{"course":{"slug":"ai-video","title":"Video with AI","url":"https://addaly.com/learn/ai-video","level":"intermediate","track":"create"},"lesson":{"slug":"why-clips-are-short","title":"Nobody is hiding the long version from you","url":"https://addaly.com/learn/ai-video/why-clips-are-short","position":1,"of":66,"minutes":8,"answer":"Clip length is a memory-and-attention limit, not a product tier, so plan in shots and cut before the drift rather than trying to generate through it.","keyIdea":"Clip length is a memory-and-attention limit, not a product tier, so plan in shots and cut before the drift rather than trying to generate through it."},"body":{"format":"markdown","text":"## The limit is arithmetic, not marketing\n\nMost people assume the few-second cap on generated clips is a pricing tactic: that a longer version exists behind a higher tier. It does not. The length falls out of how these models are built, and understanding why changes how you plan everything else.\n\nA video model does not make frames one at a time. Frame 90 has to agree with frame 12 about where the wall is, so the model generates the whole clip as a single object. To make that affordable it compresses first: a video autoencoder squeezes the picture by roughly eight times in each spatial direction and by four times or more along time, turning a few hundred frames into a much smaller block of latent tokens. A transformer then denoises that entire block, attending across all of it.\n\nAttention cost grows with roughly the square of the token count. Double the duration, roughly double the tokens, roughly quadruple the attention work, and the whole block has to sit in GPU memory at once while it is denoised. A ten-second clip is not twice the price of a five-second one. It is several times worse, and past a point it does not fit on the card at all.\n\n> Figure: What another second of clip actually costs. Bars (relative attention work, where a five-second clip is 25): 2 seconds 4; 4 seconds 16; 6 seconds 36; 8 seconds 64; 10 seconds 100. Attention work grows with roughly the square of the token count, and tokens grow with duration. Doubling the length does not double the bill, it quadruples the attention work, and the whole block has to sit in GPU memory at once while it is denoised. Five to ten seconds is where quality, price and memory currently meet.\n\nThat is the entire explanation. Five to ten seconds is where quality, price and memory currently meet.\n\n## Why the last second is the worst second\n\nIn image-to-video the first frame is given to the model. Every later frame is inferred, and each one is anchored by the frames before it rather than by anything fixed. Error accumulates with distance from the anchor.\n\nYou will see this in a consistent pattern:\n\n- The first quarter-second is a settle. Motion starts slightly wrong and corrects.\n- The middle is the good part.\n- The last second is where the face loosens, the background churns, the colour creeps warm, and any object that left frame comes back subtly changed.\n\nSo generate long and deliver short. If you need six seconds on screen, generate ten. Treat the head and tail as offcuts, the way a printer treats bleed.\n\n## Extending compounds the damage\n\nEvery tool offers an extend button. What it does mechanically: takes the last frame of the previous generation (which is already a decoded, lossy output) and uses it as the conditioning image for a fresh generation. That frame gets re-encoded into latent space, generated from, decoded again.\n\nThis is a photocopy of a photocopy. Contrast climbs. Saturation creeps. Fine texture smooths into a painted look. Three extends in, the room is a different colour from where it started, and no prompt fixes it because nothing in the prompt changed.\n\nIf you must extend, do one of two things. Colour-match the sections back together in the edit afterwards. Or better, treat each extension as a separate shot and cut between them: a mismatch that reads as a cut is invisible, while the same mismatch inside a continuous move is glaring.\n\n## Plan in shots, and put the cut where the model is weakest\n\nA three-minute explainer is thirty shots. Write it as thirty lines with a duration next to each one before you generate anything. This is not an AI workflow, it is a shot list, and it is the same document a crew would work from.\n\nThen place your cuts deliberately. If the hand goes wrong at six seconds, the shot is five seconds long. You are not compromising; you are doing what editors have always done, which is end the shot before it stops working.\n\nYou can write this list on a phone in a notes app. It costs nothing and it is the highest-value thing in the whole pipeline.\n\n## When longer clips arrive, less changes than you expect\n\nModels will keep stretching the window. It matters less than it sounds. Commercial editing runs at two to four seconds a shot; a sixty-second unbroken take is not something most projects want. The constraint you are fighting is already close to normal film grammar.