Use Generated Images in a Treatment Without Passing Them Off as Research
Use Generated Images in a Treatment Without Passing Them Off as Research
A generated frame can earn its place in a treatment the way a scout's photograph does — by answering a question someone actually has. The trouble starts when the question quietly changes and nobody notices. A synthetic image can propose a composition with complete conviction and, inside the same picture, imply that a product works a certain way. It has no obligation to be honest about which of those two things it is doing. You do.
So here is the method before the reasoning. Decide what the image is for. If its job is to propose a look — framing, palette, light, mood — it can do that as a declared proposition, and there is nothing wrong with it. If its job is to show how something behaves, in a real place, with a real product, a real hand or a real mechanism, it cannot do that job at all, however convincing it looks. Inspect the one relationship the scene depends on. Keep a record of how the image was made separate from your reading of what it shows. And when the open question is whether something can actually happen, go find observed material or run the smallest authorized test that would settle it.
One clarification saves a lot of argument later: a disclosure that the image was generated describes one piece of its history. It does not establish that the depicted place exists, that the mechanism operates, or that you have permission to use the image or anything visible inside it. Three separate questions, three separate processes.
Name the image's job before judging its beauty
Two jobs get folded together constantly. This is the composition we propose is one. This demonstrates how the material behaves is another. A synthetic frame can carry the first without touching the second, and it will look equally authoritative either way, because plausibility is not a property of the job — it is a property of the picture.
The workable habit is to write the intended use beside the image and then ask what an uninformed recipient would take it to prove. Somebody who has never seen your notes, who doesn't know where the file came from, opens the treatment at that page. What do they now believe that they didn't believe before? If the honest answer includes a fact about a real place or a real product, the caption is carrying more weight than the image can bear, or the image is doing work nobody assigned it.
This is where the word reference fails. A page headed "reference" can hold a photograph of an actual location and a generated frame of an imagined one with nothing to distinguish them. The category collapses a proposed shot and an observed example, and once they're in the same pile, whoever reads the treatment next will treat them the same way. Name the role instead: proposed framing, proposed palette, or unverified depiction of the bottle's opening action. It takes eleven more words and it prevents a scheduling meeting three weeks out.
The example I'll use for the rest of this article is invented, and I'll fix its facts now so nothing shifts underneath it. A team is pitching a spot for a 300 ml foaming hand soap bottle. The bottle has a pump collar that must be turned a quarter turn out of a locked position before the plunger will travel; while the collar is locked, the plunger does not move at all. The team has a generated frame — call it frame twelve. Everything below is a stipulated situation, not a report from a job.
Inspect the relationship the scene depends on
You are not auditing the image for every anomaly. You are looking for the one relationship whose failure would change the film. In a product scene that is usually one of four things: contact between a hand and an object, the geometry of a moving part, fidelity to a supplied design, or whether the space holds together from one look to the next. Pick the relationship the shot hinges on and examine that one properly instead of scanning the whole frame for six-fingered hands.
Frame twelve is beautiful in a way that makes this harder. Low angle, bottle on a wet stone ledge, window light from the left, amber liquid glowing in the glass. A hand enters from the left, holds the bottle at the shoulder, fingers wrapped around the body, thumb resting on top of the plunger. The collar sits on the far side of the bottle in shadow, and its two alignment marks are not legible at this scale. No second hand. Nothing in the frame shows the collar being turned, and nothing in the frame shows the result of a turn.
You could object that no still image can show a turn, and that is exactly the point. The frame is not failing to show movement; a still cannot. The frame is failing to show movement while presenting a pose that reads as complete. It is the completeness that does the damage — one hand, one grip, bottle already awake, the plunger sitting at rest in its ready position. A reader fills in the rest. The image never states that the operation is one-handed. It just doesn't leave room for anything else.
Which matters depending on what the spot does next. If the storyboard's opening beat is "she picks it up and pumps it in one continuous motion," then that grip has to survive the whole sequence from locked collar to soap in the palm, and frame twelve has not tested any part of that. That is the relationship the scene depends on. The reflection in the stone and the exact shade of amber are not.
Watch for atmosphere covering the gap. Wet stone, morning light through a window, liquid catching the sun — all of it signals care, and care reads as accuracy. A plausible mood can carry an impossible grip past three rounds of review. That is not a reason to distrust atmosphere; it is a reason to check the grip separately from the mood.
