Image-to-video failures get fixed faster when you name the symptom and change one variable. A warped product, a drifting face, flicker, bad hands, a melted label, a camera that ignores the still, or a lighting jump each points to a different next test. Keep the same still and the same prompt until that one change has a clear result. Another full rewrite, stacked on a new model, spends credits and hides what actually helped.
This page is a diagnosis and retest guide for clips that already failed. It is not a prompt-writing manual, not a still-image artifact catalog, and not a map of the whole production pipeline. If the wording of motion is the open question, use the image to video prompt guide after the symptom is stable enough to test language on purpose. If the still itself is already broken, repair that file with the common AI image artifacts workflow before you animate it again.
Cliprise can help you run the retest in one image-to-video workflow: same still, a controlled prompt, and a model you can name in your notes. Check current models and pricing before you batch repairs, because availability and credit cost can change.
What you are diagnosing
An image-to-video model animates a still. The still is the identity of the shot. The prompt is a request for change over time. The failure is whatever makes that clip unsafe to publish, even if parts of it look impressive.
Creators often treat every bad clip as a prompt problem. That habit produces long prompts and similar failures. A diagnostic pass separates four sources:
- The still already contains the defect, and motion makes it obvious.
- The still is acceptable, and the requested motion breaks it.
- The prompt asks for two changes that fight each other.
- The model redraws details over time even when the request is modest.
You only need the dominant source for the next run. Secondary issues get a later pass. If you fix lighting, crop, model, and hand pose together, you will not know which spend to repeat on the next product.
Write the dominant symptom in plain language before you open settings again. Useful labels sound like this: "The bottle shoulder bends after the push begins." "The face matches at frame one and becomes a different person by the end." "The label letters swim while the cap stays put." "The camera orbits even though the prompt asked for a locked frame." "The room shifts from daylight to warm tungsten halfway through."
Vague labels waste the retest. "It looks AI" is not a symptom. "The highlight on the glass pops every few frames" is a symptom.
Symptom-to-fix decision table
Use this table before you rewrite anything. Pick the row that matches what you can see when you pause. If two rows match, choose the one that would make a buyer, client, or viewer reject the clip first. A melted logo outranks a slightly soft background. A warped product silhouette outranks a pleasant camera move.
| Symptom | What you see when you pause and scrub | Likely cause in the animation | First variable to change | Abandon the still if |
|---|---|---|---|---|
| Warped product | Edges bend, caps become uneven, handles multiply, the silhouette pumps in scale | Motion shear, a camera move across a thin object, or a prompt that makes the product perform | Remove the camera move, or shorten duration, and keep the still | Frame one already has the wrong silhouette, cap, or proportions |
| Identity drift | Eyes, jaw, age, ears, glasses, or hairline change across the clip | A new angle or expression, a long duration, or a still whose likeness is already weak | Reduce head and face motion, then shorten only if the drift is late | Frame one would not pass as the same person |
| Flicker | Texture boils, edges crawl, highlights pop, or the background shimmers | Noisy detail, sparkle language, or a surface the model redraws every few frames | Delete shimmer language and reduce background motion | The still is so noisy or patterned that every locked test still boils |
| Bad hands | Finger count changes, a grip fuses into the object, or a gesture melts | A new hand action, occlusion, or hands that were already unclear | Remove the hand action from the prompt | Hands are the subject and the still cannot show a believable grip |
| Label melt | Letters wave, a logo morphs, a barcode crawls, type leaves the package plane | Parallax across type, a moving label, or a prompt that treats packaging as alive | Lock the camera and keep the label plane still | Exact text must stay readable and a locked short test still melts it |
| Camera ignores the still | A new angle, crop, dolly, or orbit appears that the still did not set up | Conflicting camera words, vague cinematic wording, or a shot the framing cannot support | Delete camera instructions and rerun once | The composition you need is a different crop, angle, or subject size |
| Lighting jump | Exposure, color temperature, or shadow direction changes and then holds | Mixed light in the still, or time-of-day words that imply a second grade | Keep one light description and remove time-of-day conflicts | The still already contains two light directions you must preserve |
The first variable is a starting test, not a promise. Models differ. A shorter clip may hide a late warp without curing it. A locked camera may hold a label and kill the only motion the placement needed. Score the symptom, then decide whether that tradeoff is acceptable.
