We tested one 30 second product brief in two generation modes: one session used 8 reference frames from the brand's existing visual archive, while the other used the same script with no reference frames and relied entirely on text. The results looked like different products. A creative director immediately recognized the reference frame output as belonging to the brand. The text only version looked generically professional: competent, well lit, appropriate, but not distinctly tied to anything. For a new brand creating a visual identity from scratch, that second result is fine. For a brand with an established visual language, it leaves value unused.
The central question is whether the project needs to belong to an existing visual world or remain visually open.
How reference frames affect the generation pipeline
Reference frames are not templates for the model to copy. They condition the range of possible outputs by shaping its probability distribution. From the set, the model reads a visual signature: color temperature, lighting direction, surface texture quality, and overall contrast ratio. It then favors visual choices that match that signature.
The frames do not have to come from the campaign itself, show the product, or even belong to the same category. They must be visually coherent with the world you want the output to occupy. Teams have produced strong results from editorial photography, architecture shoots, and film stills because those images expressed the visual quality they wanted.
Reference frames do not control the script's narrative structure, set scene level motion behavior, or guarantee particular compositional elements. The frames establish visual atmosphere, while the script supplies narrative logic. Strong output needs both, and neither can fully make up for weakness in the other.
When reference frames raise output quality
Reference frames reliably improve results in three situations. First, when a brand has a distinctive visual identity that text alone cannot convey. Precise color palettes, texture preferences, and lighting character are difficult to describe well enough for consistent reproduction. Showing the model the intended look is faster and more accurate than explaining it.
Second, they help when a project belongs to a series requiring consistency across multiple pieces. Producing 6 campaign ads across a quarter means making them look as if they came from one shoot. Using the same reference set in each generation session creates that consistency, while separate text only runs drift toward generic defaults.
Third, they help when the brief requires an atmosphere defined in another medium. If brand photography established a warm, soft, natural light aesthetic, provide those examples rather than trying to describe it in text, allowing the model to learn the signature directly.
When text alone works better
Text only generation fits two specific situations. First, it suits exploration beyond the brand's previous territory. If a campaign concept intentionally breaks from the existing visual identity, references that express that identity will pull the result back toward the familiar. Text only generation allows more exploration because reference conditioning removes some of the model's creative latitude.
Second, it fits projects without a visual archive to use. New products, new brands, and one off campaign concepts may have no material matching the intended direction. Mismatched frames, such as a competitor's visual language or references with the wrong tone, can hurt results more than a carefully written text description. The model balances reference conditioning with text input, so contradictory references can make the output incoherent.
We are not saying text only generation is automatically inferior. For creative exploration and brand agnostic campaign concepts, it can produce more interesting results than a reference conditioned session. Quality comparison matters only when brand consistency is required.
Video series visual consistency
For teams making ongoing video content, the question changes from whether to use frames to how to manage the reference set over time. A set built from a photo shoot in 2023 will pull results toward that period's brand aesthetic. As the brand changes, the reference set must change too.
Our practical recommendation is to keep a named reference set for each major phase of the brand's visual identity. After updating brand photography or visual identity, create a new set from that material and use it for all new campaign generation. Do not combine old and new frames, because the model averages them, and the result often resembles neither aesthetic.
This is why we built the style profile system: a named reference set paired with script parameters that can load into any generation session. Teams producing high volumes of content benefit from an agreed style profile representing the current brand language, preventing sessions from diverging based on who assembled the frames that day.
The hybrid method: references plus focused text
Experienced teams get the strongest results by using frames for atmosphere and lighting signature, then precise text for motion character and pacing. Frames communicate visual quality better, while text communicates sequence and motion energy better. Assigning each input its proper role directs the result more specifically than either input alone.
For example, 6 warm editorial frames can establish the lighting and color world. The script can add scene level motion directions: "opens slowly, camera holds still, then transitions quickly to the product close up, held tight for 4 seconds." The frames provide atmosphere, the text provides motion intent, and together they create output that fits the brand while pacing the story for the placement.
A calibration exercise for new teams
When establishing a workflow, run the same 15 second script twice: once with your strongest 6 reference frames and once with a text only description of the visual world. Compare the results directly. This reveals more about how the model responds to your particular reference material than general advice can. Teams that do this early in onboarding usually gain a clearer sense of when their reference set helps and when it works against them.
The aim is not to choose one method permanently. It is to match each project type with the right method, making the choice automatic instead of a debate at every session's start.