One more distinction, because it changes what you're allowed to do about it. A declared fantasy is fine. If the pitch has the bottle floating above the ledge, that is a stylistic proposition and the label says so. The quarter turn is different in kind, because it is not the team's invention. It is a constraint of a design someone else supplied. Getting it wrong is not a creative choice, it is an accidental claim about a client's product — the kind of claim a client is entitled to correct.
The good news is that the illustration often survives. Reframe the hand so it isn't doing anything the product can't do, crop the mechanism out, or move the composition to a beat where the collar's state isn't the point. What will not work is polishing. No amount of rendering quality adds an observation that was never made.
Keep creation history beside the interpretation
Three columns, written down, on the page where the image lives: what is generated, what is observed, and what is composited or still unverified. Those are not three shades of the same thing, and none of them is a statement about whether the product works.
Alongside them, keep the trail that lets a collaborator retrace the file inside your own authorized workflow — which inputs went in, what was changed, at what stage. That trail belongs to you. It is not the same as a public credential, and it is worth keeping whether or not one exists.
The C2PA 2.4 explainer covers this distinction directly. Its goals and non-goals, and the sections asking whether provenance is always complete and whether provenance can determine whether content depicts truth, all point at the same separation: a creation-history record describes how a file came to be, not whether the scene inside it is factual, feasible or cleared. The explainer also notes that provenance can be incomplete. Two practical consequences follow. First, do not upgrade a disclosure into a universal reliability badge — a visible label tells you something about one part of an image's lineage and leaves everything else exactly where it was. Second, treat the distinction as durable reasoning rather than a fact about any one tool's current badge. That reference is to one version of one document, C2PA 2.4, and not to the latest specification, and nothing here rests on having tested an implementation.
Concretely, here are two captions for the invented pitch. They do different jobs, and neither is called "reference."
Frame twelve — generated composition. Proposed framing, light and palette for the opening beat. The collar is a stand-in for the supplied design and its alignment marks are not visible at this angle. Whether this grip can unlock and operate the product is unverified. See the product photograph opposite.
Product photograph — client-supplied. Shows the collar, its two alignment marks and the locked position. It does not show an operating sequence, and nothing in it indicates how the bottle is opened.
Notice what the second caption refuses to claim. The photograph is real and attributable, and it still settles only geometry and mark placement. A still photograph of a bottle contains no motion, so it cannot demonstrate that a hand can unlock and pump in one movement. Observed material is better than synthetic material, but it is not automatically an answer; it is an answer to the question it actually addresses.
Labels are also not stickers applied once. If someone later composites the real bottle into frame twelve's background, the caption changes again: background generated, product composited from the client photograph, hand generated, contact between them unverified. The moment a composite is assembled, the reliable label is the one describing that composite, not the ones describing its pieces.
Everything in this section concerns truthfulness about origin and status. Permission is a separate matter with its own questions — whether client inputs were authorized to go where they went, whether a recognizable person's likeness is involved, what the terms of use allow. A provenance label does not resolve any of them, and this article does not need to name a generator to make that separation, because the separation holds for all of them.
Replace the image when the question requires evidence
Once the question is "can this actually happen," the image has stopped being the tool. You have two ways forward.
If the action already exists somewhere in the world, find an attributable observed example under conditions that resemble yours. Not a mood board. An example you can point at, with a source and a date, whose relevance you can defend. The client's own product photograph is a modest version of this; genuine footage of the mechanism on a real unit is a better one.
If the action is new, or no example can be found, define the smallest authorized experiment that would distinguish the plausible routes rather than the experiment that would produce the best footage.
For the bottle, there are three plausible routes, and they are genuinely different productions. One hand does both: the character unlocks the collar and pumps without the grip changing. Two motions: the grip in frame twelve survives, but only as the second half of a beat, so the storyboard needs a shot for the first half and the edit needs a place to put it. Or the bottle is prepared off camera between takes — unlocked by someone on set, then handed over or reset. That third route is legitimate filmmaking. It becomes a problem only when the treatment describes it as observed product behavior, which is the purest version of the mistake this article is about: a real thing happening on a real set, written up as though it were the product's own capability.