Read the "Abandon the still" column as a stop rule, not as a failure of effort. Another run on a still that is already wrong mostly teaches the same lesson at a higher credit cost.
How to watch a failed clip
Play the clip once at the size you would actually ship. A full-screen review can make small texture shimmer look catastrophic, and a tiny preview can hide a bent bottle. Then scrub. Diagnosis happens on paused frames more than on the vibe of playback.
Check four moments:
- Frame one. It should still be the still. If identity, product shape, crop, or light is already different, you are not looking at a motion problem yet.
- The first moment of motion. This is where camera disobedience and early shear often show up.
- The middle. Drift, label swim, and lighting jumps often become obvious here.
- The last few frames. Many clips collapse at the end: a face softens, a hand merges, a package twists, or exposure lifts.
Pause on the risk area and compare it with the source still beside it. For a product, look at the outer silhouette, then openings, then the label, then materials. For a person, look at the eyes and hairline, then ears and jaw, then hands. For a scene, look at horizon, shadow direction, and whether the room gained or lost windows.
Name one dominant symptom out loud or in a note. If you cannot, you are not ready to rerun. You are still browsing.
Also note what held. A bottle that keeps its cap while the label melts is a different repair from a bottle whose whole geometry pumps. A face that holds while the background flickers may be publishable for a social teaser and unacceptable for a beauty ad. The job decides the bar. The symptom decides the test.
Warped products
Product warp is a change in the object's geometry over time. The bottle shoulder bends. A phone's corners stop being square. A shoe sole waves. A mug handle grows a second curve. A box face becomes a trapezoid that no longer matches the front panel. The product can also pulse, growing and shrinking as if it were breathing.
Confirm it on the silhouette before you blame the label. Cover the type in your mind and trace the outer edge at frame one, the middle, and the end. If the edge moves in a way a rigid object would not, you have warp. If the edge holds and only the print moves, skip to label melt.
Warp often starts when the prompt asks a rigid object to travel, spin, pour, unfold, or "come alive," or when the camera trucks across a thin form. Push-ins are not automatically safe. A push toward a round cap can turn the ellipse into a wobble. Orbits are a frequent break on handles, straps, and spouts because the hidden side has to be invented.
The first retest for a rigid product is usually to take the camera move away, or to cut duration if the warp only appears late. Leave the product description alone on that pass. You are testing whether motion shear is the break. If the silhouette holds after that change, keep the calmer motion and only then decide whether the placement still has enough life. Steam, a shadow, a slow background shift, or a tiny push may be enough. Those are later tests, one at a time.
If frame one is already warped, stop. Animation will rarely repair a bad silhouette, and it often makes the error move, which is more noticeable than a still defect. Go back to the source photo or generate a cleaner product still. The product photo to AI video workflow is the place to rebuild that handoff. Still-level geometry, text, and edge problems belong in the image artifact fixes guide, not in another video prompt.
Soft goods are a special case. Fabric, foam, food, and hair are allowed to move. The failure is unmotivated geometry: a sleeve that changes seam count, a pastry that gains layers, a cushion that loses its piping. For those, name the allowed motion ("the sleeve shifts slightly") and delete extra action. Do not ask the garment to dance and expect the pattern to survive.
A practical example: a skincare bottle looks correct as a still. The clip uses a slow orbit plus "cinematic reflections." By the middle, the shoulder is oval and the pump leans. Pass one removes the orbit and keeps the same prompt otherwise, including duration and model. If the shoulder holds, the orbit was the break. If the shoulder still bends on a locked frame, the next single change is duration or a tighter crop of the same still, not a new slogan about premium glass. If a locked short clip still bends the pump, abandon that still or repair the product render before you continue.
Identity drift
Identity drift means the person in later frames would not be accepted as the same person who appears in the still. Expression change is narrower. A blink, a slight smile, or hair moving in a breeze can stay on-model. Drift changes structure: eyelid shape, eye spacing, nose, jaw, age, facial hair, earrings, glasses, or the hairline.
Scrub identity in this order. Frame one against the still. Then a frame one second in. Then the middle. Then the end. Early drift means the model left the still immediately. Late drift means the clip may contain a usable opening if you shorten it, which is a valid one-variable test. Do not call a shortened clip a likeness win if the job needs the full duration. In that case, revert the duration cut and reduce face motion instead.
