A product can stay geometrically stable and still look the wrong size. Scale in AI video is the relationship between the product and a known reference, such as a hand, a table, a doorway, or another SKU. It is not the same problem as warp or flicker. A bottle can keep straight edges, a readable label, and a steady highlight, and still read as a travel vial or as a jug because the desk, the hand, or the neighboring pack is the wrong size beside it.
That distinction changes the fix. If the silhouette bends, the label melts, or the picture flickers, use the image-to-video failure fixes guide and change one symptom at a time. If the form holds and the object merely looks like the wrong product, you have a scale problem. More motion, a sharper model, or a longer prompt will not install a size cue the frame never contained.
This page is for marketers, ecommerce leads, brand designers, and photographers who already know which SKU they are showing. The job is to choose a scale anchor, write the prompt so the model does not hero-size the object, review frames against the still, and recognize when the still itself taught the wrong size. It is not a full product-photo pipeline. The wider path from a catalog photo to a short clip lives in product photo to AI video. Use this guide as the size gate inside that path, or inside any image-to-video test where a shopper could tell that the object is too big or too small.
Cliprise fits as the place to make or refine the still, animate that still, and compare image and video routes before you spend a wider batch of credits. It does not measure the product. No model on the model list should be treated as a source of true-to-life size. A useful workflow is stricter than the prompt: one anchor, one still that already shows the relationship, restrained motion that keeps the anchor in frame, then a side-by-side review.
If the still does not exist yet, build it in the AI image generator and stop when the still would survive as a photo next to a real reference. If the still is already approved, move it through the image-to-video AI generator with the anchor language intact. If you also need a text-only concept before any still exists, the AI video generator can explore a shot, but that exploration is not a size proof.
What believable scale means
Believable scale means a careful viewer can place the clip next to the reference still and accept the same size relationship. It is not a caliper match, and it is not a frozen silhouette. It means you would publish the clip beside the product page.
Judge four relationships. Product to anchor is the one that matters most: bottle to notebook, shoe to chair, carton to doorway. Product to frame is how much of the picture the object fills. A tight crop is honest only while the anchor is still readable. Product to sibling SKU is travel against family, phone against tablet, sample against full jar. If those heights collapse, two products became one. Product to the claim is the caption. Travel size, family size, or "fits in a pocket" has to agree with the picture. The model will not police that sentence.
A clip can be sharp, well lit, and free of warp, and still fail because the cleanser looks like a perfume on a doll desk. Readable framing and believable size are different jobs. First frame design asks whether the opening communicates. This page asks whether the object is the size the SKU actually is. You can frame a correctly sized bottle so it is easy to see without turning it into a giant.
Scale is not warp, flicker, or a bad hand
Keep the failure names separate or you will fix the wrong thing.
Warp is a geometry change. Edges bow, the pump leans, the bottle pinches, the box stops being a box. Flicker is a high-frequency redraw: boiling texture, crawling highlights, a label that shivers. A bad hand is a hand with the wrong anatomy: extra fingers, fused joints, a palm that does not connect. Those symptoms have their own retest in the failure guide. Reducing motion is often the first test there.
Scale can sit beside those problems, and it can appear alone. A perfect hand can still be twice as wide as it should be next to a pump. A bottle can stay rigid while it grows relative to the desk during a push-in. A label can stay sharp on a carton that is as tall as the door. If you only ask "is it clean?", you will approve a clip that merchandises the wrong size.
Color is a third gate. A bottle can be the right size and the wrong hue, or the right hue and the wrong size. When the shopper buys a shade, run the color review in AI video color consistency after the size relationship is acceptable. Do not let a pretty color pass a clip that looks like another format of the product.
Label melt, a swapped cap, or a new logo is an identity failure. A correct logo on a bottle that has become miniature is a scale failure. Note both if both show up, and name which one blocks publishing.
