A person behind the product is not scenery. Viewers read them as a customer, a user, or proof. Decide whether the clip needs a person before you generate.
That reading happens with the sound off. A figure at a shelf can look like someone who chose the product. A person behind a counter can look like staff. A face turned toward the lens can look like an endorsement. A crowd can look like demand. A glance, a uniform, a hand on the pack, or a busy aisle is enough. A background person can still imply a customer, an employee, or an endorsement when they were only meant to fill space.
This page is the set decision for a product clip: when a background person helps, when that person becomes a claim, what to do when faces melt, how hands on the product change the shot, why crowds are a separate risk, and how to brief an empty set.
Creator-led ads belong in the AI UGC video generator workflow. There the creator is the subject, the hook, or the voice. Label wording and placement belong in AI video disclosure labels. Open that guide after the picture is chosen, when a person, a hand, or a motion path will be published. Hiring and casting sit outside this job.
A real likeness needs permission you actually have for that person and that use. A place in the background does not lower the bar. A close imitation is the same problem as a named face. If that permission is missing, do not generate the person. This page does not explain how to impersonate a founder, a customer, a creator, or a public figure.
Cliprise is an AI creative platform at https://www.cliprise.app. The AI video generator turns a prompt, or a still plus a prompt, into a short clip. The image-to-video AI generator is the path when an approved product photo should control the frame. Decide the set, then generate. Each run spends credits. Check the current models list and pricing before a batch. A plan does not grant rights in a face, a uniform, or a picture that says someone bought the product.
Decide before you generate
Write one sentence before the first render: "A viewer can reasonably think this person is _____." Fill the blank with what the picture will support. Customer. User. Employee. Endorsement. A passerby who shares the room and does not interact. Or nobody, because the subject is the product.
If the honest blank is customer, employee, or endorsement, you need the fact behind it and, where a real person is implicated, permission you actually have. A small face is not that permission. "They would not mind" is not that permission. If you cannot support the blank, brief an empty set before you spend credits. Taking a stray shopper out later usually means a new still.
"Make it lifestyle" and "make it feel popular" hide a person and a claim in the same request. Replace them with one fact: human scale beside a lamp, depth between a bottle and the back wall, or a hand you will review as a performance. Energy, atmosphere, and social proof are not facts the model can honestly install.
Three checks sit on the sentence. If the pack, the shelf, or the room already explains the offer, refuse the optional person. The product photo to AI video path turns that still into a clip, and this page is the people gate on it. A mute viewer will trust a hand on the bottle over a caption that says "concept only." If a founder, a staff member, a buyer, a creator, or anyone a viewer could name is implicated, do not clear a missing permission by describing that person more precisely. That description is how imitation starts.
A child is not a background prop. Do not generate a real child, and do not invent a family using the product as proof that households buy it. A filmed plate that already contains a child is a rights question for the people who clear footage.
When a background person helps
A person helps only when they add one fact the product and the place cannot show, and a viewer can read that fact without taking it as a testimonial. If the face, the hand, or the crowd becomes the story, you have left this page.
Human scale is the cleanest fact. A lamp, a suitcase, or a speaker can look like a toy or a monument until a person stands at a known distance, fits the room, and ignores the camera. That is a size reference, in the same family as AI video product scale. If the still already has a door, a chair, or a known pack, use that anchor and skip the person. A shopper added to prove a bottle is bottle-sized becomes a customer. Depth is the next fact: a shoulder or a back at the far edge can separate a flat shelf from the rear wall, if the body stays off the product and does not stop, reach, or turn. A browsing path turns the depth cue into a shopper.
Cover the person with your thumb on a pause frame. If the product and the place still make sense, remove the person unless the one named fact disappears too. Write that fact in the generation log. A note that says "more human" needs a clearer frame, not a generated smile.
When that person becomes a claim
The same body becomes a claim when a viewer can treat it as evidence of a purchase, a job, a recommendation, or a result. No on-screen type is required.
