Pitch a Commercial With Generated Footage, Not Just an Impressive Generated Clip
Pitch a Commercial With Generated Footage, Not Just an Impressive Generated Clip
A generated clip is very good at one thing: it shows that a look exists. That is not a small contribution to a pitch. A room that has been arguing about tone for two weeks will stop arguing when it sees forty seconds that feel like the right film. The sample did its job.
Then the client says: "Love it. Same thing, but the shelves become a window, and can the fan be the matte one?"
Now the room is quiet for a different reason. Nothing in that sample answers the request. It was never built to. It was built to demonstrate a possibility, and possibility is the cheapest thing in the room. What the treatment still owes the client is a route — a description of what survives when something changes.
That is the distinction the pitch has to make visible. Not "can generation look like this," but "which parts of this film are held steady, by what, and what happens on the day the background moves."
A sample proves mood, and mood is the easy part
The failure is quiet because it doesn't look like one. The generated sample is beautiful, the references are coherent, the deck reads well. Everything is about look, and look is exactly where generative video is most persuasive and least diagnostic. Approximating a mood is the first and best thing these systems do. Holding a product's grille identical across three setups, through a revision, is a different class of problem, and a mood reel is not evidence about it.
So a sample can be excellent support for one claim — this is the register we want — and worthless support for the claims the production actually rests on. When sample quality gets substituted for route description, the substitution is hard to catch, because both sound like confidence. The clip is confident. The treatment, in that spot, is just hopeful.
The fix is unglamorous. It starts with an inventory.
Name the relationships the film cannot casually change
Take an invented example to keep the categories concrete. A desk fan is the hero product of a short spot. Nothing about this example is observed from a real production; it is a construction for reasoning, and no footage described below has been generated anywhere.
Four things in that spot cannot drift.
Product form. The grille — blade count, spacing, the wire pattern — and the control layout. Suppose the fan has four buttons in a row and a dial to their right. That arrangement is what a product manager will freeze-frame on, and it also carries the claim: this is the product in the box. A grille with one fewer ring, or a control row that has become three buttons, is not a rendering slip. It is a different object being advertised.
Subject identity. If a person appears in more than one shot, their face, hair, and wardrobe are a continuity problem that spans every route you might pick. This is the requirement most likely to be waved away by a smooth-looking sample, because a single clip never has to prove it.
Geography. Where the fan sits on the desk, how far the notebook is from its base, which side the room's light comes from, where the camera stands relative to all of it. Geography is the thing that makes an edit read as a space rather than a sequence of pictures. Get the fan's position loose between two shots and the audience may not name the problem, but they will feel the film come apart. Light belongs in this category as direction rather than intensity: within a version, the source stays on the same side of the frame, so the fan's shadow and the notebook's fall the same way across a cut. Replacing the shelves behind the desk with a window is not slack. It is a change of source, it is a deliberate revision, and it reopens the lighting — which the routes below have to account for.
Action and timing. The head tilts, the oscillator starts, and the sweep carries across the cut. If the idea of the spot is that the fan cools the room without anyone noticing, then the relationship between the sweep and the actor's reaction is the idea. Timing isn't decoration on that; it's the load.
Everything else can move. The clutter behind the fan, the weather implied outside, how many pedestrians cross, the exact grade of the wall, whether the desk has one pencil cup or two — that's atmospheric slack, and a good treatment should spend some of it freely. Slack is how the wall takes the light, not which side the light comes from. But the treatment needs to say, in writing, which category each thing is in. That distinction is not visible in a deck that is all mood, and it is not recoverable after the shoot.
None of this is exciting to read. It is the part of the pitch that a sample cannot improve, a client cannot argue with, and a revision request will test immediately.
Test a sequence and a revision request, not a single clip
One clip proves an output. A sequence test proves a method, and it costs almost nothing to describe. Instead of a hero clip, put three connected shots and one demanded change on the table.
Back to the invented fan. Shot A is wide: the fan on the left third of the desk, notebook to the right, an actor entering frame. Shot B moves closer as the head tilts and the oscillator starts, control row legible. Shot C is the product view — front on, fan centered, sweep continuing.
Held steady across all three: the grille pattern, the four-button-and-dial layout, the fan's position relative to the notebook, the direction of the light, and the sweep — same arc, continuing through the cuts rather than restarting at each one. Free to move: the clutter behind, the grade of the wall, how many objects sit on the desk.
The revision request is the shelf-to-window swap.
Now watch what each route has to admit.
If the shot is generated whole, a second pass with a new background re-decides everything, including the fan. It may keep the grille. It may not. You will not know until you look, and you have to look at the entire action — not the first frame and the last frame, which can both be perfect while the middle grows a fifth button and invents a second hand.
If a controlled fan element is captured or built and the background is generated behind it, the revision touches the background only. The control row is whatever the object has, in every pass, because no model is being asked about it. That route can absorb the change more cheaply. Not for free, though — the window is a change of source, and a change of source reopens the lighting you were holding. The plastic housing's reflections were recorded against shelves, and now someone is relighting the fan, replacing those reflections, and cleaning edges around the grille. That is real labor on a real day, and it belongs in the treatment rather than in a line item the client discovers later.