Prompts that ask for a new emotion, a head turn, a look to camera, or a different angle are common drift triggers. So are long clips of a three-quarter face, because the far side of the head has to be sustained. Beauty stills with heavy retouching can also drift, because the face is already close to a synthetic average. A reference that merely looks "similar" is not the same as a still of the specific person. If likeness is the job, the still has to carry that likeness before video starts. The image reference upload guide covers how reference images support consistency. This page only tells you when the moving result has left that reference.
First retest: remove head turns, expression changes, and any line that asks for a new mood. Keep the camera as it was, unless the camera itself is an orbit around the head. If the camera is the thing moving the face into a new angle, the camera instruction is the one variable, not the expression line. Pick the change that matches the moment drift begins.
Glasses, hats, and earrings are useful tells. If the frames change shape while the background holds, you are watching identity or accessory redraw, not a lighting jump. If the whole grade shifts and the face structure holds, go to the lighting section.
Abandon the still when frame one is already a different person, when you need a specific real individual and this still only resembles them, or when two restrained tests still drift. "Restrained" here means you already removed the extra face action and you did not stack a model change into that same run. A new portrait, generated or photographed with a clearer angle and fewer overlapping props, is often cheaper than five more near-copies. You can build that replacement with the AI image generator, then return to image-to-video only after the still passes a likeness check at full size.
Do not promise that any setting locks a face. Review every frame you might pause on, including the end.
Flicker
Flicker is rapid, repeated change. Texture boils on skin, walls, or fabric. A highlight on metal pops on and off. An edge crawls. A background pattern shimmers. Leaves or city lights strobe even though you did not ask for strobe. The subject can stay recognizable while the surface looks nervous.
Separate flicker from a lighting jump. Flicker does not settle. A lighting jump changes state and then holds: daylight becomes tungsten, the exposure lifts, a shadow switches sides. If you pause three times and the light direction is different each time but stable between pauses, call it a jump. If the same cheek highlight is present, gone, and present again within a short stretch, call it flicker.
Flicker often comes from detail the model cannot keep stable. Fine grain, heavy noise, tiny tiles, dense foliage, glitter, rain, and "sparkling" or "shimmering" prompt language are frequent contributors. A dirty or over-sharpened still can boil even with a locked camera. Compression artifacts in the uploaded still can also crawl once they are animated.
The first retest is to remove the words that ask for high-frequency light or texture, and to stop background motion if that was requested. That should be one theme of change: "delete the flicker invitation." Do not also switch models in that run. If the prompt never asked for shimmer and the still is noisy, the next single test is a cleaner export of the same composition, not a new scene. Mild denoise or a cleaner source file is a still repair. Inventing a new background is a new still.
What can stay: a slow, broad change such as a soft shadow or a gentle push, once flicker is under control. What often makes flicker worse: adding more detail words in the hope of "stabilizing texture." Extra micro-detail gives the model more to redraw.
Placement matters. A slight background boil may be acceptable in a small organic post. It is often unacceptable on a beauty face, a jewelry highlight, or a UI screen. Screens are their own flicker case. If text or interface elements must stay readable, treat swimming UI as label melt, not as harmless atmosphere.
Abandon the still when a locked, short test on a clean file still boils the hero surface, especially on repetitive patterns, heavy grain, or a texture that is the product. A simpler background or a cleaner product crop is a design change. The first frame design guide helps you rebuild that opening so the hero has space and the busy texture is not the whole picture.
Bad hands
Hands fail in image-to-video in two different ways, and the fix depends on which one you have.
The still is already wrong. Extra fingers, fused fingers, or a grip that merges with the cup are present before motion. Video usually animates the error. Pause on frame one. If the hand would fail a still review, do not buy another video run to "let it settle." Repair or replace the still. Finger-level still repair is covered in the common AI image artifacts guide. This page starts once the hand's motion is the new problem.
The still is acceptable and the clip breaks the hand. A finger count changes, a thumb slides through a handle, a wave melts into the cheek, or a hand that was in a pocket crawls out because the prompt asked for a gesture. That is a motion failure. The first variable is the hand action. Delete "waves," "points," "holds up," "gestures," or "grabs" and rerun. If the prompt did not ask the hands to move, and they move anyway, the next single test is to shorten the clip or to crop the hands out of frame if the job allows. Cropping is a new still file in practice, so log it as a still change, not as a prompt tweak.