A customer is someone offered as a buyer or a chooser: stopping, reaching, comparing packs, smiling at the label, carrying a bag, or waiting in a queue. A user is that claim in action. Pumping, pouring, spraying, wearing, opening, or applying counts with the face out of frame. If you did not film it, the clip should not look as if you did. A blurred figure who slows beside the shelf is still a shopper. Log the reject as an invented customer.
An employee is anyone the wardrobe or the position marks as staff. An apron, a lanyard, a badge, a branded shirt, a headset, a place behind the counter, or a restocking pose can do it with the back turned. A brand can look as if it employs that person. A retailer can look as if the aisle is its store. Generate either picture only when the role is true and the person is permitted. A generic associate invented to make a shelf feel staffed is a fictional employee offered as a picture of the business.
An endorsement is attention that recommends: a face to camera, an approving look at the product, a thumbs-up, a presenting pose, or a person the viewer already knows. Mute autoplay is where many of these clips are seen. A small background face is easy to skip in review and easy to recognize in public. If someone can pause and name the founder, or say the face looks like a creator you did not book, remove it. A prettier prompt is the wrong repair. A closer match to a real person is impersonation, and this page does not explain how to make one.
A real likeness needs permission you actually have for that person and that use. Founders, staff, customers, creators, and public figures count, including a background face you hope is only close. A store photo is not permission for a synthesized ad, and a license for one filmed spot is not permission for a new generated performance. Without that record, brief an empty aisle. A synthetic person can still read as a customer, an employee, or an endorsement. AI video disclosure labels can say the person was generated. A label does not create permission or a true testimonial.
If the person is there so skin, hair, or a body can look improved, you are showing a result. Use the before-and-after AI video workflow only for a change you can support. A label cannot rescue a caption about customers or a uniform that invents staff. If any person or hand remains, the label is a separate pass. This page stops at the set.
A decision table for the set
Use the table as a studio default before you generate. It is not a statute, a platform rule, or a price list. When a contract or a live policy is stricter, follow that. When two rows fit, follow the stricter default. A face in a crowd is both. A branded hand on a pump is a hand and an employee cue. An empty shelf with a person in the door glass is no longer empty.
| Set | What the viewer can read | Helps when | Becomes a claim when | Default |
|---|---|---|---|---|
| Empty set | The product and the place are the subject | The pack, shelf, or room already explains the shot | A later cut adds a person and keeps the empty approval | Start here. Put the absence in the still and the prompt |
| Distant body, face unreadable, no contact | Someone shares the place | You need depth or scale, and they do not stop or wear the brand | They pause, queue, point, or become recognizable | Optional. Check the start, middle, and end |
| Back or side, face out, off the product | A person is present and is not addressing the viewer | The body is only depth or scale | The path reads as shopping, restocking, or presenting | Keep the body off the product. Remove a storytelling pose |
| Readable face | A specific person, a user, or an endorsement | Rarely, in a product-first clip | The face is recognizable, looks at the lens or product, or melts | Do not generate a readable background face |
| Hand on the product | A user performing an action | The shot is a demo you will review as a performance | The hand was called scenery, or the product changes | One reviewed hand, or none |
| Crowd, queue, or busy aisle | Popularity, traffic, or demand | You are cutting a real plate you have the rights to use | The people are generated or stand in for proof | Do not invent a crowd |
| Real likeness or a close imitation | That person is involved | Permission you actually have covers this person and this use | Permission is missing, or assumed because the face is small | Do not generate them. Do not imitate them |
| Uniform, apron, badge, or counter | Staff, or a store that is yours | The role is true and the person is permitted | The outfit invents staff, a retailer, or a job | Treat branded clothing as a claim |
The default column is an instruction for the still and the prompt. Cropping an empty keeper does not allow a shopper to be painted back in. A uniform, or a quick person on the next run, needs a new inference sentence in the note. Reflections, posters, screens, and mannequins inherit the row of the person they show, so brief those surfaces with the people.
