Defects and their causes
Almost every complaint about a generated clip maps to a decision the prompt left unnamed. A generator has to fill every gap, and it fills them with the average of what it was trained on, so an undescribed background becomes a cluttered one, an unnamed camera drifts, and a face on an audio capable model starts talking. This page lists the defects people actually report, the specific reason each one happens, and the clause that fixes it. Find the symptom, change only what the cause names, and regenerate. A defect report is not a reason to rewrite the whole prompt.
Anatomy and morphing
Hands, faces and shapes that come apart. These are the defects people notice first, and they are almost always about how many things are in frame and how much time the model has to keep track of them.
| What you are seeing | Why | What to change |
|---|---|---|
| "too many fingers", a hand that gains and loses digits as it moves | Fingers are small, fast moving and self occluding, so a model given no anatomy exclusion and no statement of what the hands are doing decides frame by frame. Every extra person in frame is another pair of hands to get wrong. | Say how many people are in shot, give the hands one specific job, and add the anatomy exclusion. If the hands are not the point of the shot, frame them out, because a shot that does not contain hands cannot get them wrong.One person in frame, both hands resting flat and still on the table, five fingers on each hand. No extra fingers, no extra limbs, no additional people. |
| "there is an extra hand", "extra limbs", an arm that belongs to nobody | No anatomy exclusion, and usually more people in frame than the duration can hold. Each additional figure multiplies the joints the model has to track, and a 5 second clip with 4 people has more limbs in it than it can resolve. | Cut the head count to the smallest number that still tells the shot, name what each person does, and add the anatomy exclusion. Where a crowd is the point, keep the extra people far from camera and out of focus.Two people in frame and nobody else, the woman on the left pouring, the man on the right watching. No extra limbs, no extra hands, no additional figures. |
| "the hands look melted", fingers fusing into a mitten, a hand that reads as putty | The hands are in frame doing something the prompt never named, so the model animates them freely, and free hand animation is where generators fail first. A hand with a purpose holds together, an idle hand deforms. | Give the hands one action with an endpoint, or crop them out of the frame. A hand gripping a named object is far more stable than a hand gesturing.Both hands hold the mug by the handle and do not move, fingers closed around it for the whole clip. |
| "the face is distorted", features sliding around, eyes at different heights | The face occupies too few pixels for the model to resolve, which is what happens on a wide or full body shot, and naming the face then asks for detail there is no room to render. A wide focal length with the camera close does the same thing by stretching the features outward, which is real lens behaviour rather than a fault. | Either go closer, to a medium close-up or tighter, so the face has pixels to spend, or stop describing the face and let it stay a silhouette at that distance. Where the shot has to be wide, keep the face away from the edge of frame and use a longer focal length.Medium close-up from the shoulders up, 85mm, the face filling the upper third of frame, features sharp and stable throughout. |
| "the subject warps", "it morphs halfway through", the shape holds for 2 seconds and then comes apart | The action does not fill the duration. The model finishes what the prompt described with time left over and invents a continuation. Longer clips fail this way more often, because there is more time to fill. | Either shorten the clip to the length of the action, or simplify to one action that completes inside it. Two things happening is two shots.5 seconds, one action only: the lid hinges open to 45 degrees and stops. Nothing else in the frame moves. |
Motion and pacing
Camera movement and subject motion are two separate decisions and are constantly confused for one. Name only the camera and the subject stands still. Name only the subject and the camera invents a move of its own.