\n\nThe thing a longer window does not fix is scenes. Two people talking across a cut, matching eyelines, the same jacket, the same key light: that is the unsolved problem, and it is about consistency between generations, not length within one. A later lesson deals with it honestly.\n\n---\n\nToday: take whatever you actually want to make, write it as a shot list with a duration on every line, and count the lines. That number, times your cost per shot, is your project."},"check":{"question":"A team generates an eight-second clip and then presses extend three times to reach thirty seconds. The final section is noticeably more contrasty, more saturated and softer in detail than the opening. What happened?","options":[{"text":"Prompt adherence decays over long durations, so the later sections followed the description less closely than the first.","correct":false,"tempting":"Attributes a compounding image-quality loss to instruction-following, when the prompt is re-applied identically at each step."},{"text":"Each extension conditions on the last decoded frame of the previous generation, so encoding and decoding error compounds like a copy of a copy.","correct":true},{"text":"The clip passed the model's trained duration, so the tool fell back to frame interpolation to fill the remaining time.","correct":false,"tempting":"Invents a fallback mechanism and confuses a training window with a different technique entirely."},{"text":"Later generations were served from a lower-cost tier once the session's high-quality credits ran down.","correct":false,"tempting":"Reaches for a billing behaviour where a signal-processing mechanism explains the result."}]},"prev":null,"next":{"slug":"latent-space-for-video","title":"Your clip is compressed fifty times before anything happens","url":"https://addaly.com/learn/ai-video/latent-space-for-video"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Nobody is hiding the long version from you', in Video with AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-video/why-clips-are-short (Accessed: 31 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Nobody is hiding the long version from you","courseTitle":"Video with AI","module":{"index":1,"slug":"how-video-models-work","title":"What the machine is actually doing"},"answer":null,"keyIdea":"Clip length is a memory-and-attention limit, not a product tier, so plan in shots and cut before the drift rather than trying to generate through it.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure 1.1","figure":{"type":"bars","unit":"relative attention work, where a five-second clip is 25","title":"What another second of clip actually costs","series":[{"label":"2 seconds","value":4},{"label":"4 seconds","value":16},{"label":"6 seconds","value":36},{"label":"8 seconds","value":64},{"label":"10 seconds","value":100,"highlight":true}],"caption":"Attention work grows with roughly the square of the token count, and tokens grow with duration. Doubling the length does not double the bill, it quadruples the attention work, and the whole block has to sit in GPU memory at once while it is denoised. Five to ten seconds is where quality, price and memory currently meet."},"alt":"Figure: What another second of clip actually costs. Bars (relative attention work, where a five-second clip is 25): 2 seconds 4; 4 seconds 16; 6 seconds 36; 8 seconds 64; 10 seconds 100. Attention work grows with roughly the square of the token count, and tokens grow with duration. Doubling the length does not double the bill, it quadruples the attention work, and the whole block has to sit in GPU memory at once while it is denoised. Five to ten seconds is where quality, price and memory currently meet."}},"blocks":[{"t":"h2","id":"the-limit-is-arithmetic-not-marketing","runs":[{"text":"The limit is arithmetic, not marketing"}],"role":"limit"},{"t":"p","runs":[{"text":"Most people assume the few-second cap on generated clips is a pricing tactic: that a longer version exists behind a higher tier. It does not. The length falls out of how these models are built, and understanding why changes how you plan everything else."}]},{"t":"p","runs":[{"text":"A video model does not make frames one at a time. Frame 90 has to agree with frame 12 about where the wall is, so the model generates the whole clip as a single object. To make that affordable it compresses first: a video autoencoder squeezes the picture by roughly eight times in each spatial direction and by four times or more along time, turning a few hundred frames into a much smaller block of latent tokens. A transformer then denoises that entire block, attending across all of it."}]},{"t":"p","runs":[{"text":"Attention cost grows with roughly the square of the token count. Double the duration, roughly double the tokens, roughly quadruple the attention work, and the whole block has to sit in GPU memory at once while it is denoised. A ten-second clip is not twice the price of a five-second one. It is several times worse, and past a point it does not fit on the card at all."}]},{"t":"p","runs":[{"text":"That is the entire explanation. Five to ten seconds is where quality, price and memory currently meet."