Occlusion makes hands worse. Hands around a phone, a strap, a collar, or another person's hand ask the model to maintain several edges at once. If the hero is the face or the product, a crop that removes the hands is often the most reliable repair. If the hero is the hand, a clearer still with a simple grip and an uncluttered background is the repair. Do not ask that hand to perform a multi-step action in the same clip that must keep the grip honest.
Watch the contact point, not only the finger count. A hand can have five fingers and still pass through a mug handle. That is a failure. Pause where skin meets the object. If the edge doubles or the object bends to follow the fingers, treat it as warp of the object plus a hand failure, and pick the one a viewer would notice first. Usually that is the broken grip.
Abandon the still when you need a specific gesture the current pose cannot support, when tools or jewelry make the fingers unreadable, or when a no-action retest still melts the hands. Generate a new pose. Then animate with almost no hand motion until that pose survives.
Label melt
Label melt is type or branding that will not sit still on its surface. Letters thicken and thin. A word gains or loses a character. A logo mark morphs. A baseline waves. A barcode or QR-looking pattern crawls. The package shape may hold while the print behaves like liquid. That split matters. If the package shape is also bending, you may need to fix warp first, because a moving surface will take the type with it.
Sharp source type does not guarantee stable type in motion. Image-to-video redraws frames. Small letterforms are easy to redraw wrong. A camera move that creates parallax across a curved label is a common trigger. So is any prompt line that treats the label as a living graphic: pulsing type, animated packaging, glowing letters, or a logo that "reacts."
First retest: one change aimed at the label plane. If a camera move slides across the type, remove that move. If the camera was already locked and the prompt asked the package to spin or the label to shine, remove that action or shine line. Do not do both in one run unless you have already decided to abandon the experiment and start a clean baseline. A clean baseline is allowed. Call it a reset, record it, and then change one variable again after you see the reset.
Read the result at a pause, at the size of a product page or a paid placement, not only in playback. A label can look fine while moving and fail when a viewer pauses. If the brand or the legal line must be exact, a human has to read it. There is no setting on this page that guarantees spelling, trademarks, or claims.
Some jobs should not animate the type at all. You can keep a calm product move and plan to add or restore critical type in editing after the picture is approved. That is a production choice, not a prompt trick. If the only message is the words on the package, and a locked short test still melts them, abandon animation of that label. Use the still, or build a composition where the critical words are not depending on generated frames. For a full product handoff, stay with the product photo to AI video path and treat this page as the failure check inside it.
UI text, captions burned into the still, and signs in a street scene fail in the same family as labels. If the words are the message, prefer a still that leaves room for real type, which is a first-frame decision, and keep generated motion away from that region.
A camera that ignores the still
This failure is not "the camera move is a bit strong." It is a shot change. The still is a front-on product on a table. The clip dollies to a three-quarter view the photo never showed. The still is a waist-up portrait. The clip zooms out and invents a room. The prompt says the frame is locked. The clip orbits. The crop jumps so the subject is suddenly smaller, or the horizon tilts.
Confirm it by matching frame one to the still's crop and angle. Then watch the first second. If the viewpoint leaves the still's geometry, the camera is disobeying the source. If the viewpoint stays and only the subject bends, you are in the warp or identity section.
Causes often sit in the prompt even when the picture is strong. Several camera verbs in one line fight each other: push, orbit, handheld, and aerial in the same breath. Vague words such as cinematic, dynamic, or epic invite a new shot. A requested move can also be impossible inside the still. A large orbit needs the sides of the object. A crane-up needs headroom and a floor. If the still is cropped tight, the model may invent the missing set.
The first retest is blunt. Delete camera instructions and run again with the same still, duration, and model. You are testing whether the model will hold the still when you stop asking for a shot. If this baseline holds, add one modest motion later, and take the wording for that motion from the image to video prompt guide instead of stacking adjectives. If the baseline still invents a new angle, the next single test can be a model change with the prompt frozen, or a still whose framing already is the shot you want. Do not do both.
Aspect ratio can contribute. A vertical still forced into a wide generation, or the reverse, asks the model to invent image that was not in the frame. Match the output shape to the still when the still's composition is the point. If you need a different aspect, design that crop as a still first. That work belongs with first frame design: subject size, headroom, and what the opening must already contain.
