On Tuesday afternoon, a two person marketing team received a brief. By Thursday morning, it had a client ready first cut: a 45 second brand awareness video with three format variants for different placements. The inputs were a one page script, six reference frames from the brand's existing photo library, and a style note describing desired lighting energy. There was no camera, production company, or editor. This was one of the more compact turnarounds we saw among teams in our 2025 early access cohort. It is not typical for every project, and it shows what becomes possible when production is reorganized around a generative workflow and the specific inputs listed here.
This workflow has a defined structure. It is not simply a matter of writing a prompt and seeing the result. Teams that produce consistent quality follow a sequence of steps, while skipped steps are often what cause downstream trouble. This is what we learned from watching teams begin the process from scratch.
Making the brief a deliverable
Most marketing teams use the brief for internal alignment. In traditional production, the brief is written, presented to a production company, and interpreted by the director. Creative choices take place on set, while experienced crew members resolve gaps through judgment calls.
A generative workflow has no production crew to fill missing details. The model interprets the material it receives. The brief therefore must address questions traditional production answers in person: the lighting mood, pacing energy, visual palette, and emotional tone set by the opening 5 seconds.
We call this the script plus requirement. The raw script gives the model the sequence of events. Reference frames define the visual world, and style direction supplies the energy. All three inputs are needed for an intentional result instead of an algorithmically generic one. Teams that leave out any one of them consistently say the output misses their intent. They are correct, because that intent was not communicated.
Writing scripts for generative video
Teams new to this workflow most often write scene descriptions in production language instead of visual language. "Cut to product shot" instructs production. "Camera holds close on the product, warm side light, surface texture visible" describes something the model can interpret. This distinction matters because generative models respond to descriptions of the visual world, not directions for shooting it.
A useful structure for 30 to 60 second ad formats:
- Scene-by-scene visual descriptions of 2 to 5 sentences each, describing what the viewer sees and feels, not what the camera does
- Atmosphere and lighting notes at the scene level, not the overall level
- Motion energy descriptors for each scene: slow and held, dynamic and sweeping, intimate and close
- Transition notes only when a specific transition matters to the narrative
Voice over copy may go in the script or be added in post. Teams needing precise on screen text treatment or VO timing often generate the visual track first, then add VO as an overlay. Others include VO from the start when scene pacing must align with spoken words. Neither method is always better. The choice depends on whether pacing or visuals lead the creative concept.
Reference frames that guide results
Reference frames give the model its visual anchors. They define the output's color palette, lighting environment, texture quality, and overall visual register. Uploaded frames do not need to come from the campaign itself. They can be mood board images that set the visual language.
The strongest reference frame sets we have seen share two traits. First, the images are visually coherent, with similar color temperature, lighting quality, and surface texture. A set of 6 frames combining bright studio white backgrounds, dark moody cinematic shots, and warm golden hour natural light gives contradictory guidance. The model averages those directions, rarely producing a strong result.
Second, the frames focus on what matters to the campaign. When a brand's visual identity centers on warm earthy tones and natural texture, the set should strongly represent those qualities. When the campaign should feel faster and more energetic than the brand's usual style, the frames should clearly signal that change.
Six frames is generally enough for a strong first generation. Fewer than four makes the reference set too narrow. More than ten adds contradictions that undermine consistency. Frame quality matters more than quantity.
Inside the generation session
A 30 to 60 second ad generation session generally has two stages. First, the full script and reference set produce candidate directions, usually 3 to 5 variants exploring different interpretations of the inputs. This is the creative selection stage. Review the candidates to determine which direction has the right energy, pacing, and visual quality for further development.
The second stage develops chosen direction through targeted changes: different motion energy in one scene, a slightly warmer palette, or tighter pacing during the first 10 seconds. Refinement generation is faster than the initial candidate set, typically taking under 30 minutes per iteration.
Most teams reach a final quality version within 2 to 4 iterations after choosing a direction. The count rises when the initial script is imprecise or the reference frames lack consistency. Teams needing the most iterations usually change creative direction mid session instead of refining the selected one. Such pivots are possible, but they effectively restart selection and add time.
Reviewing without a director present
Traditional production reviews follow significant sunk cost: a shoot day has already been spent, so the discussion concerns making the most of what was captured. Generative reviews occur at zero marginal cost per iteration. That changes the review dynamic.
The practical result is that review sessions can be more exploratory. If someone thinks the pacing may be too fast, generate a slower version and compare it directly instead of debating from memory. Comparison variants cost little enough that "let me show you what I mean" becomes practical.
A clear decision maker remains necessary. Generative production increases the versions available for exploration, but it can also make final decisions harder when everyone has veto power and nobody has authority. We have seen teams produce excellent creative, then go through 12 internal review rounds because the approval process was undefined. Generation is fast. Organizational decision making is not always.
Format options from one session
Format flexibility is a structural advantage of generative production. One session can create multiple aspect ratio outputs: 16:9 for broadcast and pre roll, 9:16 for vertical social placements, and 1:1 for square mid feed. The creative direction remains consistent because all versions use the same inputs. Visual framing changes with the format.
Teams used to single format output often find multiple formats disorienting at first. They have never needed a process for deciding how to use 9:16 because 9:16 was absent from their production pipeline. Setting format decisions in advance prevents a team from having every variant but no agreement on where or when each one should be used.
Friction points we have observed
Script quality and reference frame coherence are the two most common friction points. Both are preventable, and both stem from treating inputs as administrative paperwork instead of the creative brief that determines everything downstream.
A subtler third issue is an incorrect expectation of what "first generation" means. The first candidate outputs from a session are starting points, not finished deliverables. Teams expecting a presentation ready cut may see rough directions and mistake that for a model quality problem. It is like expecting an editor's first rough cut to be final. Candidate generation starts creative selection; it does not end production.
Practice speeds up the workflow. A team's first session usually takes a full day for orientation, input preparation, generation, and initial review. By the third or fourth session, the same cycle typically takes under 4 hours. Nearly all of that improvement comes from better input preparation, not from learning the tool mechanics.