}]},{"t":"h2","id":"why-the-last-second-is-the-worst-second","runs":[{"text":"Why the last second is the worst second"}]},{"t":"p","runs":[{"text":"In image-to-video the first frame is given to the model. Every later frame is inferred, and each one is anchored by the frames before it rather than by anything fixed. Error accumulates with distance from the anchor."}]},{"t":"p","runs":[{"text":"You will see this in a consistent pattern:"}]},{"t":"ul","items":[[{"text":"The first quarter-second is a settle. Motion starts slightly wrong and corrects."}],[{"text":"The middle is the good part."}],[{"text":"The last second is where the face loosens, the background churns, the colour creeps warm, and any object that left frame comes back subtly changed."}]]},{"t":"p","runs":[{"text":"So generate long and deliver short. If you need six seconds on screen, generate ten. Treat the head and tail as offcuts, the way a printer treats bleed."}]},{"t":"h2","id":"extending-compounds-the-damage","runs":[{"text":"Extending compounds the damage"}]},{"t":"p","runs":[{"text":"Every tool offers an extend button. What it does mechanically: takes the last frame of the previous generation (which is already a decoded, lossy output) and uses it as the conditioning image for a fresh generation. That frame gets re-encoded into latent space, generated from, decoded again."}]},{"t":"p","runs":[{"text":"This is a photocopy of a photocopy. Contrast climbs. Saturation creeps. Fine texture smooths into a painted look. Three extends in, the room is a different colour from where it started, and no prompt fixes it because nothing in the prompt changed."}]},{"t":"p","runs":[{"text":"If you must extend, do one of two things. Colour-match the sections back together in the edit afterwards. Or better, treat each extension as a separate shot and cut between them: a mismatch that reads as a cut is invisible, while the same mismatch inside a continuous move is glaring."}]},{"t":"h2","id":"plan-in-shots-and-put-the-cut-where-the-model-is-weakest","runs":[{"text":"Plan in shots, and put the cut where the model is weakest"}]},{"t":"p","runs":[{"text":"A three-minute explainer is thirty shots. Write it as thirty lines with a duration next to each one before you generate anything. This is not an AI workflow, it is a shot list, and it is the same document a crew would work from."}]},{"t":"p","runs":[{"text":"Then place your cuts deliberately. If the hand goes wrong at six seconds, the shot is five seconds long. You are not compromising; you are doing what editors have always done, which is end the shot before it stops working."}]},{"t":"p","runs":[{"text":"You can write this list on a phone in a notes app. It costs nothing and it is the highest-value thing in the whole pipeline."}]},{"t":"h2","id":"when-longer-clips-arrive-less-changes-than-you-expect","runs":[{"text":"When longer clips arrive, less changes than you expect"}]},{"t":"p","runs":[{"text":"Models will keep stretching the window. It matters less than it sounds. Commercial editing runs at two to four seconds a shot; a sixty-second unbroken take is not something most projects want. The constraint you are fighting is already close to normal film grammar."}]},{"t":"p","runs":[{"text":"The thing a longer window does not fix is scenes. Two people talking across a cut, matching eyelines, the same jacket, the same key light: that is the unsolved problem, and it is about consistency between generations, not length within one. A later lesson deals with it honestly."}]},{"t":"rule"},{"t":"p","runs":[{"text":"Today: take whatever you actually want to make, write it as a shot list with a duration on every line, and count the lines. That number, times your cost per shot, is your project."}]}],"check":{"question":[{"text":"A team generates an eight-second clip and then presses extend three times to reach thirty seconds. The final section is noticeably more contrasty, more saturated and softer in detail than the opening. What happened?"}],"options":[{"text":[{"text":"Prompt adherence decays over long durations, so the later sections followed the description less closely than the first."}],"correct":false,"tempting":[{"text":"Attributes a compounding image-quality loss to instruction-following, when the prompt is re-applied identically at each step."}]},{"text":[{"text":"Each extension conditions on the last decoded frame of the previous generation, so encoding and decoding error compounds like a copy of a copy."}],"correct":true,"tempting":null},{"text":[{"text":"The clip passed the model’s trained duration, so the tool fell back to frame interpolation to fill the remaining time."}],"correct":false,"tempting":[{"text":"Invents a fallback mechanism and confuses a training window with a different technique entirely."}]},{"text":[{"text":"Later generations were served from a lower-cost tier once the session’s high-quality credits ran down."}],"correct":false,"tempting":[{"text":"Reaches for a billing behaviour where a signal-processing mechanism explains the result."}]}],"order":[0,3,1,2],"why":null},"prev":null,"next":{"slug":"latent-space-for-video","title":"Your clip is compressed fifty times before anything happens","minutes":10}}}