Faces that melt
A melting face is a failed take even when the bottle is sharp. Features slide. The eyes at the end would not pass as the eyes at the start. The mouth smears on a small turn. Glasses appear or vanish. Age shifts. Two shoppers share a face. A profile becomes someone else from the front. A smear at the shelf edge still reads as a person looking at the pack.
Pause on every face, including ones you hope are too small to count. Compare the first frame, a middle frame, and the last frame. If those three stills are not the same person, the symptom is identity drift. Retest it the way image-to-video failure fixes describes: name the symptom, keep the still and the prompt, change one variable. Less motion is often the first change. A new model in the same pass hides what helped.
Skip the retest in three cases. The still already contains the broken face, so repair the still. The face is stable and belongs to a real person you lack permission to use, so stability is not consent. The face is stable and the inference sentence says customer or endorsement, so clarity made the claim stronger. Reject that take because the claim got clearer.
Move the head out of frame only when the remaining body still passes the inference sentence. A reaching shopper is still a customer with the eyes cropped. Do not repair a melt by naming a real person or a lookalike. Fewer people, a turned-away body that is not shopping, a cleaned still, or an empty set are the allowed repairs. Glass, phones, posters, and screens invent faces after the set looked clear. A person who shows up only in a reflection, or only as the camera shifts, is still in the export.
Hands on the product
A hand stops being background when it touches the product, the shelf that holds it, or a tool that operates it. Viewers read that contact as use. A hand on a pump is a demo.
Sort contact before you judge the render. Beside the product, and not touching, a hand can be a scale anchor if the fingers are plausible and the size does not miniaturize the bottle. Extra fingers go to the failure-fixes guide. A clean hand at the wrong size goes to the product-scale guide. On the pack, the cap, or the shelf lip, the hand is a user even if it never presses, and a hand across the logo hides the product. Pumping, pouring, spraying, or applying is a performance. If you did not film it, mark it illustrative and do not offer it as measured proof. If the product foams or transforms, that is the before-and-after workflow plus a claims review by the people who own the claim.
A bare hand reads as a user. A uniform cuff, a marked glove, or a branded sleeve reads as staff with no face required. If that role is fiction, remove the costume.
Count digits at the start, the middle, and the end. Reject a sixth finger, a finger fused to the bottle, a second hand entering to help, or a grip that swaps hands halfway through. Prompt for one hand or for none. "Natural hands using the product" turns atmosphere into an unreviewed demo. If that hand is meant to be the creator, use the UGC workflow. The creator is the subject there.
When an empty still grows a hand, the model added a subject. Keep the still, strip lifestyle wording, and retest with less motion before a model switch. Log "hand entered from the left and rested on the cap," not "looks premium." If the hand survives, the delivery note says it is generated, whether it touches the product, and whether any product change is illustrative. Then classify it in the disclosure guide.
Crowds, queues, and busy aisles
A crowd stacks claims: popularity, foot traffic, a busy store, a line worth joining. A queue at the door, a row of heads on a shelf, or three people aimed at one SKU can say demand with no copy.
A melted crowd is an obvious reject. Cloned faces, fused legs, and people walking through the pack should never ship. A crisp crowd of smiling shoppers can pass a visual check and still be a false picture of demand. Ask what the crowd is allowed to prove before you ask whether the anatomy held. If you cannot pause on every face at the start, the middle, and the end, including glass and the back row, there are too many people. "Only look at the hero" means publishing the people you refused to see.
Watch density change. An aisle that fills, a line that grows, a duplicated person, or more heads whenever the camera moves is not a plate of a real afternoon. Go back to an empty set. Do not spend the next run on a more consistent crowd.
Notes often hide the crowd in a mood: energy, lived-in, Saturday, social proof, a store that looks busy. Translate the note with the client revision notes workflow, which turns comments into one change against a named parent. On an empty parent, one variable is light, distance, or duration. A crowd is a new shot. Refuse it when the inference is demand you did not film.
A filmed crowd can be a plate only when you actually have the rights for that footage and that use. A generated crowd made to resemble the photo is a different asset, and identifiable people in the reference need permission you actually have. Do not brief happy customers, influencers, or fans in a line. Use an empty aisle or one body you can check. Glass will multiply whoever you keep. If the route adds crowd noise nobody recorded, treat that sound as another person, and note in the log when the route has no audio.