Choose the scale anchor before you prompt
A scale anchor is a reference the viewer already knows how to read, in a still you would accept as a photograph. It is not a caption and not a measurement typed into the prompt. An exact height in the prompt does not create a ruler. The picture needs an object whose size people already carry in their heads.
Pick one anchor for the shot. A second anchor helps only when it agrees with the first. A hand, a notebook, and a doorway that imply three different sizes make the review impossible. The coin row is in the table so teams stop reaching for it.
| Anchor | Use it when | It fails when | What to say in the prompt |
|---|---|---|---|
| Adult hand | The still already shows a hand holding, wearing, or resting beside the SKU | The hand is malformed, or it is anatomically fine but giant or tiny next to the product | Keep the hand the size it is in the still. No new grip |
| Coin | A real coin is already in the photograph | You add a coin in the prompt, or the coin changes size or design | Do not invent a coin. If one is already there, do not feature it |
| Known pack | A carton, refill, or sibling SKU already has an approved size relationship | The prompt averages two SKUs, or the pack was never sized | Keep the relationship in the still. Do not ask for a new format |
| Furniture | A desk, shelf, chair, floor, or doorway is fully readable | The furniture turns dollhouse-scale, or a push-in crops it out | Keep the furniture in frame. Keep the product the same size against it |
Why a coin is a bad idea to invent
A coin feels like a ruler because people think they know its size. In an AI clip it is a weak proof. Coins are small, so a slight redraw is a large relative error, and the model invents thickness and artwork that never belonged to a real coin. A coin that was not in the still is a new object hallucinated beside the product you were trying to protect. If a real photograph already contains one, leave it. Do not add "next to a coin for scale" to a still that has no coin.
A known carton, refill, or previous bottle is stronger, because shoppers already carry that size. Furniture is stronger for a speaker, a lamp, or a suitcase. A hand is stronger for anything worn or applied.
What not to use as a fake ruler
Skip invented rulers, arrows, and on-screen measurements. They warp, and they spend attention on fake type. Add a real spec in the edit only after the picture already looks the right size. A banana, a can, or a phone that is not in the still is another invented ruler.
Write the prompt so the model does not hero-size the object
Hero-sizing is the default pressure in product prompts. Words like macro, extreme close-up, cinematic hero, product fills the frame, luxury oversized, and dramatic push-in tell the model to spend the picture on the object and to drop the context that made the size believable. For a poster, that can be a choice. For a shelf product, a desk bottle, a shoe, or a device that must match a product page, it is how the SKU becomes a prop.
Write the prompt as a conservation instruction. You are asking the model to keep a relationship that the still already shows, and to add only a small change over time.
A practical prompt shape:
- Name the product once, in ordinary language, matching the still.
- Name the anchor in the same sentence.
- Say the relationship holds: same size as in the still, both objects fully visible.
- Name one small motion that does not crop the anchor: a slow hold, a slight living drift, a hand that stays still.
- Remove camera moves that travel toward the product until the anchor leaves.
- Say what you do not want, briefly: no macro, no fill-the-frame crop, no second product, no new prop.
Plain language beats fake precision. "The pump bottle stays about the height of the closed notebook beside it" is reviewable, because both objects are in the picture. A volume and a centimeter height in the prompt are not. Keep the real spec in the brief. Do not expect the clip to certify it.
Match the words to the shot. A shelf or desk product needs a medium view with the shelf gap, desk edge, or neighbor visible the whole time. A held product needs the hand to stay the size it is in the still, with no new grip. A large product needs the doorway, chair, or floor line in frame. A low angle is fair only when the anchor shares it. A camera that makes the product tower over a normal door is a scale fail. If you need surface texture, make a second shot and label it as a detail in the generation log. Do not let that detail become the size master.
"Giant product" and "oversized prop" are scale instructions. Use them only for a clip that is openly surreal. If a mute hook needs a tighter crop, cut it from a master that still contains the anchor. Do not animate that tight file as the next source.
