Why Pre-Production Still Matters in AI Video
AI video tools have collapsed the distance between an idea and a moving image. A few years ago, producing a cinematic clip meant hiring a crew, renting gear, and spending days on set. Today, a carefully written prompt can generate photorealistic footage in minutes. The result is a creative environment that is faster, cheaper, and more accessible than anything that came before it.
But speed has a side effect. When generating is easy, planning is often skipped. The typical failure mode of AI video is not technical quality — modern models are stunningly good — it is narrative incoherence. Creators generate clip after clip, each one beautiful on its own, and then struggle to stitch them into something that feels like a story. The footage is right; the plan was missing.
That is why pre-production matters now more than ever. In traditional filmmaking, the director, the cinematographer, and the storyboard artist translate a script into a visual plan before a single frame is shot. AI video needs exactly the same discipline, adapted to a workflow where the camera is a model and the crew is a set of prompts. The creators who produce consistent, repeatable, character-driven videos are not better prompters. They are better planners.
This guide walks through a practical pre-production workflow for AI video: defining your shots, building a storyboard, keeping characters and scenes consistent, and choosing the right model for each part of the story. None of these steps require a film school degree, but together they separate one-off clips from projects that look directed.
What Shot Design Means in an AI Workflow
Shot design is the practice of deciding, before you generate, what the viewer should see in each moment. It covers framing, camera movement, duration, and the emotional intent of a scene. In traditional production, this is the director's job, supported by the cinematographer and the storyboard artist.
In AI video, shot design translates into three concrete artifacts:
- A shot list: a table of every shot in the video, with the action, framing, camera movement, and duration.
- A storyboard: a visual sketch of each shot, showing composition and sequence.
- A style and consistency sheet: a record of character appearance, environment, lighting, and color so that every generation matches.
The shot list is the backbone. For a sixty-second video, you might have ten to fifteen shots. Each row should describe what happens, who is in frame, where the camera is, how it moves, and how long the shot lasts. Writing this down forces you to think in sequence rather than in isolated images. It also gives you a checklist: when a generation fails, you know exactly which shot to regenerate instead of starting over.
The storyboard adds the visual layer. It does not need to be polished art — stick figures with arrows for camera movement are enough. The goal is to see the composition before spending compute on a generation. Many creators skip this step because it feels old-fashioned, yet it is the cheapest way to catch problems. A storyboard that shows a confusing camera angle or a mismatch between two consecutive shots costs minutes to fix; a generated clip that has to be thrown away costs time and money.
The consistency sheet is the piece that AI workflows often forget. Write down the character's hair, clothing, and accessories; the setting's layout; the lighting mood; and the color palette. When you later need to regenerate or extend a scene, this sheet becomes your source of truth.
Turning a Script into a Shot List
Start with a short written outline of the story: beginning, middle, end, and the emotional beat of each section. Then break the outline into beats. A beat is the smallest unit of story — a character enters a room, a door closes, a phone rings. Each beat becomes one or more shots.
For each shot, answer five questions:
- What is happening? (the action)
- Who is on screen? (characters and props)
- How is it framed? (wide, medium, close-up, extreme close-up)
- How does the camera move? (static, pan, tilt, dolly, handheld)
- How long does it last? (seconds)
Write the answers in a table. Keep each action simple enough for a model to understand. "The detective walks toward the window and stops" generates more reliably than "The detective contemplates the rain with quiet melancholy." Emotion is communicated through action, lighting, and music — not by asking the model to read minds.
When the shot list is complete, review it for logic. Does shot four follow shot three in space and time? Would a viewer be confused about where the characters are? If the answer is yes, adjust before generating.
Building a Storyboard for AI Generators
A storyboard does not need to be beautiful, but it needs to communicate. For each shot in your list, sketch three things: the frame (what is in the composition), the camera (where it points and how it moves), and the connection (how this shot enters and exits — cuts, fades, transitions).
If you cannot draw, you have options. Generate a reference image for each shot using an image model, then arrange the images in sequence. This produces a lookbook storyboard that doubles as your visual bible. Many creators do this anyway to lock down character design before generating video.
When using generated references, keep the consistency rules in mind. The same character description must be used across reference images, or the character will drift between shots. Fix the seed, the model, and the style prompt if the tool supports them. Reference images are not final footage; they are the contract between you and the generation step.
The storyboard also reveals pacing problems. If ten shots in a row are close-ups, the sequence will feel claustrophobic. If every shot is a wide, the video will feel static. Vary shot size and camera movement to match the energy of the scene: fast, handheld movement for tension; slow, locked-off frames for calm.
Keeping Characters and Scenes Consistent
Consistency is the single biggest quality problem in AI video. A character's face changes between shots, a room rearranges itself, or the lighting shifts scene to scene. The cause is usually simple: the model treats each generation as a new scene and invents details.
The fix is a pre-production consistency system. Build it before generating, not after.
First, define the character completely. Hair color and style, eye color, body type, clothing, and any distinctive accessory. Write the description as a reusable character card. Use the exact same phrasing in every prompt that includes the character — small wording changes cause visible drift.