| What you are seeing | Why | What to change |
|---|---|---|
| "my generated video is drifting", the camera wanders slowly for no apparent reason | Camera movement was not named, and an unnamed camera is not a still camera. The model supplies a small arbitrary move, and because nothing in the shot motivates it, it reads as an accident. Handheld sway is a common default in the training material, which is why an unnamed shot often judders as well. | Name one movement, or name the absence of one. Both are clauses, and leaving it out is neither. State the tripod positively rather than banning the shake, because on a model with no separate negative field a ban can act as a mention.Locked-off tripod, perfectly still, no camera movement at all, no zoom. |
| "nothing happens in the clip", 5 seconds of a still image with a faint push in | Only camera movement was named. Subject motion is a separate decision, and where it is missing the model holds the subject still and moves the camera instead. | Write both, in separate clauses: what moves inside the frame and how fast, and what the camera does.Steam drifts upward from the cup and dissipates before it leaves frame, while the camera holds still. |
| "it looks like a slideshow", a run of stills rather than a sequence | Subject motion unnamed in every shot, so each one is a photograph with a camera move on top. It is the same cause as a clip where nothing happens, spread across a set. | Give every shot one action that completes inside it, and change the shot size or the angle between consecutive shots so the eye has a reason to read a cut.SHOT 1 wide, the door swings open. SHOT 2 close-up, the hand turns the key and stops. |
| "the camera is doing weird things", a move that starts one way and changes direction partway | Two camera movements were named for one shot, so the model attempts both. A push in plus an orbit is two instructions competing for the same seconds, and a focal length that fights the framing has the same effect. | One movement per shot. If the shot needs a second move, it is a second shot.One continuous move: a slow push-in from a medium shot to a tight detail, no orbit, no pan. |
| "the ending is abrupt", the clip stops mid gesture | No pacing or beat structure, so the model spread one undifferentiated move across the whole duration and ran out of time rather than arriving anywhere. Nothing described a last frame, so there was nothing to land on. | Say where the shot starts and where it ends, make those two states different, and ask for it to settle and hold.Three beats: reveal, detail, settle. All elements ease into their final positions and the last second holds completely still. |
| the brightness or the colour pulses through the clip, or the image flickers | Exposure and white balance were never pinned, so the model re-decides them as the shot progresses. It shows up worst on a slow move across a scene that has a bright side and a dark side. | Ask for locked exposure and white balance, minimal motion blur, and no flicker. It is one clause and it applies to almost every shot.Locked exposure and white balance throughout, no flicker, minimal motion blur. |
Things that appeared uninvited
A generator cannot render nothing. Every part of the frame you did not describe is a part it will fill with something plausible, and plausible means whatever it saw most often.
| What you are seeing | Why | What to change |
|---|---|---|
| "there are objects that were not in my prompt", furniture and clutter nobody asked for, or something appearing behind the subject partway through | The background was never described, and an undescribed background is where a model puts whatever it likes. Describing it only for the first frame has the same result later in the clip, because nothing said it had to hold. | Describe the background even when it is meant to be nothing, say it stays fixed for the duration, and keep it low in detail. A named empty background is a decision, an omitted one is a lottery.Plain charcoal studio background, low in detail, unchanged for the whole clip. One object on an empty surface, nothing else in frame. |
| "there are people I did not describe", figures in the background of a room that was meant to be empty | The space was implied to be empty rather than stated as empty, and a room in the training material usually has people in it. | Say the space is empty in positive terms. On a model with no separate negative field, "no people" can act as a mention and pull a figure into frame.An empty room, uninhabited, empty of figures, nobody anywhere in frame. |
| an object in shot gets picked up, opened, swung or switched on when it was meant to sit there | An object the model has mostly seen in use tends to get used. Naming it without naming its state reads as an invitation to do the thing it is for. | Name the object as inert and say what it does not do. State the limit of any movement it is allowed.The tool stays inert on the bench for the whole clip and is never picked up, never raised, never used. |
| "there is an artifact in the generated video", "there are glitches", tearing or smearing in one region of frame | Usually a conflict rather than a model failure. Two clauses ask for things that cannot both be true, such as a locked off camera and a parallax move, or the action is too complex for the seconds available. | Read the prompt back as a list of instructions and find the two that contradict. Delete one, then cut the shot to a single action. Adding more description does not fix a contradiction, it buries it. |
Consistency across shots
A generator has no memory between calls. Shot 4 does not know what shot 1 looked like, so the only thing carrying a character, a look or a light across a cut is the same words appearing in both prompts. Not similar words. The same ones.