A worked example: bottle on a desk
This example is illustrative. It is a teaching scene, not a measured test, not a lab result, and not a promise that any model will return these frames. Do not copy the objects as if they were a real SKU spec. Copy the decisions: one anchor, a still that already has the relationship, a prompt that refuses hero-sizing, and a review against that still.
The scene. An everyday pump bottle on a desk, not a sample and not a jug. The anchor is a closed hardcover notebook, fully in frame, with the desk edge visible. A book is a size people already know. No coin. No second bottle.
The still, before any video. Use a real photo or a generated still. The bottle stands on the desk, base planted, in a height relationship you can compare with the notebook. Both are fully visible in a medium view, not a macro of the pump. Build that still in the AI image generator if you do not have the photo. Reject a floating bottle, a notebook the size of a coaster, or a desk that has blurred into a gradient. If you would reject the still as a photo, do not animate it.
A prompt that hero-sizes, for contrast. "Macro cinematic hero of a luxury pump bottle, extreme close-up, product fills the frame, dramatic push-in, sparkling commercial, highly detailed." That prompt never mentions the notebook. It asks the model to crop toward texture and to treat the bottle as a poster. Even if the still was honest, this wording fights the still.
A prompt that protects the relationship. "The same pump bottle remains on the wooden desk beside the closed notebook. The bottle stays the same size relative to the notebook as in the still. Both the bottle and the notebook stay fully visible. Soft daylight, no new props. The camera holds a medium view. No push-in, no macro, no fill-the-frame crop." Run that as image-to-video from the approved still, not as a fresh text-to-video scene that may invent a new desk.
An illustrative pass. Frame one matches the still, and the last frame still shows the bottle and the notebook at the same height relationship. The pump geometry held. That is a size pass for this teaching scene, not a promise about another model or another SKU.
Illustrative fails. The bottle drops to perfume height beside a normal book, or grows into a jug while the notebook shrinks. The notebook leaves during a push, or the desk becomes a toy surface under a huge bottle. The bottle can stay clean and rigid. The relationship failed. If frame one already disagrees with the still, stop. If the hero wording caused the crop, retest once with the conservation prompt and the same still. If the still was already macro, redesign or reshoot.
Log the still, the anchor, the prompt version, and a reason that names the relationship in the AI video generation log, including the credits the screen showed. Do not turn that note into a price list.
Review frames against the still
Review at the size a shopper will see, and also at a size where you can compare edges. A desktop zoom can hide a dollhouse desk. A phone can make a slightly large bottle feel like a different format. Watch both.
Use a contact sheet, not a vibe check while the clip plays. Play once so you know whether the relationship changes over time. Then pause.
Frame one, the middle, and the last frame. Frame one is the inheritance test. If it is already the wrong size, the still was wrong or the model replaced the crop before the action started. Do not hope a later frame repairs it. The middle is the drift test. A push or a living camera can enlarge the product faster than the desk while the bottle itself stays rigid. The last frame is the exit test. If the anchor is gone, the clip will not hold scale on a loop or a product page.
Compare ratios, not moods: bottle to notebook, shoe to chair seat, carton to door, hand to jar. A ruler on the screen is only a consistency check between the still and the frame. It is not a physical spec. A lower camera makes vertical objects look taller. That is acceptable when the product and the anchor share the angle. It fails when the product's perspective and the room disagree, or when the product dwarfs an anchor that should contain it.
Do not repair scale in the grade. Color, contrast, and warmth can move in a grade. Relative size does not. Cropping in the edit can hide an anchor and make a correct master look hero-sized. If you crop, keep a version that shows the anchor, and do not let the tight crop become the only reference for the next generation.
If the shopper also buys a shade, mark color on the same sheet. Two gates, one sheet, two notes. A passing size does not pass a hue.