Abandon the still when the shot you actually need is a different angle or a different subject size. No amount of "keep the original framing" language will reliably film the back of a package from a front photo. Create the angle you need, or switch the concept to text-to-video if you do not have that still. The AI video generator is the path when you are no longer animating a locked frame.
Lighting jumps
A lighting jump is a new illumination state that arrives and stays for a while. The clip starts in cool window light and becomes warm interior light. Exposure pumps up and plateaus. A shadow that fell to the left falls to the right. The background grade shifts while the subject is lit by a second, unrelated source. Skin tone moves from the still's color into a different cast.
Do not file this under flicker unless the light strobes. A jump is a continuity error. It reads as a cut even when there is no cut. On a product, it can also shift perceived color, which matters for anything a shopper might compare with a real SKU. Color continuity across a finished piece is a wider topic. Use the AI video color consistency guide when you are matching several approved shots. Use this section when one image-to-video clip will not keep the still's light.
Check the still before you blame the prompt. Two practical lights in one photo, a flash on the face plus a sunset behind, or a composited subject with a mismatched shadow will often animate into a jump. If you can already see two shadow directions in the still, fix the still. Motion did not invent that conflict. It revealed it.
If the still's light is coherent, look at the prompt for a second time of day, a second color temperature, or an environment line that implies a different grade. "Golden hour," "neon," "studio white," and "moody tungsten" in the same request are a conflict. The first retest keeps a single light description that matches the still and deletes the rest. Do not add a new style sentence to "stabilize" the grade. Adding style is how jumps get written in.
Environment motion can drag the light with it. A prompt that moves clouds, turns practical lamps on, or sends a sunset across the sky is a lighting change by design. If you did not want that design, remove it as the one variable. If you did want a motivated shift, judge it as a creative choice only when the subject remains readable and the product color does not wander. An unmotivated shift is a failure even if each frame is pretty.
Abandon the still when you must preserve two conflicting lights, or when a single-light, low-motion retest still regrades the hero. A cleaner plate, a consistent shadow, and a first frame that already has the grade you want will outperform another adjective.
Symptoms that masquerade as each other
A wrong label on the symptom sends the retest to the wrong variable. Use these splits when the clip feels generally broken.
Warp versus label melt. Trace the outer contour. If the contour holds, stay on the label. If the contour moves, fix geometry before you judge spelling.
Identity drift versus lighting. Structure first, color second. A warmer face with the same eyes is a grade issue. A new eye shape in the same grade is drift.
Flicker versus jump. Rapid repeat versus a new state that holds.
Camera disobedience versus warp. If the viewpoint changes, you cannot fairly judge shape, because you are seeing a side the still may not contain. Lock the view, then judge the product.
Hand failure versus object warp. Look at the contact point. If the object bends only where the hand touches it, you still have a hand-led failure. If the object bends on the far side with no contact, treat it as product warp.
End-frame collapse versus a global failure. If the first two thirds are usable and the last moment melts, a shorter duration is a fair first test. If the break is immediate, duration is the wrong first lever.
Background swim versus subject failure. A morphing backdrop behind a stable product may be acceptable for a draft and unacceptable for a clean commercial. Name the subject as dominant when the subject is the thing you sell or the person you must recognize. Name the background as dominant when the swim pulls attention from a stable hero.
Write the dominant symptom, then the first runner-up. The runner-up is what you watch so a "fix" does not secretly trade one rejection for another.
The one-variable retest protocol
A one-variable retest means the next clip differs from the failed clip in one deliberate way. Same still file. Same prompt text, except the single edit you are testing. Same model, unless the model is the variable. Same duration, unless duration is the variable. Same crop, unless crop is the variable.
Change variables in an order that protects your budget and your ability to learn:
- Motion amount or duration. Use this when the break appears after motion starts, or only near the end. Shorten, or remove the largest move. Do not do both.
- Camera instruction. Use this when the viewpoint leaves the still, or when the move crosses a fragile edge, label, or face. Deleting camera language is one change.
- One conflicting phrase. Use this when a second action, mood, light, or "alive" instruction explains the symptom. Delete that phrase only.
