Ask a first-time feature director where the money went and you will rarely hear “cameras.” You will hear “the days we lost.” A location that read beautifully in a scout photo and turned out to be useless at the time of day the scene needed. A chase sequence that made sense on the page and fell apart once the DP tried to block it. A financier who could not picture the film from the script and quietly stopped answering emails.
Every one of those problems is a pre-visualisation problem, and pre-visualisation has historically been a luxury of productions that could afford a previs team. That is the part of filmmaking where generative video has made a genuine, unglamorous difference over the last year, and it is worth looking at closely because it is not where most of the online argument about AI in film is taking place.
What previs used to cost, and why indies skipped it
On a studio picture, previs is a department. Artists build rough 3D versions of key sequences so the director, the DP, the stunt coordinator and the VFX supervisor can agree on lenses, camera moves and timing before anyone is paid a day rate. It saves enormous amounts of money on the shoot because the expensive decisions have already been made.
Independent productions have always done a cheaper version: storyboards, a shot list, maybe a phone-video rehearsal in a car park with friends standing in for the cast. These work, but they share a weakness. They describe the shot without showing how it moves, and movement is exactly where plans fall apart. A storyboard frame of a slow push-in on a face says nothing about whether the move should take four seconds or twelve, or whether the background will read at all.
Generative video sits in the gap between the storyboard and the full previs department. It does not replace either. What it does is let a director turn a storyboard frame into a few seconds of motion, cheaply, and look at it.
The image-to-video habit
The workflow that has caught on among small productions is image-to-video rather than text-to-video. Text prompts produce footage that looks like a film, but not like your film. Feed the model a frame you already have instead, and the output stays tethered to your project.
That frame can be almost anything. A location scout photo shot at the right time of day. A storyboard panel. A concept painting the production designer knocked out over a weekend. A still pulled from a test shoot. The model animates from it, and suddenly the director can see whether the camera drifting left past the doorway actually reveals anything interesting, or whether it just looks like someone forgot to lock off the tripod.
Directors who use this describe it less as “making footage” and more as “testing an idea at the speed of conversation.” One editor working on a horror short described the change as the end of the meeting where everyone nods at a storyboard and imagines a different shot. Now the shot is on the screen, and the disagreement happens before the shoot instead of during it.
Where it actually helps on set
Blocking complicated movement. Anything involving a moving camera and a moving subject at the same time is where storyboards lie the most. A ten-second generated clip of a tracking shot through a market gives the DP and the grip something concrete to argue with.
Timing and pacing. Drop generated clips into an animatic with temp music and you learn whether a sequence breathes or drags. It is the cheapest way yet to find out that your opening montage is ninety seconds too long.
Light and time of day. Asking for the same frame at dawn, midday and blue hour is a fast way to settle a scheduling argument. It will not be photometrically accurate, but it is accurate enough to decide which call time to fight for.
The pitch. This is the use nobody likes to admit to and almost everybody relies on. A two-minute sizzle reel built from generated previs clips communicates tone to a financier far better than a lookbook full of reference stills from other people’s films.
Where it visibly does not help
It is worth being blunt about the limits, because they are consistent and they decide where the tool belongs.
Performance is off the table. Generated faces in close-up still feel wrong in ways an audience senses even when it cannot explain why. Nobody serious is using this to previs a dialogue scene for emotional content; at most it is used to check eyelines and coverage.
Continuity across shots is unreliable. Generating the same actor, in the same costume, in the same light from two angles remains hard, which is why most directors treat each generated clip as a standalone test rather than a sequence.
Physics is approximate. Water, fire, glass and fabric look roughly right and specifically wrong. That is fine for previs, where “roughly” is the job, and fatal for anything meant to end up in the final cut.
The working rule among the directors I spoke to is simple: generated video is for deciding, never for delivering.
The practical side: tools, cost and workflow
Most directors begin in a consumer app, and that is fine for an afternoon of experiments. It stops being fine the moment an assistant director wants forty variations of the same shot, named by scene and shot number, sitting in a shared folder before the production meeting. At that point the model has to be reached through an API rather than a chat window, usually with a small script or a no-code automation doing the repetitive work.
That changes how you budget. Video models are billed per second of output rather than per seat, and you will throw away most of what you generate, so the meaningful number is what a hundred discarded attempts cost rather than what one clip costs. The current generation of cinematic models also differs a lot in how faithfully it follows a reference frame, which matters more for previs than raw image quality. Productions comparing options tend to look at image-to-video fidelity, maximum clip length, cost per second, and whether the same endpoint lets them switch engines as better ones ship; the published rates for the Kling 3.0 API alongside competing video models are a reasonable place to calibrate what a week of previs will actually cost.
The most common advice from people who have done it on a real schedule: assume a ten-to-one ratio of generated to useful, never let a generated clip become the only version of a plan, and keep every test labelled as a test so nobody mistakes it for a promise.
What this means for the people who used to do the job
There is an honest cost here. Previs and storyboard artists on low-budget productions were often the first people hired and the cheapest line in the budget, and some of that work is being absorbed by directors generating clips themselves. That is a real loss for those artists, and it is mostly invisible in the excitement.
The counter-argument, which has some weight, is that most indie productions never hired a previs artist in the first place. For a film made on credit cards, the alternative to generated previs was not a professional; it was nothing, and a shoot day lost to a plan that did not work.
Both things are true. What is not really in dispute is that the planning stage of independent filmmaking has changed faster in the last year than the shooting stage has in a decade. The camera department looks much the same as it did. The production meeting does not.
The sensible position
If you are prepping a short or a micro-budget feature, the useful question is not whether AI video is good enough to be in your film. For anything that matters, it is not. The useful question is whether a few hours of generated tests could save you a shoot day, and for most productions with any camera movement at all, the answer is yes.
Use it to settle arguments early. Use it to show a financier the tone. Use it to find out that the beautiful oner you have been dreaming about does not actually show the audience anything. Then put it away and go and shoot the film.