And here is the sentence that keeps the pitch honest: a new attractive variation is not a successful revision. If the second pass is lovelier and the control row shifted, the revision failed the brief. Write the pass/fail list down before the test so nobody has to argue about taste afterward — the fan either kept its four buttons and its position, or it didn't. Keep the original input, the output, and the revision request together as a set, so that when a collaborator asks what actually changed, there is an answer rather than a recollection.
Know which controls actually coexist, and who does the work between them
A feature list is a menu, not a promise that everything on it can be ordered at once. Adobe's documentation for generating videos from images records first and last keyframe inputs and notes that adding image keyframes disables specified controls, with adding an end frame also disabling Motion options. That page, updated June 16, 2026, is documentation — not a generation test, and not a statement about quality, continuity, availability, or rights. But it is enough to make a point a pitch should absorb early.
Suppose the fan's sweep depends on keyframing the first and last frame and on motion controls to shape the arc. The documentation for at least one route already suggests those may not be available together. That is the kind of thing to discover while writing the treatment, not after it's approved. If your method leans on a combination, either demonstrate the combination or describe the fallback — because "we'll figure it out in the test" is a schedule risk wearing a confident expression.
Two habits follow. First, don't infer independently adjustable controls from a list of features. The specific combination matters, and it moves with the version. Second, when the test is eventually run, record the model, the feature, and the control combination that produced the result. A finding without a version attached is not a finding; it's a memory.
Then account for the work that happens between any two outputs. Somebody reviews the full action frame by frame. Somebody selects among takes. Somebody roto's the grille, rebuilds a patch of desk, composites the fan over the new background, and matches the grade across the cut. Somebody decides that take seven, not take eleven, is the one. None of that appears in a prompt, and all of it appears in the budget and the schedule. A prompt is an input to a production method. It is not the method.
Keep the other reviews separate while you're at it. Permission to use an input, the right to use an output, likeness, and provenance are four different questions with four different owners, and a route can work perfectly and still be unusable. Provider language about commercial safety is marketing until your own review says otherwise, and it says nothing about price, speed, or feasibility that you can act on.
Put generation where the film can survive a change
Three routes, compared on one question: what can each of them revise while preserving the relationships you listed?
Whole-shot generation buys the most freedom per shot and holds the least. Every revision reopens everything. It's a defensible choice when the subject has no external owner — no product team, no recognizable face — and when the brief is willing to accept a family resemblance rather than an identity.
A bounded generated element inside controlled material flips that. Generate the environment, capture or build the product. Revisions land on the element you generated, and the product is stable because it was never generated. You pay for it up front — the fan has to exist on set or in a build, and every generated background has to be integrated — but the cost is legible, and it doesn't move when the client changes their mind.
Recording or building the passage is the deterministic option. A practical fan, a real window, a motion-control pass. Revisions mean a reshoot or a rebuild, which is expensive and slow, and also exactly what the client asked for, with no argument about whether the grille drifted.
The useful asymmetry: the background is often the safest thing to generate, because nobody in the audience is checking the shelves. The product is the worst thing to generate — not because generation is bad at products, but because the product is the one element in the frame with an outside party whose job is to check it.
What the treatment still has to write
Generation can be asked to play a bounded role here: supply the variable environment, extend the space behind the action, produce the atmospheric variations that would be tedious and costly to shoot. That is a role a pitch can defend, because it names what is allowed to change.
The controls you can honestly claim are the ones you have run and documented — and if the test is still proposed rather than performed, that number is currently zero, which is a perfectly good thing to write down. The unresolved question is narrow and worth stating plainly: can a named model, at a named version, hold the fan and the timing through the requested change? That gets answered by running the sequence test and keeping the failures, not by liking the clip.
So let the sample stay in the deck. It earned its place. Just put next to it the sentence that says which parts of this film you are willing to lose an argument with. Generation can hold the weather. It should not be asked to hold the product.
Frequently asked questions
What can a generated sample prove in a pitch, and what can it not prove?
It can show that a look or register exists and can stop a room arguing about tone. It cannot show that product form, subject identity, geography, or action and timing will hold when something changes, because a mood reel was not built to answer those claims.
Which relationships should a treatment name as unable to drift, and what can move freely?
Product form (grille pattern, button layout), subject identity (face, hair, wardrobe across shots), geography (fan position relative to notebook, camera position, light direction), and action/timing (head tilt, oscillator sweep across a cut). Atmospheric slack - clutter, implied weather, wall grade, number of desk objects - can move if the treatment says which category each item is in.
Why test three connected shots and a demanded change instead of one hero clip?
One clip proves an output; a sequence proves a method. A three-shot test with a change such as shelf-to-window can show whether the grille, control row, fan position, light direction, and sweep survive, and whether the background revision reopens lighting or drifts the product. A new attractive variation is not a successful revision if the control row shifted.
What does the Adobe documentation cited in the piece say about keyframes and controls?
It records first and last keyframe inputs and notes that adding image keyframes disables specified controls, with adding an end frame also disabling Motion options. The page, updated June 16, 2026, is documentation, not a generation test, and not a statement about quality, continuity, availability, or rights.
How should a production document controls and findings?
Record the model, feature, and control combination that produced a result; a finding without a version attached is a memory, not a finding. Keep the original input, output, and revision request together as a set, and write the pass/fail list before the test so taste disputes do not replace a failed revision.