- Crop of the same scene. Use this when the hero is fine and the failure lives at the edge of frame, in the hands, or in a busy background. Save it as a new still file and say so in the log. It is a still change.
- Model, with everything else frozen. Use this only after a restrained test has failed, or when the baseline already ignored the still. One comparison, not a tour of every option. Check the current list on models before you assume a given model is the right repair.
- A new or repaired still. Use this when frame one fails, when the shot you need is not in the photo, or when the early tests do not move the dominant symptom.
Do not start at step 5 because a timeline comment said another model is "better at products." That claim may be out of date, and it may be irrelevant to your symptom. A model switch can also hide a prompt conflict. You then learn nothing about the prompt you will reuse next week.
What to record before the rerun
Open a note, or add a row in the AI video generation log, with:
- Still file name and crop
- Model name as shown in the product
- Duration and aspect ratio
- The prompt, copied, not paraphrased
- Dominant symptom in one sentence
- The single variable you will change
- What result would count as better
That last line keeps you honest. "Better" might mean "the pump stays vertical at the middle frame," not "the clip feels more premium."
How to score the pair
Watch the failed clip and the retest back to back, then pause both on the same relative moment. Score only the dominant symptom: worse, the same, better, or gone. Then answer three yes or no questions:
- Did frame one still match the still?
- Did a new blocker appear?
- Is the clip usable for this placement if you stop here?
Keep the change only when the symptom is better or gone, frame one holds, and any new issue is truly minor for the placement. Revert when the symptom is the same or worse. Revert when the symptom improves but a new blocker is worse than the original. A calmer clip with a melted logo is not a keep if the logo was the job.
If you keep a change, the next test builds on that kept version. The kept prompt becomes the new baseline. You still change only one thing. If you revert, go back to the previous baseline and pick the next variable. Do not keep half of a failed idea because the lighting felt nicer.
Stop spending on the same idea
Stop the loop when any of these is true:
- Frame one fails the job.
- The symptom is unchanged after a motion reduction and after a camera removal.
- You already compared one other model with the still and prompt frozen, and the same symptom remains.
- The usable version no longer contains the motion the placement required, and a different composition would solve that more honestly.
- You cannot say which variable you changed.
Two or three informed runs are a diagnosis. Eight runs that each change several words are a search. Searches can work, and they can also burn the credit budget for the rest of the set. Check pricing for current credit costs before you batch a repair across many products. This guide does not list prices, because they can change.
A worked retest
A ceramic mug still is clean: one handle, readable wordmark, soft window light from the left. The failed clip asks for a slow orbit, a hand entering to hold the mug, and "living reflections." By one second the handle thickens. By the middle a hand has fused into the handle and the wordmark swims. The light also warms.
Dominant symptom: the handle and the hand fuse. Runner-up: the wordmark swims. The warm grade is noted and ignored for this pass.
Pass one changes only the hand. The orbit, the reflection line, the still, the model, and the duration stay. Result: no hand enters, the handle still thickens on the orbit, and the wordmark still swims. The hand removal helped the fusion and did not fix the mug. Decision: keep the hand out, because a fused hand is a blocker, and continue from this new baseline. The dominant symptom is now the handle warp.
Pass two, from that baseline, removes the orbit and does not touch the reflection line. Result: the handle holds, the wordmark is calmer but still soft, and the grade stays closer to the still. Decision: keep the locked framing. New dominant symptom: soft wordmark.
Pass three deletes only the living-reflection line. Result: the wordmark holds at a pause, the handle holds, and the clip is quiet. Decision: usable for a product teaser that does not need a move around the mug. If the placement required the orbit, this still cannot support that shot. The next action is a new angle designed as a still, not a fourth adjective on the orbit.
Notice what this sequence refuses to do. It does not switch models in pass one. It does not shorten, relight, and rewrite together. It does not call the quiet clip a failure just because it is quiet. It also does not call the orbit "fixed" when the only version that holds has no orbit.
A second pattern, for late identity drift: a portrait holds for the first moments and changes jaw shape near the end. Pass one only shortens the clip to the last clean moment. If that duration meets the placement, stop. If the placement needs the original length, revert the shorter duration and, as the next single change, remove the head-turn line. You never shorten and rewrite the head turn in the same generation, or you will not know which one saved the likeness.
