Orelon logoOrelon
Tarifs

Designing an AI Video Workflow That Actually Finishes

5 nov. 2026 · Par Orelon Team

Explorez les modèles vidéo IA

Parcourez quelques créations de la communauté pour trouver l’inspiration, puis ouvrez n’importe quel modèle pour continuer à créer dans Orelon.

A stage-by-stage AI video pipeline: briefs, treatments, shot lists, motion prompts, consistency anchors, editing, sound, and the fixes that save a project.

Most disappointing AI video does not fail at the render. It fails ten minutes earlier, when a vague idea is typed into a prompt field and everyone waits to see what the machine invents. Modern video models can produce convincing motion, believable light, and camera moves that once required a crew. What they cannot do is decide what the piece is about, keep a jacket the same colour in shot six as in shot one, or notice that your cut has no rhythm.

That gap is what a workflow fills. Below is a repeatable pipeline for AI video production — brief, treatment, boards, shot list, prompts, routing, assembly — with the decision points spelled out at each stage. It is written for people who want finished videos, not a folder of attractive fragments.

Why AI Video Projects Stall Before the First Render

Three symptoms show up again and again, and none of them are model problems.

Symptom one: the folder of orphans. Twenty beautiful clips, no sequence. This happens when each generation is treated as an experiment rather than a shot that belongs to a list. The fix is boring: write the list first, then generate only what is on it.

Symptom two: the drifting character. The face is close in shot one, softer in shot three, and wearing a different collar in shot five. People respond by writing longer prompts. Longer prompts do not fix drift; anchoring frames and a written continuity sheet do.

Symptom three: the endless take. One shot has been regenerated fourteen times. If a shot has failed six times, the problem is almost never the wording. It is usually conflicting camera instructions, an impossible duration, or the wrong engine for that content type.

What a real pipeline buys you

A pipeline is not bureaucracy. It buys three things that compound.

First, debuggability. When the finished video feels flat, you can trace the cause backwards: a shot list that described feelings instead of actions, a treatment with no point of view, or a brief that never named an audience. Without artifacts, the only diagnosis available is "the model is bad today."

Second, spent compute in the right place. Most projects burn the majority of their generation time on shots that get cut. That is a planning failure, not a rendering failure.

Third, cheap revisions. When a client asks for a different ending, a documented pipeline turns that into a twenty-minute job rather than a restart.

The stages overlap in practice. You will rewrite a prompt after seeing a render, and you will re-board a sequence after the edit reveals a hole. The point is not a rigid waterfall; it is knowing which artifact is currently wrong.

The Pipeline in Five Stages

Each stage ends with a locked artifact you can review and hand to someone else.

Stage Input Output artifact Typical time (30s piece)
Brief Goal and audience One page of constraints 10 min
Treatment Brief One page of visible action 20 min
Boards and shot list Treatment Reference frames plus a table 20–30 min
Prompt pack and renders Shot list Versioned prompts, 2–3 takes per shot 30–45 min
Assembly Clips Cut, sound, grade 30–60 min

Two habits make this work. Name files with the shot ID so a winning render is findable a week later. And keep prompts in a shared library next to the boards, not in a chat window that scrolls away.

Pre-Production: Briefs, Treatments, and Reference Frames

The first hour decides most of the outcome. Spend it on paper, not on generation.

The brief answers a question

A brief is not a mood board. It answers four questions in plain sentences: who watches this, what should they do afterwards, where will it be seen, and what must never appear. That last item matters more than people expect — a logo that cannot be cropped, a colour that clashes with a brand palette, a hand gesture that reads badly in one market.

Write the brief in bullet points, five to eight lines. If it runs past a page, you are writing a pitch deck.

The treatment describes what the camera sees

A brief says what the video should achieve. A treatment says what the frame contains. Write it as prose, in present tense, one page, describing locations, light, actions, and the emotional beat of each section.

Ban adjectives that cannot be photographed. "Premium," "dynamic," and "engaging" mean nothing to a model or a viewer. Replace them with observable facts: brushed steel against dark seamless, water beading on a cold surface, a slow push-in from waist height.

A usable treatment paragraph looks like this:

Dawn light through a tall window, dust visible in the beam. A woman in a linen shirt sets a ceramic cup on a steel bench and steps back. The camera is low and close, the cup in the left third. Steam rises. Cut to black before the room sound arrives.

That paragraph already contains a shot, a lens position, and a rhythm. Twelve of them and you have a thirty-second video.

Timebox the treatment. Twenty minutes is enough for a short piece, an hour for a two-minute one. If you catch yourself writing marketing copy instead of describing images, stop and restart with the question: what does the frame show?

Reference frames settle composition

Adjectives are open-ended; images are not. This is the most useful principle in AI video: an existing image constrains a model more tightly than a paragraph of description. That is why boards come before prompts.

