Free Resource
What AI Video Can and Cannot Do for an Ad Brief
A producer's map of where generative video is reliable, where it still breaks, how a studio works around each limit, and what that means for the brief.
By Unreal Media GroupEditorial standardsReviewed
Planning guide Written by the studio from its own practice; where it states a view, the view is the studio's. It cites no third-party research and attributes no performance figures to the studio.
See cinematic brand films and commercials the studio has made
Burger Build Wide · 15s
Most writing about generative video is either a vendor's reel or a critic's collection of failures, and neither helps a marketer decide whether their brief suits the method. This guide is a producer's map instead: what generation does reliably, what it still gets wrong, what each failure looks like on screen, how a studio works around it, and what a brief should say so the workaround is planned rather than improvised. It is written from Unreal's own practice producing AI video for brands and is honest about the limits because the review that decides delivery is built around them. It names no tools and ranks none; capabilities change between one review date and the next, and the limits below are described by what they do to a frame, not by which model produced it.
What's included
- The exact-versus-evocative test for every element of a brief
- Where generation is reliable today
- Seven limits, what each looks like, and the workaround
- The pre-delivery review built around the limits
- Definitions a reviewer and a producer can share
- A brief-readiness checklist
Start from the brief, not the reel
The useful question is not whether generative video is good. It is which elements of a particular brief have to be exactly true and which have to be evocative. A product's shape, label and finish have to be exactly true, because the viewer will compare them with the thing that arrives. A claim demonstrated on screen has to be exactly true, because the demonstration is the claim. A real founder or a real customer has to be exactly that person. The mood of a kitchen at dawn, the feel of a road at night, the weight of a fabric in motion, the scale of a landscape: those have to be right, not identical to any place that exists. Generation is strong on the second list and has to be managed on the first, and a brief that sorts its elements into the two lists has done most of the work of deciding the production method.
Where it is reliable
Environments and atmosphere: interiors, exteriors, weather, time of day, the look of a place that does not need to be a specific place. Camera language: moves, lenses, depth of field, the grammar of a commercial. Light: consistent, motivated, graded. Product as an object in space, from a supplied still, in short shots where the product does not have to be handled. Synthetic performers in short takes, disclosed as synthetic, where the brief wants a person who does not exist rather than one who does. Motion derived from an approved still frame, so that what moves is what was signed off. Abstract and graphic sequences, transitions and textures. In each of these the generated frame is judged on whether it is right for the brief, and the studio can iterate until it is.
Where it breaks, and how each break is handled
The failures are well known and specific. Each row names one, describes what it looks like so a reviewer can spot it, states how a studio handles it, and says what the brief should supply or decide so the handling is planned. None of these is a reason not to use generation; each is a reason to scope it correctly.
| What breaks | What it looks like on screen | How a studio handles it | What the brief should say |
|---|---|---|---|
| Product fidelity | Shape drifts between shots; a finish or colour shifts; a branded item becomes a generic one; a clasp, a cap or a seam is wrong | Generation from the brand's own product photography as reference; the real product composited into generated scenes for hero shots; a review of logo, label, proportion and material against the references before delivery | Supply the approved product photography and packaging files; name the details that must be exact |
| Text, logos and interfaces | Garbled type on packaging or screens; a logo that is nearly right; an app interface that does not exist | Text, logos and interfaces kept as a graphics layer added in the edit over the generated frame, never generated as pixels; screens and labels replaced in compositing | Supply vector logos, the exact copy for any on-screen text, and screen recordings or design files for any interface |
| Hands and physical interaction | Wrong finger counts; a grip that does not hold; a pour, a spread or a squeeze that behaves wrongly | Shot design that keeps hands out of the critical frame or holds them still; short takes cut around the interaction; a filmed insert for the one gesture the claim depends on | Say whether the product must be shown in use, and which use; that decision may make one shot a filmed element |
| Continuity across shots | The same person or room looks different from cut to cut; wardrobe, hair or props change | A character or environment sheet approved before generation; every shot derived from the same references; the edit built to tolerate small drift with cuts rather than long holds | Approve the character and environment sheets; decide how many distinct setups the piece needs |
| Physics and complex action | Liquids, fabric, collisions and fast movement behave implausibly; objects pass through one another | Action kept simple and brief; motion derived from approved stills; the difficult second of movement filmed or replaced with a cut | Flag the action beats; expect the studio to propose a filmed element or a simpler staging for the hardest one |
| Dialogue and lip sync in long takes | Mouth shapes drift from the words; expression flattens over a long take | Short takes with cutaways; performers written for brevity; captions carrying the script; a filmed presenter where a long, continuous piece to camera is the brief | Say whether the piece needs a continuous presenter; if it does, plan to film that person |
| Real people and real places | A likeness that is nearly someone; a location that is nearly somewhere | Not attempted. Real people are filmed, or licensed under a written rights agreement; recognisable real locations are filmed or replaced with an unnamed environment. Synthetic performers are disclosed as synthetic | Name every real person and every real place the piece must show |
The review that decides delivery
Because the limits are known, the review is built around them. Before delivery the studio checks each generated shot against the approved references for the product (logo, label, proportion, material and colour), checks every frame that carries text, checks hands and interactions in the frames where they appear, and checks continuity across the cut. Anything that cannot be resolved is documented so the brand knows what it is approving. This is the same review described in what is included in AI video production services and on how it works, and it is why the brief's reference material is not optional: a review against nothing approves anything.