Second, lock the environment. Describe the room, the city, the lighting, and the time of day in a reusable environment card. If a scene takes place in a specific location, generate a reference image first and use it as an input image for video generation where the tool supports image-to-video.
Third, use multi-image fusion and keyframes where available. Several models now accept multiple reference images to lock identity, and keyframe controls let you fix the first and last frame of a clip. This is the closest thing AI video has to a storyboard inside the generation step: you define the boundaries, and the model fills in the motion between them.
Finally, keep a log of what worked. When you find a prompt combination that produces a consistent character, save it. Consistency is a repeatable process, not luck.
Choosing the Right Model for Each Shot
No single model is best at everything. Modern AI video platforms route different jobs to different engines, and the director's job is to pick the right tool per shot. As a general rule:
- Use high-realism models for footage that must look photographic: products, live-action-style scenes, and anything where texture matters.
- Use motion-focused models for dynamic sequences: action, camera moves, and fast transitions.
- Use stylized models for animation, illustration, and branded looks.
- Use image-to-video models when you already have a locked reference or storyboard frame.
Examples of engines you will meet in most platforms include Flux for still-image quality and photorealistic stills, Runway for creative video generation and control, Sora for long, coherent scenes with strong physics, Kling for prompt adherence and a distinctive aesthetic, and Pika or Luma for fast iteration and playful motion. Each has strengths; none is universally best.
Budget is part of the decision too. Some models consume far more compute than others, so match the model's cost to the importance of the shot. Reserve the most expensive engine for hero shots — the moments the viewer will remember — and use cheaper models for filler and transitions. This is exactly how a director allocates a real production budget.
A Step-by-Step Pre-Production Workflow
Putting it together, here is a repeatable workflow:
- Write a one-page outline of the story with a clear beginning, middle, and end.
- Break the outline into beats and build a shot list (action, framing, camera, duration).
- Create a consistency sheet: character cards, environment cards, lighting, and palette.
- Generate reference images for key shots and arrange them into a storyboard.
- Review the storyboard for pacing, logic, and composition; fix problems now.
- Choose the model for each shot based on realism, motion, style, and budget.
- Generate shot by shot, using reference images and keyframes to hold consistency.
- Review the assembled sequence, identify weak shots, and regenerate only those.
Most creators who adopt this workflow report that they generate less and keep more. The planning time is repaid many times over in reduced rework.
Common Mistakes and How to Avoid Them
Skipping the shot list. Without a list, you cannot know what is missing. Spend ten minutes writing it.
Generating before locking the character. The character is the anchor of the story. Lock it in reference images first.
Overloading the prompt. One action per prompt. Add style and environment from the consistency sheet, not from improvised adjectives.
Using the same model for everything. Match the engine to the shot, and protect the budget for hero moments.
Ignoring duration limits. Models generate clips of limited length. Plan shots inside the clip-length limit, or plan to extend with continuation tools.
Planning for the Platform
Pre-production should also account for where the video will live. A nine-by-sixteen vertical video for social feeds needs tighter framing and faster pacing than a sixteen-by-nine widescreen video made for a website or a TV-style screen. Decide the aspect ratio in pre-production, because it changes shot design: vertical video favors single-subject compositions and closer framing, while widescreen gives you room for environments and multi-subject shots.
The same logic applies to duration. A thirty-second social video can sustain two or three beats; a three-minute narrative needs a full arc with setup, escalation, and resolution. Match your shot list to the target length, and keep the platform's playback habits in mind — mobile viewers often watch without sound, so plan for captions and visual storytelling that works muted.
Sound design is worth planning too, even in AI video. Many tools generate clips without usable audio, so decide early whether the final video will use voiceover, music, or effects. Leave room in the timeline for those layers, and write the shot list with pacing that lets the audio breathe.
FAQ
Do I need to know how to draw to storyboard? No. Stick figures, arrows, and generated reference images are enough. The purpose is communication, not art.
How long should my shot list be? For a short social video, eight to fifteen shots. For a longer narrative, break the video into scenes and build a list per scene.
Can I keep a character consistent across different models? Sometimes. Using the same reference image and character card improves the odds, but moving between very different engines still causes drift. Prefer one primary model for a character's scenes.
What is the difference between a keyframe and a storyboard? A storyboard is a planning document you create before generating. A keyframe is a fixed frame you feed into the model to control the start or end of a generated clip. They serve the same goal — control — at different stages.
How many shots can one prompt generate? That depends on the platform and model. Read the tool's limits and plan your shot list to fit them, using continuation or additional clips when a shot needs more time.
Final Thoughts
AI video has removed most of the technical barriers to filmmaking, but it has not removed the need for direction. The gap between a pile of impressive clips and a video that feels intentional is planning. Shot lists, storyboards, consistency sheets, and deliberate model selection are not bureaucratic overhead — they are the craft. Spend your planning time before you spend your generation budget, and the footage will finally look the way you imagined it.