How to brief an empty set
An empty set is a believable shelf, counter, table, or floor with a product on it and nobody who can be drafted as proof. Write the fields down so the next prompt version cannot grow a shopper.
- Product position. Where the pack sits, how much of it shows, and that it stays a pack. It does not rise, open, or perform.
- Place. The real surface, plus neighboring objects that are allowed, especially other packs that carry scale.
- Allowed motion. A locked frame or a very small move, and a light shift you will call illustrative. A push down the aisle discovers people who were "off camera."
- Absence. No people, shoppers, staff, hands, fingers, faces, silhouettes, crowds, mannequins, posters of people, or screens showing people.
- Reflective surfaces. Glass, mirrors, chrome, polished stone, phones, and televisions contain no person.
- Printed people on the approved pack. A face already on the real label stays artwork. Do not add a person beside it or replace that face. When the type melts, use the failure-fixes guide.
- Edges. Nobody enters from the left, the right, the top, or the foreground.
The still should already meet the list. Image-to-video animates what you approve, so a person or a person-shaped highlight in the still is an instruction to keep it. Clean the file, then move it through the image-to-video AI generator. The AI video generator can sketch an empty aisle from language alone. Keep that sketch out of the approval folder. When the SKU has to stay the SKU, rebuild from the photo.
Take lifestyle, busy, popular, candid, shoppers, customers, crowd, influencer, and energy out of an empty-set prompt. A model may still add a person. If it does, keep the still and change one thing, usually the amount of motion. A second failure means you fix shadows, cropped posters, and glass in the still instead of spending another near-duplicate run. Switch models later, with the still and the prompt frozen, only if that route is offered that day, and copy the name from the screen. Foam or other product motion still goes to the disclosure guide when a viewer could take it as a filmed demonstration.
Worked example: a retail shelf, illustrative
This walkthrough is illustrative. The soap, the filenames, and the notes are fictional. They are not a client result, not a measured outcome, and not a prediction of any model. The sequence is the point.
A team needs a vertical clip of a pump bottle of liquid soap they already sell. The approved photo shows the bottle on a retail shelf, neighboring bottles visible, label readable, no shopper. Nobody filmed a customer. No creator is booked. A client is waiting on a paid social concept. The brief asks for "a little life," meaning a shopper passing or reaching.
They write the inference sentence first. A person who looks or reaches reads as a customer, and they have neither the filmed event nor permission for a real shopper. Neighboring bottles already show scale, so a person would add a chooser. The row is the empty set. They animate the approved photo in the image-to-video AI generator instead of sketching the bottle from text.
The illustrative prompt on that still:
"The approved pump bottle stays centered on the retail shelf. Neighboring bottles stay in place. Soft store light. Very slight light shift. No people, no shoppers, no staff, no hands, no fingers, no faces, no reflections of people, no posters of people, no mannequins. Nobody enters the frame."
They do not ask the model to paint a disclosure line, and they do not add "premium lifestyle." The AI video generation log gets the prompt version, still filename, the route label shown that day, duration, aspect, and the credits the screen showed.
Run one is a reject. The bottle holds. Near the end a blurred figure enters from the left and slows at the shelf. The face never resolves. The path is still a shopper noticing the product. Reason: invented person, customer inference, final second. They do not rewrite the prompt and switch models together.
Run two changes one variable: less motion, same still, same prompt. A hand comes in along the lower edge and rests on the shelf lip. The fingers are plausible. The claim is not. Contact with the display is a user, and the hand was not in the photo. Reason: invented hand on the shelf. Cropping it "later" is refused, because the client file is the one that gets recut.
They look at the still again. Endcap plastic holds a person-shaped highlight they had called a fixture reflection. They clean it, save a new still version, and change only the still. Run three keeps the bottle readable, the neighbors stable, and the glass, edges, and last frame empty of people and hands. The log says: empty set, no person, product readable, motion limited to a light shift. That keeper is a set decision. It is not a traffic claim and not a legal sign-off.