Generate one reference frame per shot in the target aspect ratio with an AI image generator. They do not have to be beautiful. They have to settle composition, palette, subject, and lens. Once those frames exist, image-to-video becomes a much smaller problem — describing movement inside a known frame rather than inventing everything at once.

Keep the frames as a contact sheet in shot order. Reading the sequence as thumbnails tells you in five seconds whether the piece has visual variety or repeats the same mid-shot eleven times.

Shot Lists That Constrain the Model

A treatment is prose. A shot list is a machine-readable plan. Build it with fixed columns so nothing important lives only in your head.

Field What it captures
Shot ID Stable reference used in filenames and prompts
Duration Target seconds on the timeline
Subject Who or what the frame is about
Action Movement within the shot, one sentence
Camera Static, pan, dolly, handheld, crane, orbit
Lens Wide, normal, long, macro; approximate focal length
Light Direction, quality, time of day
Transition How the shot leaves
Notes Continuity, wardrobe, colour, sound cue

Two rules that keep a shot list honest

Keep shots short. Most generated clips are strongest between two and five seconds. A sequence of short shots cuts better than a few long ones, and it hides small inconsistencies because the eye never has time to study them.

Write actions a camera operator could obey. "She turns toward the window" is directable. "She feels nostalgic" is not. If the action column contains a feeling, the model will invent a gesture to represent it, and the gesture will probably be a shrug.

Sort the list by shooting efficiency rather than story order when you generate — all the shots that share a set, a wardrobe, or a lens family together. It costs nothing and reduces drift, because each batch inherits the same visual context.

Prompts That Describe Motion, Not Just Scenes

A still-image prompt describes a frame. A video prompt has to describe what changes inside that frame. Most weak AI video comes from prompts that describe a photograph and hope motion appears on its own.

A prompt order that survives iteration

Write the prompt in a fixed order so you always know which line to change:

  1. Subject and wardrobe — who is in frame and what they wear.
  2. Action — one clear movement with a beginning and an end.
  3. Camera — position, height, and movement, stated once.
  4. Lens and depth — wide, normal, long, macro; shallow or deep focus.
  5. Light — direction, hardness, colour temperature.
  6. Environment — location, weather, background activity.
  7. Style — film stock feel, grain, contrast, palette.
  8. Pace and duration — slow, brisk, hold, cut on movement.

A worked product shot using that order:

Macro shot of a matte black camera body resting on dark slate, faint dust visible. The camera slowly arcs fifteen degrees right at table height, shallow depth of field, focus held on the lens mount. Single soft key from upper left with a cool rim from behind. Muted palette, subtle grain, no text, five seconds, hold to the end.

Three mistakes that flatten motion

Noun stacking. Listing objects — "city, neon, rain, cyberpunk, cinematic" — hands the model a scene but no event. Add one verb and the shot gains a shape.

Conflicting camera instructions. "Slow dolly in while orbiting handheld" is two shots pretending to be one. Pick a single movement per clip. If you need both, cut between them.

Overlong prompts. Past roughly eighty to a hundred words, extra clauses begin cancelling each other out. If a shot needs more than that, it is probably two shots.

Iterate in threes and change one variable

Generate three takes of the same shot, then change exactly one thing — camera height, or light direction, or the verb — and generate three more. Changing three variables at once teaches you nothing about which one mattered. Log the winning version in a prompt library with the shot ID so it is reproducible rather than remembered.

A practical note on scale: if you are generating dozens of takes with identical parameters, scripting against an API pays for itself. For a single hero shot, manual iteration is almost always faster than wiring anything up.

Consistency Across a Sequence

A sequence feels professional when the shots look like they came from the same production. Consistency is not a single setting; it is a small set of locked decisions made early.

Lock the look before shot ten. Palette, contrast, grain, black level, lens family. Write them into a short style block you paste into every prompt: "muted teal and warm amber, soft contrast, fine grain, 35mm equivalent, shallow focus." Ten words, applied everywhere, do more for coherence than any single clever prompt.

Anchor every shot to a first frame. Render the opening frame of each shot as an image, then use it as the start frame for image-to-video. Characters and sets drift far less when every clip begins from a controlled still.

Keep a continuity sheet. Wardrobe, hair, props, time of day, and which side of the frame the subject exits. Ten lines prevent the classic mistake of a character wearing a different jacket in every shot, and they prevent the subtler error of exiting left in one shot and entering left in the next.

Grade at the end, not in the prompt. Models approximate colour badly. Get shots close, then unify everything with one grade in the edit. A shared look applied across the timeline does more for perceived quality than another round of rendering.

Routing Shots and Choosing Engines

No single model is best at everything. Some handle human performance and dialogue well; others are stronger on stylised motion, product macro, or wide landscape. Routing means matching each shot on your list to the engine most likely to render it cleanly on the first or second attempt.

