How to make on-brand AI video ads that keep your character consistent.

A prompt can describe a person, product, logo, lighting, setting, and camera movement all at once, but the model has to reinvent each of them on every run. When the same details must recur across a campaign, keep them as separate saved references: the character, the product, the logo, and the location or style that should stay recognizable.

References don’t make every frame identical. Models read them differently, and each model accepts a different set of inputs. Treat them as a reusable starting point, and inspect the result before it becomes a customer-facing ad.

Collect the images you are allowed to use: a clean product photo, the logo in a legible form, and a character reference if a recurring spokesperson matters. Add a location, prop, or style reference only when it contributes to the shot. Label each asset clearly so whoever writes the prompt can tell the product from the logo and the presenter.

Write down what must not change, such as the package color, product silhouette, logo placement, character identity, or visual tone. Then write down what can change: the room, camera movement, background props, or time of day. A short list of fixed and flexible elements is easier to review against than a prompt with several competing instructions.

Build one clear shot first. Name the subject, action, product, setting, and composition, and mention only the saved references that should affect the result. For example: “Vertical product ad. @Maya holds @Serum beside the bathroom mirror in soft morning light. Keep the bottle label facing camera.” The @mentions point Studio at saved references instead of asking the model to infer everything from words.

Review the image at the size and crop you plan to publish. Check that the product shape and label look right, the character is the intended one, the logo is present and readable if required, and there is room for any captions you add later. If a key brand element is wrong, fix the still now. A video started from a wrong image carries the mistake into every frame.

Use the approved image as the first frame instead of downloading and uploading another copy. If the workflow extracts a character from that image, save the character to the Brand Kit and use the saved reference when staging the next scene. Later shots can then change the setting and action while keeping the same person.

Keep the next prompt focused on motion: what moves, what the presenter does, and how the camera behaves. Don’t restate every detail from the first image unless it needs to change. When the ending composition matters, choose a model and settings that support an end frame, because video models don’t all accept the same references or controls.

Put approved dialogue in quotes and include the intended language. Read it aloud before generating, and check names that are easy to mispronounce, punctuation, offer terms, and numbers. A short line is easier to deliver clearly than a paragraph, especially in a vertical social video.

Choose a model that supports sound for a talking video, then compare the generated speech with the approved line. Agenda Panda sends the spoken dialogue word for word, but delivery can still vary from run to run. Check pronunciation, lip movement, product visibility, and the opening and ending frames before scheduling.

Watch the full clip with sound. Can a viewer identify the product, understand the offer, and hear the line without extra context? Check that the brand mark stays legible and that the crop works on the target network. Fix any factual or visual problem before the post goes on the calendar.

Keep a short approval record: which product reference and script you used, who reviewed the clip, and what changed after review. You can repeat a strong format from that record, knowing the next run of the same prompt can still come out differently.