Previs is a question machine
AI previsualization helps a cinematographer choose between shot approaches before committing crew time. The useful output is a decision pack: a scene brief, comparable moving studies, annotated frames and a record of what still needs physical verification.
It is a question machine. Can the scene hold from this angle? Does the blocking need a practical source? Is the camera too close too early? Does the location feel expensive for the right reason? Does the weather idea create drama or just noise? Can the director and producer finally see the same problem?
Google Flow is interesting because it puts generative video inside a creative studio frame rather than only a prompt field. For cinematographers, that is the important shift. Previs needs iteration, ingredients, references, and a place to compare versions.
The point is not to replace prep. The point is to make prep more specific.
Look development that moves
Moodboards are useful, but they can hide motion problems.
A still reference can sell colour, texture, contrast, and composition. It cannot show whether the camera move feels motivated, whether rain destroys the silhouette, whether the subject disappears against the background, or whether the grade idea collapses once the shot moves through space.
AI video can turn look development into moving hypothesis. That is valuable before a scout, before a deck is locked, or before a director sells a tone nobody has stress-tested.
The output still needs suspicion. A moving reference can feel more authoritative than it deserves.
Blocking is where previs earns its place
The strongest cinematography use case is blocking pressure.
If a character crosses a kitchen, turns into backlight, opens a fridge, hears a sound, and decides not to speak, the shot is not just about style. It is about timing, geography, performance, practical sources, and where the cut could happen.
AI previs can help test those choices cheaply. It can show that a planned move is ornamental, that a wide shot needs one more beat, or that a location will not hold the emotional geography the script assumes.
That is useful because it makes production questions visible early.
A six-field previs decision card
Purpose: name the decision the clip should help settle. For the kitchen scene below, the question is whether the audience should discover the doorway before or after the character hears a sound.
Geography: draw the fridge, doorway, actor marks, camera start and end, and line of action. Use a location photo or measured plan where available; label invented dimensions as assumptions.
Camera: record framing, height, distance, intended lens perspective and movement. A lens number in a prompt is a visual request, not evidence of an optically accurate camera simulation. Use the lens perspective guide to compare observable framing.
Light: identify the motivated source, direction, quality and subject exposure priority. Keep the fridge practical and doorway backlight in the same places across versions; use the lighting prompt card and exposure anchors.
Timing: mark the action beats and proposed cut point. Separate actor movement from camera movement, then test motion cadence in a short edit.
Acceptance: name the reviewer, version and pass condition, plus the physical checks still open. The card should be readable without access to the generation account.
Worked scene: a kitchen, a fridge and an unseen doorway
This is an illustrative prep exercise, not a report of a completed production or tool test. A character crosses a kitchen, opens a fridge, hears a sound from the doorway and stops. The story decision is when the audience discovers the possible threat.
Prepare one reference frame with the fridge screen left, doorway deep right and an unobstructed actor path. Draw the plan separately. Make a baseline locked wide, then a second version with a slow lateral move revealing the doorway at the sound cue. Keep the reference, blocking, duration and light intention fixed; change only the camera approach.
For a Runway Gen-4 study, begin with a simple motion instruction such as: The subject crosses to the fridge and opens it. Cool practical light reaches their face. They pause and turn toward the doorway. Locked wide shot. Treat this as a starting prompt, not a guaranteed result; split the action into separate studies if the model cannot hold the beats.
Runway’s Gen-4 prompting guide says the input image establishes composition, colour, lighting and style, while text should focus on motion. It recommends adding one element at a time. Google DeepMind’s Veo guide also describes framing, camera movement and lighting as prompt controls. These are ways to request a study, not guarantees of continuity or physical feasibility.
Cut each approach against the same reaction insert. Compare whether the doorway reveal steals attention from the performance, whether the turn reads, and whether the cut preserves screen direction. If the geometry changes midway, reject that study for the blocking decision even if its lighting looks attractive.
Acceptance review: watch the sequence, then scout the solution
First pass, story: can the director identify the intended reveal and emotional beat without reading the prompt? Second pass, continuity: sample the opening, action turn and closing frames for actor identity, door position, eyeline, light direction and exposure. Third pass, execution: can camera, grip, lighting and art departments explain how to reproduce the chosen approach?
Record each result as accepted for intent, revise, or unresolved. A study can pass the story review while remaining unresolved for camera clearance. Do not let a beautiful last frame conceal a doorway that moved during the shot.
At the location, verify room dimensions, door swing, operator and rig space, lens field of view, practical output, reflections, sound and safe access. If the measured room cannot support the move, retain the story goal and revise the shot. AI previs does not certify those conditions.
What the director, producer and crew should receive
Hand over the scene brief and plan, numbered shot cards, a short comparison edit, annotated contact sheets, and a decision log naming the selected version and approver. Include source references, prompts, model/version where available, export settings and asset permissions so another person can reopen the work.
The director receives the chosen reveal and performance beats. The cinematographer and crew receive camera and light intentions plus unresolved scout checks. The producer receives assumptions that affect rig, location, schedule and post scope. Label each clip as a moving reference rather than a promise of the final shot.
For a commissioned pilot, agree one scene, a bounded set of variants, named reviewers and the handoff format before generating. Review my film and television work for cinematography context, or discuss a reviewable AI production workflow when the team needs controls, approvals and a portable handover.
Impossible shots need honest labels
AI previs is especially good at impossible shots: crowds, weather, creatures, period streets, explosions, impossible camera rigs, or abstract memory sequences.
Those tests can unlock a conversation. They can also oversell a plan. A generated clip does not answer stunt safety, legal clearance, VFX cost, performer work, art department scale, insurance, weather, sound, or post schedule.
Every generated pitch asset needs an honest label: sketch, reference, temporary plate, technical target, or final candidate. A convincing camera path is still a proposal until the operator and crew establish access, clearance and a safe way to execute it.
If the label is wrong, the previs becomes a budget trap.
A cinematographer's rule
The rule is simple: use AI previs to create better decisions, not premature certainty.
Keep the prompt, references, version notes, and approval context with every clip. Mark what the clip proves and what it does not prove. Share it with the director as a route to questions, not as a finished answer.
Previs is powerful when it protects intent before production pressure arrives. It is dangerous when it becomes a beautiful excuse to skip thinking.
The useful version makes the shoot sharper.
Build
Need a repeatable AI production workflow?
Mike designs the tools, review loops, and publishing systems that make it usable.
Launching a business of your own? Founder Launch OS connects the brand, offer, website and visual campaign in one guided Codex or Claude Code workspace.
Keep reading