When the still itself taught the wrong size
Image-to-video animates the still you give it. If the still is a poster, the video is a moving poster. Teams blame the video model for a size error that was finished before the first credit on motion.
The still taught the wrong size when:
There is no reference object. Infinite white, a blurred studio, or a gradient background can be a legitimate pack shot. It does not establish size. The model then applies a prior that likes large, centered products. When you later ask for a room, the room arrives around a product that was never sized to that room.
The crop is already macro. A shelf product photographed as a texture fill has no shelf. The video cannot honestly restore a full shelf and a normal carton from a frame that only shows the front panel.
The hand is the wrong size. Fingers may be countable and still too large, so the jar looks like a sample. Or the hand is a delicate miniature, so a normal bottle looks like a prop jug. Anatomy can pass while scale fails. If the fingers are also wrong, stop and use the hand section of the failure guide before you debate size.
Two SKUs were normalized, or the furniture is toy scale. A travel bottle and a family bottle at one height have already lost the range. A tiny chair, a short door, or a table the size of a coaster is a room error even when the product geometry is clean.
The anchor exists only in the prompt. You upload a tight product cutout and type "on a real desk next to a notebook." The video pass now has to invent the desk and the book while clinging to a cutout that fills the square. Expect a fight. Put the anchor in the still first.
A concept pack was treated as a sized SKU. Packaging explorations are allowed to be loose. AI packaging concept video is the place for directions that are not yet a final dieline. Once a concept frame is dropped into a product gallery, people read it as the real size. Label concept frames as concepts. Do not animate them as proof of the finished pack.
How the error gets baked in
One path starts with a still that wins because it looks premium and simple. The background falls away. The bottle fills the square. That file becomes the source for a lifestyle clip, and the model preserves the oversized bottle while it invents a kitchen that is too small around it. Reviewers blame the motion. The size was already in the still.
A thumbnail crop causes the same problem. It was fine in a grid and wrong as a source, because it deleted the hand or the shelf. Animate the fuller photo. Before any video run, write what in the still tells a stranger the size. If the answer is nothing, do not generate yet.
Mistakes that make scale fail
A macro prompt on a shelf product. Cereal, cleaner, shoes, speakers, and devices that sit on a shelf need the shelf, or a stand-in the shopper understands. Macro language removes it. If you need the print texture, shoot a labeled detail. Do not make the detail the establishing shot.
No reference object at all. Without a hand, a known pack, or furniture, "the right size" is undefined. The clip can look polished and still be indefensible next to the product page. Add an anchor to the still, or accept that the file is a graphic and keep size claims out of the copy.
Mixing two SKUs in one prompt. "Show the travel bottle and the family bottle" without a still that already has the true height difference invites the model to make them match. One SKU per shot is the clean test. If the offer is a range, the still has to show the range first, and the prompt has to say the difference stays.
The follow-on mistakes. Fill-the-frame language, an invented coin or ruler, and a grade used as a size fix all leave the relationship untouched. A hook crop that wins the meeting becomes the next still, and the anchor is gone. Change one variable at a time. A pixel comparison to the still is not a fill volume or a fit, and "true to size" does not belong in the caption before the review.
When to stop and reshoot or redesign the still
Stop when more video runs cannot add information the still does not have.
Stop when the still has no anchor, when frame one already shows the wrong relationship, or when one conservation retest still resizes the product or crops the anchor away. A single retest is fair only if you can name the cause, usually the macro wording, and you keep the same still. Another route is later. Freeze the still and the prompt, change only the route, and record the name shown that day. Check models first, because availability moves. Stacking adjectives is not a test.
Reshoot when the physical product is in the room and the still is a guess. Put the real hand, carton, or table in a medium view where both objects are sharp enough to compare. Use that photograph as the image-to-video source.
Redesign the still when you are not on set. Return to the AI image generator, place the anchor fully in frame, and reject any still you would not publish as a photo. Then go back to image-to-video. Two formats need two stills. One clip should not invent the size gap.