Criterion Question to ask
Content type Dialogue, human action, product, landscape, or abstract?
Motion complexity Simple push, or crowds and cloth and water?
Consistency need Must this shot match a character or set exactly?
Turnaround Does it need to be right in one pass, or can it iterate?
Duration Does the engine hold quality past four seconds?

Build a routing sheet: shot ID, chosen engine, aspect ratio, duration, fallback engine. The fallback column matters more than it looks. When a shot fails twice on the primary engine, switching engines is usually faster than a third attempt at rewording. Different training data produces different failure modes, so a change of engine often solves a problem that more words cannot.

If you are comparing platforms before committing, an alternatives overview is a reasonable starting point, and head-to-head pages such as Orelon vs Kling AI help when one specific capability is the deciding factor. Judge engines on your own footage, not on demo reels: run the same three shots — one face, one product macro, one wide exterior — through each candidate and compare.

Assembly, Sound, and Post-Production Fixes

The edit is where footage becomes a video. Four habits carry most of the weight.

Trim the edges. Generated clips often warp in the first and last few frames. Cutting six to twelve frames off each end makes the artefact disappear without a re-render.

Cut on motion. Place the transition at the moment the subject or camera is moving. Motion masks the cut and hides small differences between takes.

Design sound early. Ambience, foley, and music do more for believability than another generation pass. Room tone, a door closing, cloth moving — these tell the audience the space is real. Add them before deciding the picture is finished, because half your picture problems are actually silence problems.

Check small. Watch the sequence at thumbnail size on a phone. Problems that vanish at that scale were never worth another render, and problems that survive it are the ones to fix.

Then grade, add titles, and export. Keep the project file organised by shot ID so revisions do not require hunting through a downloads folder.

Failure Modes, Fixes, and a 90-Minute Worked Example

Symptom Likely cause Repair
Shots look like stock footage Treatment had no point of view Add an observer: low angle, foreground object, deliberate imperfection
Character changes between shots No first-frame anchor Anchor every clip to a still plus a continuity sheet
Everything feels slow Clips are too long Shorten to two to four seconds and cut faster
Renders look mushy Motion complexity too high Simplify the action, reduce elements, or switch engines
Whole video feels generic No palette or angle constraint Lock a style block, pick one unusual angle, remove one element per shot
Repeated failure on one shot Impossible duration or conflicting camera Split into two shots or change engine

A thirty-second teaser, start to finish

Six shots, one person, one fictional coffee brand.

0–15 minutes, brief and treatment. One page: a dawn kitchen, three actions, one product close-up, one exit. Six beats, indoors, no dialogue.

15–30 minutes, boards and shot list. Six reference frames at 16:9, then a table with durations of two to four seconds. The product shot is written as a macro arc; the exit shot is written as a static wide with the subject leaving frame right.

30–60 minutes, prompts and renders. Three takes per shot, one variable changed between rounds. Two shots move to a fallback engine after repeated warping at the edges.

60–80 minutes, edit. Cut to a temporary track, trim edges, reorder two shots so the product appears later and the ending lands harder.

80–90 minutes, sound and grade. Room tone, a cup placement, one warm grade across all six clips.

The result is not a masterpiece, but it is coherent, and every decision can be traced. Changing the ending from there is a twenty-minute job.

FAQ

Do I need storyboards if I am generating from text? You do not need drawings, but you need a decision about composition before generation. A rough reference frame removes ambiguity that no adjective can fix.

How many takes per shot is reasonable? Three per round, two rounds, with one variable changed between rounds. If a shot has failed six times, the plan is wrong, not the prompt.

What clip length works best? Two to five seconds for most shots. Longer clips are possible but drift more, and short shots cut better.

Can one engine handle an entire video? Sometimes, but routing usually wins. Different shots reward different strengths, and a fallback saves an afternoon.

How do I keep a character consistent across a series? Anchor every shot to a first-frame image, maintain a continuity sheet, and constrain wardrobe and lens in the prompt. Consistency is a system, not a phrase.

Where does the time actually go? Editing and sound, not generation. Budget at least as long for assembly as for rendering, and your first estimate will still be optimistic.

How do I stop prompts from growing over time? Version them and cap the length. When a prompt passes about a hundred words, split the shot instead of adding clauses.

What should I do first if the video feels flat? Go back to the treatment. Flatness is almost always a point-of-view problem, not a rendering problem.

Start With One Shot

The pipeline above is deliberately unglamorous. It replaces hoping for a lucky render with a series of small, checkable decisions: a brief that says something, a treatment you can read aloud, a shot list a stranger could follow, prompts that describe motion, and an edit that hides the seams.

You do not need to adopt all of it at once. Pick a single shot, write the prompt in the fixed order above, generate three takes, and cut them into something twenty seconds long. Then repeat the exercise with a shot list, then with a style block. The second pass will be faster, and the third will feel like a process rather than an experiment.

When you are ready to put the system to work, start generating with Orelon — build one sequence, keep the prompt log, and let the next project inherit what you learned.