Most marketing teams try AI video the same way the first time. They type a sentence describing the ad they want and hope the model reads their mind. It almost never does. The output looks generic, the product color drifts, the logo is subtly wrong, and the clip feels like it belongs to no brand in particular. The problem isn’t the model — it’s the method. A reliable AI product video workflow starts with references, not a clever sentence. That’s what “reference-first” means, and for marketing teams, it separates a fun demo from a repeatable production process.
Why text-only prompts lose brand intent
A text prompt compresses everything you know about your brand into a few dozen words. But your brand isn’t a few dozen words. It’s a specific blue, a particular product silhouette, a house lighting style, the way your packaging catches light, a tone of motion that feels premium rather than frantic. Rely on text alone, and the model fills every gap you left unspecified with the statistical average of everything it has seen — which is, by definition, not your brand.
Reference-first flips this. Instead of describing your brand in prose, you show it. You hand the model your actual product images, your lighting direction, and examples of the motion and format you want, and you save text for the one thing references can’t capture: the objective of the clip. The more of your intent you carry in references, the less the model has to guess.
This has gotten easier as tools raise how much visual truth you can feed in per generation. ByteDance introduced the Seedance 2.5 product-video workflow at its Volcano Engine FORCE conference on June 23, 2026, and the release quadrupled the prior reference ceiling — from roughly 12 inputs up to 50 combined images, video, and audio references in a single project. That headroom is what makes reference-first practical instead of theoretical: you can pin down product, lighting, motion, and format together in one project rather than picking one and describing the rest in prose.
Build the reference pack before you write anything
Treat this as the first practical step of every project. A complete reference pack has five parts:
- Product references: clean, high-resolution images of the exact product from the angles you want on screen.
- Lighting references: one or two frames that show the mood — bright and airy, moody and premium, warm and domestic.
- Motion references: an example of the camera behavior you want, such as a slow push-in or an orbit.
- Format reference: the exact aspect ratio and safe areas for the platform you’re shipping to.
- One objective: a single sentence naming what this clip must achieve — “make the new cap design the hero,” not “make a great ad.”
In practice, teams rarely need every one of those fifty reference slots; what matters is that the ceiling is high enough that brand intent survives generation instead of getting averaged away. One caveat worth knowing before you build a habit around this: dumping in fifty unrelated images without structure tends to confuse the model rather than help it, so organize references into clear categories (product, lighting, motion, format) instead of a flat pile. Assemble the pack once per campaign and reuse it across every clip so your outputs stay consistent from ad to ad.
Write a shot brief, not a vague prompt
With references in place, the text you write becomes a shot brief — the kind of instruction you’d give a videographer, not a search engine. A good shot brief names the subject, the single action, the camera move, the lighting, and the end state. For example: “Hero the serum bottle, center frame. Slow 15-degree orbit left to right. Soft key from camera-left, subtle rim light. Hold on the cap for the final two seconds so a CTA can sit beside it.”
Notice that the brief already anticipates where the call to action will go. Planning for the CTA from the start avoids the common failure where a beautiful clip has no clean space left to place the offer.
Review for brand, motion, and CTA readability
Before anyone celebrates a first result, run it against three fixed questions:
- Brand: Do the color, product shape, and any visible branding stay faithful to the references from the first frame to the last?
- Motion: Does the movement feel smooth and on-brand, or does it wobble, warp, or turn frantic?
- CTA readability: Is there a stable, uncluttered area where your offer and button stay legible, including on a small phone screen?
If any answer is no, the clip isn’t finished — no matter how impressive the rest looks.
Run a controlled revision loop
Revisions should be surgical. Change one element of the reference pack or one line of the shot brief, regenerate, and compare. Because your references carry most of the work, a good reference-first setup lets you fix a single detail — tighten the orbit, warm the light, steady the final hold — without re-rolling the entire concept. Keep a simple log of what you changed and what it did. That way the whole team builds shared knowledge instead of each person rediscovering the same lessons on their own.
Reference-first at a glance
| Stage | Input you provide | What it protects |
|---|---|---|
| Reference pack | Product, lighting, motion, format images + one objective | Brand identity and visual consistency |
| Shot brief | Subject, action, camera, lighting, end state | Direction and CTA space |
| Review | Brand/motion / CTA checklist | Fitness to ship |
| Revision loop | One controlled change at a time + a change log | Speed and team learning |
For marketing teams, the payoff isn’t just prettier clips. It’s a process that’s documented, repeatable, and owned by the team rather than trapped in one person’s prompt intuition. Build the reference pack, write the shot brief, review against fixed questions, and revise one variable at a time — and AI video stops being a novelty and starts pulling its weight in the production line.
Related: AI Prompts to Create Mini Frameworks: 9 Proven Models for Smarter ChatGPT Results
