How Filmmakers Test a Film Before They Shoot It

A brand buying AI video today can watch a moving draft of its ad within days, sometimes before anyone has agreed on a script. That speed is useful, and it raises an old question: is the draft a plan the team can test and change, or a finished-looking picture everyone quietly accepts?

Cinema has been answering that question for close to a century. The most expensive moment in any production is the day the crew, the cast and the gear are all in one place, so filmmakers learned to watch their film before they made it: on paper, with borrowed footage, in rough 3D. Generated video is the newest tool in that line. The logic holds. Make the mistakes where they are cheap.

A wall of drawings at Disney

Webb Smith joined the Disney studio as a story man in 1931. D23, Disney’s official fan club, credits him with the idea of the storyboard, used to plan an animated film instead of a script. By the early 1930s, story sketches were pinned in order on large boards, and the artists found it much easier to see a story that way than to read it.

A script tells you what happens. A wall of drawings shows what the audience will see, in order, and where the story sags. Moving a drawing costs a pin. Moving an animated scene costs weeks. D23 notes that filmmakers planning commercials and live-action sequences adopted the method widely.

In generated video, the storyboard frame has become literal input. Many AI video models can take a still image as the first frame of a shot, and some accept a last frame too, then fill in the motion between. The wall still matters more than any one frame. A model generates each shot on its own and remembers nothing from the previous one, so continuity exists only if the frames already share it. Pinned side by side, they show where the bottle changes shape or the light jumps from morning to dusk, while the fix still costs a redraw instead of a batch of new clips.

Hitchcock’s paper film

No director is more associated with planning than Alfred Hitchcock. Writing in Print magazine in 1977, the biographer Donald Spoto described his method: after the script, an illustrator drew a working storyboard, a detailed sketch of each shot. Hitchcock liked to say his films were “finished before shooting begins.”

The legend needs a correction. A 2009 exhibition at the Deutsche Kinemathek in Berlin showed camera angle diagrams, continuity memos, sketches and storyboards, some drawn by Hitchcock, others by his cinematographers. The plan was a tool the team argued over and refined.

A generated animatic deserves the same treatment. Stringing rough clips into a timed draft costs little, which is its strength and its trap: when another take is minutes away, teams keep regenerating instead of deciding. Hitchcock’s memos worked the other way, writing down angles and continuity so the set would not reopen them. On an AI project, the equivalent is a review where the team argues over the animatic, then locks shot order, durations, framing and the moment the product appears. Anything the model added on its own, a background, a gesture, a drifting camera, gets kept or cut on purpose.

Lucas and the borrowed dogfights

Star Wars had a problem that paper could not solve. Its space battles depended on effects that did not exist yet, and in a dogfight, speed and rhythm matter more than composition. According to Cory Graff, writing for Air & Space / Smithsonian, George Lucas recorded hours of wartime newsreel and movie footage, dropped them into his cut in place of the missing space-fighter shots. Friends who saw a rough cut in early 1977 watched World War II planes where X-wings would go.

The footage also briefed the effects team. Artist Paul Huston recalled being shown a Japanese Zero passing an aircraft carrier and told: “The aircraft carrier is the Death Star, the Zero is an X-wing.” One borrowed shot answered questions about pace, framing and direction that a written brief could not.

Generated video has its own version of Lucas’s trick. Some models accept a reference clip and borrow its camera move or its movement, usually called motion reference or video-to-video. Ten seconds of phone footage of a hand turning a product, or an existing shot with the right push-in, tells a model more about pace and direction than a paragraph of prompt. The 1977 rule still applies. The reference briefs the shot and stays out of the final, especially when the brand does not own it.

Bad 3D, on purpose

By the prequels, borrowed footage had become rough 3D animation. For Episode I, David Dozoretz’s team built more than 45 minutes of 3D previs before a frame was shot, fxguide reported in 2002. For Episode III, Daniel Gregoire’s team produced more than 1,700 original previs shots in the months after March 2003. Gregoire summed up the purpose: “to sell a concept and to save time and money.”

When previs artists formed their own society in 2009, they defined the work as generating “preliminary versions of shots or sequences.” fxguide called previs “the very best, bad 3D”: no texturing, no lighting, no motion blur. The roughness was useful. Nobody fell in love with it, so everyone judged the decisions: length, geography, which shots to cut before a set was built. Variety’s report on the society stressed that previs is a process, useful from planning through post.

Testing the film first: verified milestones

From pinned drawings to generated drafts

Previs milestones

Date	Milestone
Early 1930s	Story sketches pinned to boards at Disney, credited to Webb Smith
Early 1977	Star Wars rough cut uses World War II footage in place of missing effects shots
2002	fxguide reports more than 45 minutes of 3D previs made for Episode I before shooting
2003	Episode III previs team passes 1,700 original shots, begun March 2003
2009	Previsualization Society announced
2024	Deloitte reports studios experimenting with generative AI for early visualization of character and set design

Generated previs reverses that bargain. It is faster and cheaper than a 3D team, and it looks finished: real textures, real light, plausible faces. That polish invites approval of choices nobody made. It also drifts. Inside one clip, a label can warp or a hand can gain a finger; between clips, a face or a product changes because each generation starts fresh.

Large studios are mostly keeping it at the previs stage. Deloitte wrote in 2024 that studios were experimenting with generative AI for “early visualization” of character and set design, while predicting that less than 3% of the biggest studios’ 2025 production budgets would go to these tools for content creation. Netflix’s 2025 guidelines for production partners, as reported by Cineuropa, allow generated material for non-final uses such as mood boards, reference images and temporary mock-ups, and keep it out of final deliverables unless approved.

The same test, now for a 30-second ad

For most of its history, moving previs belonged to big-budget features. A brand shooting a 30-second spot worked from a storyboard, then learned on set whether the timing worked.

Now the moving draft takes days. Frames become clips, and clips become a timed cut with pace, camera moves and transitions, before anyone books a crew or commits to finished generation. The draft also settles which shots should be generated at all. A product’s exact label, texture and scale are often safer filmed and composited over a generated background. The same review sets practical rules: the frame rate every clip will be conformed to, and which shots need upscaling for large screens.

What previs never did was guarantee the final film. Performances, light and accidents on set change things, and in generated work, so does the next generation.

What to ask of any AI video project

Ask to see the whole film move before the budget is spent: a rough, timed version of every shot, not one hero clip. Ask for the key frames side by side, and check that product, light and wardrobe match across them. Agree on what is locked after that review and what the model is still allowed to vary. Ask which shots will be generated and which filmed, what reference footage was used and who owns it. Then judge the final against the plan and ask what changed and why.