- Reference fidelity
- How closely a generated product matches the brand's approved product photography and packaging files in shape, colour, finish, label and proportion. The measure the pre-delivery review applies.
- Continuity
- The same person, object or place staying the same across cuts. Managed in generation by deriving every shot from one approved set of references, and in the edit by cutting rather than holding.
- Generative artefact
- A visible error produced by generation rather than by the brief: an extra finger, garbled type, an object that passes through another. Found in review, resolved by regeneration, compositing or a filmed element, or documented.
- Compositing
- Combining separately produced layers into one frame: a filmed or photographed product over a generated environment, a graphics layer over a generated shot. How the elements that must be exact are kept exact.
- Graphics layer
- Text, logos, interfaces and end cards added in the edit as their own layer rather than generated as pixels in the frame, so they are always correct and always editable.
- Hybrid production
- A piece in which some elements are generated and some are filmed or supplied, each chosen for what it must be. The usual shape of a commercial brief rather than an exception.
The brief-readiness checklist
A brief is ready for generative production when it says which elements must be exactly true and which must be evocative; supplies approved product photography and packaging files for everything that must be exact; supplies vector logos and the final copy for any on-screen text; names every real person and real place, with the rights position for each; flags the action beats and the moments the product is handled; states whether a continuous presenter is required; approves character and environment sheets before generation begins; and names who reviews the references and who signs off. A brief that cannot yet answer those questions is not wrong; it is early, and the discovery call is where the answers are found.
Hybrid briefs are normal
The choice is rarely all-generated or all-filmed. A typical commercial brief has one or two elements that must be exact (the product in the hand, the founder to camera) inside a world that only has to be right, and the scope names the method for each: a filmed pack shot or a licensed performer composited into generated environments, the graphics as a layer, the rest generated to the art-direction document. What that changes at each stage of production, and which briefs still need a camera throughout, is set out in AI vs traditional commercial production.
What this guide does not tell you
It does not name or rank tools; the studio confirms suitable tools per project after checking access, licensing, rights and the brief, and the limits above are described by their effect on a frame rather than by a product name. It does not say that generation is always the right method; the brief decides. It does not promise a delivery time or a price; the written scope does. And it does not claim any result for any piece: Unreal publishes no performance figures and never guarantees advertising performance.
In practice
The service this guide supports: AI video production.
Frequently asked questions
Can AI video show my actual product accurately?
Only with the brand's approved product photography and packaging files as references, and even then hero shots often composite the real product into the generated scene. The studio reviews logo, label, proportion, material and colour against the references before delivery and documents anything that cannot be resolved.
Can AI video render text and logos?
Not reliably as pixels in the generated frame. Text, logos and interfaces are added as a graphics layer in the edit, from the brand's vector files and final copy, so they are always correct and remain editable.
Why do AI videos get hands wrong?
Hands and physical interactions (grips, pours, squeezes) are among the hardest things for generation to render consistently. A studio designs shots around them, cuts around the difficult second of movement, and films the one gesture a claim depends on.
Can the same AI character appear across a whole ad?
Within limits. Continuity is managed by approving a character sheet before generation, deriving every shot from it, keeping takes short and cutting rather than holding. A long continuous piece to camera is a brief that usually calls for a filmed presenter.
When does a brief still need a camera?
When it depends on a real person's identity, a recognisable real place, a demonstration where the photograph itself is the claim, a continuous presenter, or a physical interaction the product story cannot do without. Those elements are filmed or licensed and combined with generated frames; the scope names the method for each.
Start with one scoped batch.
Book a free 15-minute discovery call. We'll map the exact plan for your brand — no pitch deck, no fluff.