When to abandon the still
Abandoning the still means you stop buying near-duplicate animations of this file. It does not mean the campaign is over. It means the reference is the wrong reference for the shot.
Abandon this file when:
- Frame one already has the wrong product shape, the wrong person, an unreadable critical label, or a broken hero hand.
- The angle, crop, or aspect you need is not in the image. A front packshot will not reliably become a top-down demo.
- The still's light, pattern, or noise is the thing that boils or jumps, and a locked short test still fails.
- You need exact text, and a locked test still melts the words that carry the claim.
- The subject is buried in occlusion: stacked products, overlapping hands, a face behind a glass that already distorts it.
- The early protocol did not move the dominant symptom.
- The only versions that "pass" have removed the action the placement was for, and a redesigned opening would allow that action with less invention.
Repair the still when the scene is right and a local defect is wrong. A cleaner cutout, a true label composite, a fixed hand, or a grade that matches one light source can save the concept. That repair is an image job. Generate or edit it, review it at full size, and only then send it back through image-to-video.
Replace the still when the concept is right and the photograph is the wrong shot. Design the opening as the opening. Give the product scale, give the face a readable angle, leave the hands out or make the grip simple, and leave space if type will be added later. First frame design is the checklist for that replacement. The AI image generator can create the new frame when you do not have a photo.
Leave image-to-video entirely when you do not have a still that contains the idea. A transformation, a scene that must be invented, or a camera path through a place you never photographed is often a text-to-video job. Use the AI video generator for that concept, and judge it with video standards rather than with still-fidelity standards. Forcing image-to-video to invent a new set is how cameras start ignoring the still.
A simple rule sits under all three paths. If you muted the motion and the still would still be rejected, do not animate it. If the still would be approved and only the motion breaks it, stay in the retest protocol. If you keep hoping the model will photograph a side you do not have, stop and make that picture.
QA before you call the clip fixed
A fix is not a feeling that the new clip is better. A fix is a pass on the original symptom without a new rejection, at the size and length you will ship.
Play the clip once with the sound off if the placement can be watched muted. Many product and social cuts are judged silent. Then pause.
Frame one
- The crop matches the still you approved.
- The person, product, or pack is the same subject.
- The light direction matches.
- No new objects have appeared at the edges.
If frame one fails, the clip fails, even if a later frame looks lucky.
The symptom you set out to fix
- Warp: the silhouette matches at the middle and the end.
- Identity: start, middle, and end could be the same person.
- Flicker: a short stretch of the hero surface does not strobe.
- Hands: finger count and contact points hold where hands are visible.
- Label: the words you must stand behind are readable on a pause.
- Camera: the viewpoint stays inside the shot you asked for.
- Light: shadow direction and product color do not jump to a new state.
New blockers
Look for the runner-up symptom on purpose. A locked camera can cure warp and expose flicker you were not staring at. A shorter clip can hide late drift and also cut the action too early. A model change can hold a face and melt a logo. Write the new issue down. If it would get the clip rejected, the retest is not a pass.
Playback and pause
Watch at delivery size. A vertical ad and a product-page loop do not share a tolerance. Pause on the hero for a full second. Viewers pause. Clients pause. A label that only works while moving is not a pass for a shop placement.
Check the end frame as carefully as the start. End-frame collapse is easy to miss if you stop the review when the opening looks good.
What a pass can mean
A full pass means you would ship this file for this placement. A conditional pass means the symptom improved, a minor issue remains, and you have logged it. Conditional is allowed for an internal concept review. It is a weak status for a paid placement or a page where the product must be accurate. A fail means you revert or you abandon the still. Do not average a fail and a vibe into a pass.
QA is also a record. The generation log should receive the passing prompt, the model name, the still file, and the symptom that was cleared. The next SKU or the next portrait can start from that baseline instead of from the original failure. If several passing shots must cut together, check color after the clip passes, using the color consistency checks. Do not regrade a clip that still has a warped product. Finish the structural fail first.
Placement bars
Use a bar that matches the job, and write it down so a later reviewer does not invent a stricter one after the fact.
- Paid product ad: silhouette, label, and color must survive a pause. Mild background life can stay if it does not shimmer.
- Product page or storefront loop: the pack must match the still a shopper could compare with photos. Exact claims on generated type need a human read.