| What you are seeing | Why | What to change |
|---|---|---|
| "the character changes appearance", the same person with a different face, hair or wardrobe in shot 3 | The description was written again from memory instead of pasted, so a slightly different description produced a slightly different person, and the differences accumulate down the sequence. | Write the character description once, keep it as text, and paste the identical string into every prompt that features them. Do not improve it between shots. Anchor the cues that carry identity at small sizes: hair shape, eye construction, the outline of the clothing. List only the characters actually on screen, because naming the rest of the cast invites them into frame.CHARACTERS: a woman in her early 30s, shoulder length dark curly hair, thin silver frame glasses, olive green field jacket over a grey shirt. |
| "the shots do not match each other", "it changes between shots", a set that looks like 3 different productions | The style and grade line was rewritten between prompts rather than repeated. A paraphrase is a different prompt, and nothing tells the model that two descriptions were meant to mean the same look. Where the style was only ever described inline in the first prompt and never saved, the second one had nothing to copy from. | Keep one style string and one grade string for the whole piece, saved as text, and paste them unchanged into every prompt. Where all the shots are in one prompt, state them once in a header above the shots instead, because restating a decision invites the model to reinterpret it mid-prompt.STYLE: photoreal, shot on 35mm with fine grain, teal and orange grade, warm skin against cool shadows. |
| the light changes direction or colour between shots that are meant to be the same moment | Either the lighting clause was dropped from the later prompts, or it was carried unchanged into a shot that moved to a different location, where it no longer describes anything real. | Repeat the lighting clause word for word while the location holds, and rewrite it deliberately when the location changes rather than letting it fall out.LIGHT: single hard key from camera left, sharp shadow edge on the far cheek, no fill. |
| two people in the same shot start to look like each other, or swap features partway through | They are described in similar terms and read alike in silhouette, so the model has nothing to keep them apart and averages them. The same height, the same hair mass and the same wardrobe weight is enough to cause it. | Name one difference that survives at silhouette scale: height, hair shape, or the outline of the clothing. Keep the two descriptions in the same order in every prompt, because a generator reads position as emphasis.The taller figure in a long straight coat, the shorter figure in a bulky jacket with a wide collar. Their silhouettes never match. |
Text and branding
Letterforms are fine geometry, which is the thing these models approximate worst, and the footage they learned from is full of titles, captions and watermarks. Assume any lettering in frame will degrade over the length of a clip.
| What you are seeing | Why | What to change |
|---|---|---|
| "there is text on the video", words, captions or a lower third appearing over the frame | No exclusion clause. Video in the training material carries titles, subtitles and burned-in captions, so text is a habit rather than an accident, and nothing told the model to leave it out. | Add the exclusion clause. This is the single highest value line you can put in a video prompt, and where the model has a separate negative field it costs nothing against the prompt budget.No text, no captions, no subtitles, no on-screen type, no logos, no watermarks. |
| "there is a watermark I did not ask for", a corner mark or a stock library overlay | Same cause. A lot of stock footage carries a watermark, and excluding text alone does not always cover a graphic mark. | Name watermarks specifically in the exclusion clause, alongside logos and idents, rather than relying on a general ban on text.No watermarks, no logos, no station idents, no corner graphics. |
| there is text on a sign or a package and it is not real words, or a word you asked for is misspelled | The model is drawing glyph shapes rather than writing, and it has no spell check. The longer the word and the longer it stays on screen, the further it drifts. | Do not generate type you care about. Generate the motion with clear empty space where the type belongs, state that space in the prompt so the composition leaves room for it, then set the type afterwards. Where lettering must be generated, keep it small, keep it briefly seen, and keep the surface it sits on still.Clean negative space across the upper third of frame, unmarked. All surfaces clean and unmarked, no printed text and no signage anywhere in frame. |
| the brand mark warps, thickens or changes shape through the clip | A logo is fine geometry, it often contains text, and it is riding on a surface that moves. The damage is proportional to how large the mark is and how long it is in shot. | Composite the mark afterwards over clean generated motion. That is what most production actually does and it is the only reliable answer. Where it has to be generated, hold it still, keep it small, and say it does not change.The printed label stays sharp and unchanged throughout, the same size and position, on a surface that does not move. |
Audio
On a model that generates sound, saying nothing about sound is not a request for silence. The model fills the gap with the average of what it heard, which is a music bed, a room tone that does not match the room, or a voice. So the prompt always carries an audio line, even when that line asks for silence. On a model with no audio track, drop these clauses entirely.