Do not reshoot only because a concept is loose on purpose. A packaging exploration can stay unlabeled until the dieline exists. Mark it so it does not leak into the size-critical set. Do not keep a failing still because one middle frame looked right. The next ad will be made from the file you save.
A useful log line names the evidence: "Scale fail. Anchor left the frame," or "Scale fail. Still has no reference. Redesign before more video," or "Scale pass. Bottle-to-notebook relationship holds through the last frame." "Not realistic" does not tell the next person what to change.
Decisions in one place
| Situation | Do this | Do not do this |
|---|---|---|
| Shelf product, no size cue in the still | Add a shelf, desk, or known pack to a new still | Prompt a macro and hope the room appears |
| Need to prove size with a tiny prop | Use a hand, pack, or furniture that is already in the still | Invent a coin or a ruler in the prompt |
| Still is right, prompt says push-in and macro | Retest once with a hold and the anchor kept in frame | Rewrite the whole prompt and switch models together |
| Frame one already mismatches the still | Repair or replace the still | Generate more motion from that frame |
| Hand looks anatomically wrong | Follow the hand failure path | Argue about scale on a broken hand |
| Hand looks fine but the product looks miniature | Treat it as scale and resize the relationship in a new still | Sharpen the clip |
| Two formats in one offer | Approve a still that already shows the true difference | Let the model average their heights |
| Pass on shape, fail on size | Fail the clip for scale | Call it good because it did not warp |
Where Cliprise sits in the work
On Cliprise, create the still in the AI image generator, animate only an approved still in the image-to-video AI generator, and use the AI video generator when you are still searching for a shot and not yet claiming a size. Use models for a single-variable route comparison after the still is right. Copy the credit figure the screen shows into the log. Check current credit costs on the pricing page before a wider batch, and do not treat an older on-screen figure as a price list.
A person still decides that the notebook and the bottle agree. If the product is on the table and every generated still looks like a different format, reshoot.
Checklists
Before generation
- The format is named in one line, including what it is not.
- One anchor is chosen, and a coin is not being invented.
- The anchor is visible in the still you will upload.
- You would accept that still as a photo of the size.
- Macro, fill-the-frame, and push-in language is gone from the size master.
- Detail crops and concepts are labeled so they cannot be reused as masters.
- The prompt keeps both the product and the anchor fully in frame.
- Only one variable will change if you already have a parent run.
During review
- Frame one matches the still's relationship.
- The middle frame does not enlarge the product against the anchor.
- The last frame still contains the anchor.
- Phone and desktop both still read as the same format.
- Warp, flicker, and hands are noted separately if they appear.
- The pass or fail sentence names the anchor.
Before upload
- The master with the anchor is the source, not an unlabeled hook crop.
- Copy does not say true-to-life size or a measurement the picture cannot support.
- Detail shots and concept frames are labeled, and concepts stay out of the product gallery.
- The log has the still, the prompt version, the route, the scale result, and the credits shown for the run.
If more than one thing is wrong, order the work. Redesign a still with no anchor before you chase flicker. Repair a broken hand before you trust it as a ruler. Send a scale pass with a hue shift to the color review. A scale fail with beautiful motion is still a fail. Warp, flicker, and malformed hands stay in the failure guide. The path from catalog photo to export stays in the product-photo workflow. This page is the size gate.
Final next step
Pick one live SKU that shoppers misunderstand by size. Name the format. Choose a hand, a known pack, or furniture, and keep a coin out of the prompt. Approve one still that already shows the relationship.
Animate only that still, with the anchor kept in frame. Review the first, middle, and last frames against the still before you make the rest of the set. Create or repair the still in the AI image generator, then move a passing still through the image-to-video AI generator. Compare a route on models only when the still is already right and you can name the scale failure.
A stable shape can still be the wrong size. The still is where that mistake usually starts, and the frames are where you catch it.