- Organic social face: likeness at a glance matters more than micro texture. A new jaw or a melted eye still fails.
- Internal concept: you may accept a conditional pass to show an idea, and you should label it as a concept so nobody exports it as final.
- Multi-shot edit: each clip needs the same subject and a light direction you can cut. One hero clip that drifts will not be saved by the surrounding shots.
If you are unsure, review with the person who owns the product or the likeness, not only with the person who wrote the prompt. A prompt author can overlook a bent spout because they remember the prompt. A merchandiser will see the spout.
A studio routine you can repeat
Use the same short routine every time a clip fails. It is deliberately boring. Boring is what keeps the credit spend attached to a result.
- Export or save the failed clip where you can find it.
- Put the still next to it.
- Play once at delivery size. Scrub frame one, first motion, middle, and end.
- Write one dominant symptom and one runner-up.
- Pick the first variable from the table. Do not negotiate with yourself about a second change.
- Generate one retest.
- Score worse, same, better, or gone. Note any new blocker.
- Keep, revert, or abandon the still.
- If you keep, stop or plan exactly one next variable.
- If you abandon, write which path you are taking: repair the still, design a new first frame, or move the idea to text-to-video.
Teams can split the watch. One person names the symptom. Another confirms it on a pause. The person who spends the next credits should be able to read the note without a meeting. "Try something more cinematic" is not a handoff. "Orbit removed, handle held, wordmark still soft, next test deletes the reflection line" is a handoff.
When the same symptom appears on a whole set of images, do not repair all of them at once. Fix one representative still until you have a keep. Then apply that single kept change to the next image and watch whether it travels. A fix that worked on a tall bottle may warp a wide jar. Confirm on the second asset before you queue the rest. Queueing the rest is how a wrong diagnosis becomes a large credit bill. Pricing and included credits depend on the current plan, so check pricing when you move from one repair to a batch.
Model choice stays inside that routine, not beside it. If you switch models, the still, the prompt, and the duration stay frozen, and you switch once. Read the current names and capabilities on models at the time you run the test. A model that held labels last month may behave differently after an update. Your log is how you notice. This guide does not rank models and does not quote benchmark scores.
Commercial use, retention, and plan limits are also current-term questions. Check the plan and the model terms you are actually using before you publish client work or run a large set. Nothing in a prompt constraint is a legal guarantee that a face, a logo, or a product claim will be preserved.
What to do with a pass
A passed clip is allowed to leave diagnosis. It is not automatically a campaign. Confirm the length, the crop, and the claim. If critical words were added in editing, read them again on the final export. If the clip must match other shots, check color after the structural pass, not before.
Then store the baseline. The still, the prompt, the model name, the duration, and the symptom you cleared are the asset. The mp4 alone is not enough, because you will not be able to reproduce the pass when the next colorway arrives. The generation log is the right home for that record.
If the pass is quiet and the placement needs more life, add one motion back and treat it as a new test. A passing locked frame is a baseline, not a ceiling. You earn the right to try a small push, a little environmental motion, or a shorter camera move because you can revert to something that already held. Start that trial from the image to video prompt guide, with one camera or subject action, and run the same scorecard. If the added life brings the old symptom back, you have learned the limit of this still. Keep the quiet pass or abandon the still for a frame designed for that move.
Where this guide stops
Still-image defects that you can see before any video run, including broken anatomy in a single frame, garbled type on a poster, and pattern errors in a static render, belong in the common AI image artifacts guide. Fix them there, then return.
How to write motion, camera, and preservation language in a structured prompt belongs in the image to video prompt guide. Use it when you are choosing words on purpose, not when you are still unsure which failure you are looking at.
How reference images are prepared and uploaded belongs in the image reference consistency guide. How an opening frame is designed belongs in first frame design. How a product photo becomes a clip, end to end, belongs in the product photo to AI video workflow. How you record the runs belongs in the generation log. How approved shots keep a stable grade belongs in color consistency.
This page sits between those jobs. You already have a clip. Something in it fails. You name it, you change one thing, and you either keep a result you can explain or you stop animating that still.
When you are ready to run the next controlled test, use the image-to-video workflow in Cliprise with the still, the prompt, and the model written down. If the still is the part that failed, make a new frame with the AI image generator first. If the idea was never inside the still, start from the AI video generator instead of asking a photograph to become a different shot.