| What you are seeing | Why | What to change |
|---|---|---|
| "there is music on my clip and I never asked for music" | No audio line at all, so the model chose. The average of a lot of video with music on it is more music. | Write an audio line on every prompt for a model that generates sound. Say what the ambience is, what the effects are, and whether there is music. Silence is a clause too.AUDIO: room tone and the sound of the mechanism only, no music. |
| a voice started explaining the shot back to me | Speech was never ruled out, and a describable scene reads to an audio capable model as an opportunity for narration. | Rule out voice in the audio line itself, not only in the exclusions, because the audio line is what the model reads for sound.AUDIO: ambience and effects only, no voice, no dialogue, no narration. |
| a character's mouth is moving and there was no dialogue in the prompt | A face in shot on an audio capable model starts talking unless something says otherwise. Lip sync was not ruled out. | Say the characters do not speak and do not lip sync. This matters even for a silent cut, because a mouth moving for no reason is unusable with the sound off.Characters do not speak and do not lip sync at any point. |
| the music starts again at every join in the edit | Each clip composed its own music from nothing, so identical wording returned different takes: a different tempo drift, a different point in the phrase, a different mix. No wording prevents this, because no generation can hear the previous one. | Ask every scene prompt for ambience and effects with no music, and lay one track across the joins afterwards. Where the whole piece is generated in a single pass, the music can go in that one prompt and will genuinely play throughout.AUDIO: ambience and effects only, no music. The score is laid over afterwards. |
| the effects sound right and the clip still sounds synthetic | Ambience was left out. Effects with no room tone under them give the clip sound with no space around it, and that missing space is the part that reads as fake. | Write the ambience as its own item: the sound the location makes with nothing happening in it. Name instruments rather than a genre wherever music is wanted, because a genre is a category and categories average out.AUDIO: room tone of a small kitchen with a fridge hum underneath, and a latch clicking as the door closes. |
Commercial fidelity
Commercial work has a constraint the rest of this page does not. A cinematic clip that invents a detail is interesting. A product clip that invents a detail is unusable, and sometimes legally unusable. The governing rule is to start from a real photograph and prompt only the motion.
| What you are seeing | Why | What to change |
|---|---|---|
| it looks like my product but the details are wrong | A text to video model was asked to invent the product and then animate it, so it returned something in the same family. It never had the geometry, the label or the colour. | Use an image to video route from a real product photograph, so the model has the product and only has to move it. Then prompt motion, light and camera only. Do not redescribe what is already in the frame, because redescribing the product is how you invite it to be redrawn.Animate the supplied frame. Slow orbit around the product, one quarter turn, the product itself perfectly still, nothing in the frame redrawn. |
| the printed text on the packaging turns to mush partway through | Printed text is fine geometry on a moving surface, which is the hardest combination in the frame to hold. The damage grows with the length of the clip and with how much of the frame the label fills. | Keep the clip to 4 to 6 seconds, move the camera rather than the product so the label never has to be redrawn from a new angle, and say the label stays unchanged. Composite the label afterwards where it has to be readable.The label stays sharp and unchanged throughout, no warping of the printed text, no change to the product silhouette. |
| the stripes bend the wrong way, or the check pattern slides across the fabric | A regular pattern on a moving garment is fine geometry in motion, the same problem as text. The grid has to be redrawn every frame and nothing tells the model where it should sit. | Prefer plain or irregular fabrics for generated motion. Where the pattern is fixed, say it stays aligned to the weave, name the seam and hem positions that must hold, and hold the pose so the fabric is the only thing moving.The check pattern stays aligned to the weave and does not distort at the seams. Hem, cuff and sleeve length unchanged. The model holds the pose exactly, only the fabric moves. |