A stakeholder asks, "Can someone grab it so it feels real?" The client revision notes workflow has to translate that before the keeper moves. A grab is a new shot: a hand on the product, its own inference sentence, its own finger count, and an explicit note that nobody filmed the grab. A creator talking to the lens starts in the UGC workflow. Soap changing the look of a hand starts in the before-and-after workflow. None of those is a one-variable edit on shelf-empty-03. The empty parent stays empty.
They also refuse "a busier aisle, like a Saturday." A crowd would read as demand, and there is no permitted plate. The note is a production record, not a legal opinion: empty shelf, approved photo, illustrative light, no customer, employee, or endorser. Moving light can still look filmed, so the disclosure guide classifies the file before it leaves the team. A founder, an invented apron, a queue, a child, or a poster of a buyer would reopen the inference sentence, so none of them is added.
Mistakes that turn a background into a claim
Adding a person because the pack feels static. Stillness can be an honest product picture.
Calling a browsing path atmosphere because the face is blurred.
Using a small face, or a closer description, in place of permission you actually have. This page does not explain how to match a real person.
Treating a uniform, a badge, or a place behind the counter as neutral costume.
Publishing a melted face because the bottle is sharp.
Logging a hand on the product as scenery. Contact is use, and a product that changes is a result.
Generating a crowd, including a cleaner one, so the aisle looks busy.
Pairing a disclosure label with a caption that still says customers, fans, or staff.
Changing the prompt, the still, and the model in the same run after a person appears.
Ignoring reflections, posters, screens, and the last second.
Using a background body as the improved after state in a result clip.
Letting a note about energy, Saturday, or social proof add people to an empty parent.
A review pass you can repeat
Before the render, read the inference sentence aloud and confirm the table row. Any real likeness needs permission you actually have. An empty row whose still contains a person-shaped reflection is the wrong still.
After the render, watch once with sound and once muted, near phone size. Pause at the start, the middle, and the end. Name every person, including reflections, posters, screens, mannequins, and bodies cut by the edge. If any inference comes back as customer, employee, or endorsement and you cannot support it, the take fails. Drift between the three pauses fails a face. A face that appears only at the end fails an empty brief. Count fingers, note contact, and note whether the product changes. If you cannot look at every face in a crowd, density alone is a reject.
Log filename, prompt version, still version, the route as shown, duration, aspect, the credits shown, and a reason that names the set. "Empty, no person, label readable" can be used again. "Nice" cannot. If any person or hand remains, or if empty-set motion looks filmed, the disclosure guide is next. Stop when a cleaned still and lower motion both still invent a person.
What to do next
Write the inference sentence for the next product clip before you generate. If the answer is a customer, an employee, or an endorsement, and you lack the fact or the permission you actually need, brief the empty set and clean the still, including glass and posters. Animate that still in the image-to-video AI generator.
Use the AI video generator for a rough sketch only when no approved photo exists, and keep that sketch out of the client folder. Log rejects where a person wandered in. If a person or hand you can support remains, review it at three points in time and send the label to the disclosure guide. A creator belongs in the UGC workflow. A transformed person belongs in the before-and-after workflow. Generate in Cliprise after the set is decided, and compare one short image-to-video take with the empty still before a wider batch of credits. Read the current models list and the pricing page for what a run costs now.
Related Articles
- AI Video Disclosure Labels for Social, Ads, and Client Files
- AI UGC Video Generator Workflow: Create Safer Ad Concepts
- Product Photo to AI Video: E-commerce Workflow for Cliprise
- Image-to-Video Failures: A Symptom-to-Fix Guide
- Turn Client Notes Into One AI Video Change
- AI Video Generation Log: Keep Tests Comparable
- Before-and-After AI Video Workflow: Show Change Without Fake Proof
- AI Video Product Scale: Keep Product Size Believable
- AI Video Generator
- Image-to-Video AI Generator