to six takes per shot, then stop and review them as a contact sheet before generating more. Choosing between takes you already have is faster than describing a better one.
Name and log everything:
- Files:
SC02_SH07_takeC_v02.mp4 - A shot log with prompt text, reference images used, model, seed if available, and a one-word verdict (keep / maybe / reject)
- A single "selects" folder containing only approved takes
Batch by location and lighting. Generating all kitchen shots in one session keeps you in the same descriptive headspace and reduces drift. Batch by shot type too — do all your locked-off close-ups together, then all your wides.
Set a hard take limit per shot, usually six to ten. If a shot fails beyond that, the problem is the concept, not the take. Simplify the shot: fewer characters, less motion, shorter duration.
Post-production: where AI shorts become films
Editing is where the audience decides whether your film is real. Three priorities, in order.
1. Assemble for rhythm, not coverage. Cut on motion and on eyeline. Use the shortest version of every shot that still reads. AI-generated footage usually has a slightly soft final beat — trim it. Keep twelve-frame handles on everything for transitions and speed ramps.
2. Build sound before you polish picture. Room tone under every scene, foley for footsteps and fabric, a low bed under tension, silence where it earns attention. Sound design does more for perceived production value than any upscale. If a shot's visual is weak, a door slam or a sharp breath can carry it.
3. Finish deliberately. Upscale only what survives the cut. Apply one consistent grade across the film — matching contrast and black levels will unify footage generated on different days more effectively than regeneration. Add grain sparingly; it hides minor shimmer but too much reads as a filter.
Fixing artifacts in the edit instead of regenerating
Most artifacts are survivable. Warped hands: cut two frames earlier, or add a shadow pass. Flickering background: overlay a subtle light-leak or adjust exposure across the shot. Mismatched eyeline: flip the shot if screen direction allows, or insert a reaction shot between them. Reserve regeneration for shots that carry story weight.
Scope, schedule, and the mistakes that sink first films
Be honest about scope. A first AI short should be 90 to 180 seconds, one or two characters, two or three locations, minimal dialogue, and twenty to thirty shots. That is three to four weeks of part-time work with a real edit.
The recurring mistakes:
- Generating before the animatic is locked. Everything downstream inherits the problem.
- Too many characters. Every additional face multiplies continuity work.
- Long shots. Anything over six seconds invites motion artifacts; cut shorter and more often.
- Describing instead of referencing. Words drift, images hold.
- Skipping sound. Silent cuts feel like tests, not films.
- Polishing one shot for a week. Fix the sequence, not the frame.
- No shot log. You will lose the good take.
- Chasing a specific model. Route by shot type and move on.
FAQ
How long should an AI-assisted short film be? For a first project, 90 to 180 seconds. It is long enough to demonstrate craft and short enough to finish. Longer films are usually a sequence of well-made short scenes anyway.
Do I need a script if the visuals are generated? Yes, more than ever. Generation has no narrative memory. The script and shot list are the only places your story exists in a stable form.
How do I keep a character's face consistent? Freeze two reference images per character, paste an identical descriptive block into every prompt, keep shot durations short, and direct around the face when consistency resists — reaction shots, silhouettes, and inserts.
Can I mix footage from different generators in one film? Absolutely, and you probably should. Unify them with a consistent grade, matched black levels, and shared sound design. Audiences read continuity through tone and rhythm far more than through pixel-level consistency.
What is the biggest quality jump for the least effort? Sound design and trimming. A tight cut with real room tone and foley reads as professional even when individual frames are imperfect.
How many takes should I generate per shot? Four to six before reviewing, with a hard ceiling around ten. Beyond that, simplify the shot rather than spending more attempts.
Bring your short film to screen with Orelon
Orelon is built for cinematic ideas in motion — a workspace where your reference plates, character sheets, and shot prompts live alongside the generation itself, so continuity is a system rather than a memory exercise. Start with a board, drive it into motion, and keep every take organized from first frame to final cut.
Open the AI video generator to render your first shot, browse video templates if you want a structural starting point, and read more production workflows on the Orelon blog. Your logline is already a film. Give it a shot list.