| the stone glows instead of reflecting, the facets smooth over, the metal reads as plastic | Jewellery is mostly specular reflection and fine geometry, and a generator approximates both. Metal reads as metal only through a sharp reflection, so diffuse light turns it into plastic. Every extra second of movement is another chance to lose a facet. | Keep it to 3 to 5 seconds and a few degrees of movement, frame it macro, name hard sources so the highlight is a sharp line rather than a glow, and state that the facets, the setting and the prong count do not change.Macro on the setting, facets sharp and individually readable, two hard sources at 45 degrees. Extremely slow rotation of a few degrees only, the stone shape and the 4 visible prongs unchanged, no glowing or emissive gems. |
| the product ended up behind the caption and the buttons once I posted it | The shot was composed for the full frame, but a social app draws its own interface over the frame, and the reserved area at the bottom and down the right hand side is large. Nothing in the prompt kept the subject clear of it. | Decide the platform before composing, keep the subject inside the usable box, and leave the reserved areas empty. The recommended universal safe area on a 1080 by 1920 frame reserves 288 pixels at the top, 672 at the bottom, 120 on the left and 300 on the right, which means the usable box is not centred: it sits left of the centre line, because every platform's action rail is on the right.Vertical 9:16. The product sits in the upper middle of frame and slightly left of centre, well clear of the lower third and the right hand edge, which stay empty. |
Animated and drawn styles
The most common mistake in an animation prompt is writing it as though it were live action. Optical language on flat art produces a photographic hybrid nobody wanted. Drop the optics and name the drawing instead.
| What you are seeing | Why | What to change |
|---|---|---|
| "it looks like video, not animation", the drawing moves too smoothly and never stops changing | Nothing named the animation timing, so every frame is a new drawing and the motion is continuous. Real 2D animation holds drawings, and the holds are what read as animation rather than as a morphing video. | Name the timing and ask for held poses. A character description repeated word for word across shots gives the holds something fixed to return to.Animated on twos, with the pose holding completely still for a beat before the action, then a clean held pose at the end. |
| the outlines are black and the reference uses brown | Outline colour was not stated, and black is the default a model reaches for. Most productions use a dark brown or a dark shade of the fill colour instead, or no outline at all. | State the outline colour and its weight, say whether the background carries one, and exclude black outlines, because that default is strong enough to need naming. Carry the same clause into every shot, or the line weight changes at the cut.Dark brown outlines of even weight on the characters, no outline on the background, no black outlines. |
| lens flare, bokeh and a blurred background appeared on a flat cartoon | A lens or a depth of field was named on a style that has no camera in it, so the model tried to satisfy both the drawing and the optics. | Drop the optics from a drawn style. Say flat with no lens effects, and get depth from layers sliding at different speeds instead of from focus.Flat, no lens effects, no depth of field, no flare, no chromatic aberration. Simple horizontal pan, layers sliding at different speeds for parallax. |
| the animation moves but it has no snap or weight | Everything was eased evenly, with no anticipation before an action and no overshoot after it. Even easing across a whole move is what makes generated animation feel weightless. | Ask for the timing beats by name rather than describing the movement as a whole.A small anticipation opposite to the move, then the action, then a slight overshoot as the pose settles. Follow through on the hair and the coat after the body stops. |
The prompt was only partly followed
Two different failures look the same from the outside. Either the prompt was longer than the model allows and the end of it was cut off silently, or it was written as prose and the model could not find the instructions in it.