The smallest test for routes one and two is genuinely small. With the client's authorization, take a sample of the bottle, hold it the way frame twelve shows — fingers around the body, thumb on the plunger — and try to unlock the collar and complete the pump without regripping. A phone on a tripod is enough. One take answers the geometry question: whether the thumb can reach the collar and return to the plunger without the hand changing shape, and whether the wrist angle survives the turn.
The limits of that test matter, and they are easy to state. One tester's hand, a sample rather than the final tooling, no wardrobe, no lighting, no camera move, no edit. It tells you whether the pose is physically available; it does not validate the shot, the performance or the look. A rough test is a question-answering device, not production approval. Say which of the two it is every time you describe it.
When neither observed material nor a test is available — no example, no sample, no authorization to try — you still have two honest options. Keep the image as a conditional proposal, with the condition written down where the reader will see it: the bottle's opening action is undetermined, and nothing in this treatment should be read as a depiction of it. Or take the image out, if its persuasive force comes mainly from the part that can't be supported. What you cannot do is keep the image and leave the missing piece in someone's memory of a meeting.
Keep it, replace it, test it, or take it out
Four outcomes, and the choice follows from the image's job rather than from how good it looks.
Keep it when the job is composition, palette or mood, when it makes no claim about behavior, place or the supplied design, and when the labels say accurately what is generated, observed and unverified.
Replace it when the question is what a real place or a real product does and attributable observed material exists. The generated frame did its job during the pitch. It doesn't get to stay attached to a claim it can't hold.
Test it when the question is decidable, authorization exists, and the experiment is small enough that saying no to it costs more than running it. Write down what the test can and can't answer before you shoot it.
Take it out when the image's persuasive power is inseparable from the thing it can't establish. This is the hardest one, because it is often the best-looking frame on the page.
None of these decisions requires a verdict on whether synthetic imagery is good. A generated frame is a proposition, and propositions are useful. What goes wrong is narrower than the argument usually gets: the image gets asked a question it was never able to answer, and everyone agrees to forget which question that was. Generated is not a synonym for false, and it is not a synonym for verified. Keeping those two failures apart is most of the work.
Frequently asked questions
What is the basic method for using a generated image in a treatment?
Decide what the image is for. If its job is to propose a look—framing, palette, light, mood—it can be a declared proposition. If its job is to show how something behaves in a real place, with a real product, hand, or mechanism, it cannot do that job however convincing it looks. Inspect the one relationship the scene depends on, keep creation history separate from your interpretation, and when the open question is whether something can actually happen, find observed material or run the smallest authorized test.
Does calling an image generated make it accurate or cleared?
No. A disclosure that an image was generated describes one piece of its history. It does not establish that the depicted place exists, that the mechanism operates, or that you have permission to use the image or anything visible inside it. Those are three separate questions with separate processes.
How should I inspect a generated product image?
You are not auditing every anomaly. Find the one relationship whose failure would change the film—usually contact between a hand and an object, geometry of a moving part, fidelity to a supplied design, or whether the space holds together. In the invented bottle example, frame twelve shows one hand gripping the bottle with the thumb on the plunger and the collar in shadow, presenting a pose that reads as complete; if the storyboard needs a continuous unlock-and-pump, the frame has not tested that. Atmosphere can make an impossible grip pass review.
What labels or captions should accompany generated, observed, and composite images?
Keep three columns visible: what is generated, what is observed, and what is composited or still unverified. Name the role instead of using reference. A generated composition caption can state proposed framing, light, and palette, note that the collar is a stand-in, and say the grip is unverified. A client-supplied product photograph can show collar, marks, and locked position, but it does not show an operating sequence. If someone later composites the real bottle into the generated background, label the composite.
When should I keep, replace, test, or remove a generated image?
Keep it when the job is composition, palette, or mood, it makes no claim about behavior, place, or supplied design, and the labels accurately say what is generated, observed, and unverified. Replace it when the question is what a real place or product does and attributable observed material exists. Test it when the question is decidable, authorization exists, and the experiment is small enough that saying no costs more than running it. Take it out when its persuasive power is inseparable from what it cannot establish. A smallest test can be a phone on a tripod: with authorization, hold a sample bottle as frame twelve shows, try to unlock the collar and pump without regripping. Its limits are one tester’s hand, a sample rather than final tooling, no wardrobe, lighting, camera move, or edit; it tells whether the pose is physically available, not production approval.