| What you are seeing | Why | What to change |
|---|---|---|
| "it ignored the last part of my prompt", or "text and watermarks came back even though I excluded them" | The prompt exceeded the model's character limit and was truncated from the end, with no error and no warning. The exclusion clause sits last, so it is the first thing truncation removes, which is why the output comes back carrying exactly what you banned and nothing points at length. A model that rejects the prompt outright instead is doing you a favour, because it tells you. | Count the characters rather than estimating, and write to the model's documented limit less 10 percent. Where the limit is unknown, write to 1000 characters, which is the tightest limit among the major hosted models, so a prompt that fits it runs anywhere. Move the exclusions into the model's separate negative field where one exists, because that is a separate budget. Never cut the exclusion clause to make room: if the prompt only fits without it, it is describing more than one shot. |
| "it ignores half my prompt" | The prompt is prose rather than a list of clauses. Atmosphere spread across sentences gives the model no discrete instructions to act on, so it acts on the few it can pick out. Stacked synonyms make it worse: cinematic, filmic and movie-like is one decision written 3 times, and the model averages the three rather than weighting the decision more heavily. | Rewrite as one clause per decision, commas between them, in a fixed order, with the exclusions last. Prefer a number to an adjective wherever a number exists: 8 seconds rather than short, 85mm rather than flattering, 92 BPM rather than upbeat. Say each decision once, in the most specific words available. |
| "it looks generic", "it looks like stock footage" | Medium, grade and lighting were left unnamed, so the model returned the average of its training material, and the average of a very large amount of video is stock video. | Name the medium, the grade and the light. Those 3 clauses move a shot further from the average than any amount of extra subject description.Photoreal, shot on 35mm with fine grain, teal and orange grade, single hard key from camera left with a sharp shadow edge and no fill. |
Common questions
Why does my AI video have extra fingers?
Fingers are small, they move fast and they hide each other, so a model with no instruction about the hands decides what they look like frame by frame. It gets worse with every extra person in frame, because each one is another pair of hands to resolve. Say how many people are in shot, give the hands one specific job with an endpoint, and add an exclusion for extra fingers and extra limbs. If the hands do not matter to the shot, frame them out.
Why does my character change between shots?
A generator has no memory between calls, so shot 3 knows nothing about shot 1. If the character was described again in slightly different words, the model was given a slightly different person, and the differences build up across the sequence. Write the description once, keep it as text, and paste the identical string into every prompt that features them.
Why is there text in my AI generated video?
Because the footage these models learned from is full of titles, captions and lower thirds, so adding text is a habit rather than an accident. Nothing removes it unless the prompt says so. Add an exclusion covering text, captions, subtitles, on-screen type, logos and watermarks. It is the highest value line you can put in a video prompt, and where the model has a separate negative field it costs nothing against the prompt length.
Why is there a watermark on my AI generated video?
A lot of stock footage in the training material carries one, and an exclusion for text alone does not always cover a graphic mark. Name watermarks specifically, alongside logos and channel idents. If a watermark keeps returning even though you excluded it, the prompt is probably over the model's character limit and the exclusion clause is being cut off the end.
Why does nothing happen in my AI video?
Camera movement and subject motion are two separate decisions, and naming only the camera means the model holds the subject still and moves the camera instead. Write both in separate clauses: what moves inside the frame and how fast, and what the camera does. A shot with no motion clause at all usually comes back as a slow arbitrary drift, because an unnamed camera is not a still camera.
Why does my AI video morph halfway through?
The action you described does not fill the time you asked for, so the model finishes it early and invents a continuation. Either shorten the clip to the length of the action or simplify to one action that completes inside it. This is the main reason 5 seconds is a better default than 20: a long clip has more time to fill and more chances to drift, and a failed long clip costs several short ones.
Why is there music on my AI video when I did not ask for it?
On a model that generates audio, silence is not the default. Saying nothing about sound leaves the model to choose, and it chooses the average of what it heard, which is a music bed. Always include an audio line, even when it asks for silence, and rule out voice and lip sync explicitly, because a face in shot will otherwise start talking.
Why did the model ignore the end of my prompt?
Most video models enforce their prompt limit by truncating from the end rather than returning an error, so a prompt that is too long comes back partly unread and nothing reports it. The exclusion clause sits last, which is why the things you banned are the first to reappear. Count the characters rather than estimating, write to the documented limit less 10 percent, and where the limit is unknown write to 1000 characters.
Why does my AI product video not look like my product?
A text to video model asked to invent your product will produce something that resembles it, because it never had the geometry, the label or the colour. Start from a real product photograph and use an image to video route, then prompt only motion, light and camera. Do not redescribe the product in that prompt: the model already has it, and describing it again invites a redraw.