{"course":{"slug":"ai-video","title":"Video with AI","url":"https://addaly.com/learn/ai-video","level":"intermediate","track":"create"},"lesson":{"slug":"composition-that-survives-motion","title":"Compose for the frame you will end on","url":"https://addaly.com/learn/ai-video/composition-that-survives-motion","position":12,"of":66,"minutes":9,"answer":"A camera move changes the framing, so the still has to be composed for where the shot ends rather than where it starts.","keyIdea":"A camera move changes the framing, so the still has to be composed for where the shot ends rather than where it starts, and every edge of the frame is a place where the model has to invent what enters."},"body":{"format":"markdown","text":"## The still is the first frame, not the shot\n\nPeople compose a beautiful still, add a slow push-in, and are disappointed. The push-in cropped fifteen per cent off every edge, so the composition they liked is gone by second three and what remains is a tighter, worse version of it.\n\nCompose for the end of the move. If the shot pushes in, start looser than feels right. If it pulls out, start tighter. If it tilts up, leave the headroom you will need at the top. This is ordinary camera-operator thinking and it transfers exactly.\n\nA practical way to do it: decide the move first, then open the still in any image editor and draw a rectangle where the frame will end up. If the composition inside that rectangle does not work, the still is wrong, not the prompt.\n\n## Headroom, lead room, and why models get them wrong\n\nTwo classical rules the model has no opinion about:\n\n- **Headroom.** The gap above the head. Too much and the subject sinks; too little and they feel cramped. Models trained on centred stock footage default to too much.\n- **Lead room.** A person looking or moving to the left needs space on the left to look into. Without it the frame feels like it is pushing them into the edge.\n\nThe model will not fix either. It animates the composition it is given. Fix them in the still, where a crop costs nothing.\n\n## Every edge is a place things come from\n\nThis is the composition rule specific to generative video, and it does not exist in still photography.\n\nWhen the camera moves, new picture has to enter at the leading edge. The model invents it. It has no idea what was outside your frame, because nothing was: the still ends where it ends.\n\nSo a pan to the right invents everything on the right. A pull-out invents a ring of new scene on all four sides. A tilt up invents a ceiling or a sky.\n\nWhat follows:\n\n- **Put the difficult content in the middle.** Anything with structure (a building edge, a bookshelf, a patterned floor) should not be sitting half in and half out of frame, because the half that is outside will be invented and will not match.\n- **Prefer moves that reveal simple things.** Tilting up into plain sky works. Tilting up into an ornate ceiling produces melted ornament.\n- **A push-in invents nothing.** It only crops. That is why it is the safest move in the category and why models default to it.\n\n## Orbits and the flat-backdrop problem\n\nAn orbit or a parallax move asks the model to show the scene from an angle it was never given. If the still is a subject standing in front of an out-of-focus wall, there is no space behind them to reveal, so the model invents one as it goes and the shot reads as a cardboard cut-out sliding across a painted flat.\n\nThe fix is in the still: give the space depth before you ask the camera to move through it. Foreground element, mid-ground subject, background with real recession. A doorway, a table edge, a lamp in front, a corridor behind. A still with three planes orbits convincingly. A still with one does not, regardless of the model.\n\n## Aspect ratio is a composition decision, not an export setting\n\nMatch the model's native ratio when you make the still: this was covered before and is worth repeating because it is the most common waste. But there is a second point.\n\nIf you need both a 16:9 and a 9:16 version of a piece, you cannot crop one from the other and keep the composition. A vertical reframe of a landscape shot either loses the sides or scales everything up until it is soft.\n\nTwo honest approaches. Either compose the landscape still with a vertical safe area in mind, keeping everything essential inside a centre column, which costs you the sides as usable composition, or generate the shot twice, once in each ratio, from two stills. The second costs more money and looks considerably better, and for anything with a subject it is usually the right call.\n\n## Safe areas, still\n\nIf a platform overlays a caption, a profile picture and three buttons on the lower third of a vertical video, then the lower third of your composition is not yours. Check where the interface sits on the platform you are delivering to, and keep faces and text out of it. This is unglamorous and it is the difference between a shot that works in the app and one that works only in your editor.\n\n## The free path\n\nEverything here is a crop, a guide and a rectangle. Photopea in a browser does all of it and opens PSDs. GIMP and Krita if you can install. On a phone, any editor with a crop tool and the ability to draw a box. The thinking is the expensive part and it costs nothing.\n\n---\n\nToday: take a still you intend to push in on. Draw the end frame on it. Decide whether the shot you actually want is the one inside that box."},"check":{"question":"A generated orbit around a seated figure looks like a flat cut-out sliding across a painted background. What was wrong with the still?","options":[{"text":"The still had only one plane of depth, so the move asked the model to reveal a space the frame never contained.","correct":true},{"text":"The orbit was too fast for the model's motion range, and a slower rotation would have preserved the geometry of the scene properly.","correct":false,"tempting":"Treats a missing-information problem as a speed problem; a slow orbit reveals the same absent space more gradually."},{"text":"The subject was not centred, so the orbit rotated around the wrong point and the parallax was computed against the wrong axis.","correct":false,"tempting":"Assumes an explicit rotation axis exists; the model has no geometric pivot to get right or wrong."},{"text":"The background was out of focus, and blurred regions carry too little signal for the attention layers to keep them consistent across the move.","correct":false,"tempting":"Blames defocus, but a sharp single-plane backdrop orbits just as badly: the problem is the absence of recession, not of detail."}]},"prev":{"slug":"start-from-a-still","title":"The still frame is where the decisions belong","url":"https://addaly.com/learn/ai-video/start-from-a-still"},"next":{"slug":"lighting-the-still","title":"One light direction per scene, written down","url":"https://addaly.com/learn/ai-video/lighting-the-still"},"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) 'Compose for the frame you will end on', in Video with AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-video/composition-that-survives-motion (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Compose for the frame you will end on","courseTitle":"Video with AI","module":{"index":2,"slug":"the-first-frame","title":"The first frame, and everything decided in it"},"answer":null,"keyIdea":"A camera move changes the framing, so the still has to be composed for where the shot ends rather than where it starts, and every edge of the frame is a place where the model has to invent what enters.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"the-still-is-the-first-frame-not-the-shot","runs":[{"text":"The still is the first frame, not the shot"}]},{"t":"p","runs":[{"text":"People compose a beautiful still, add a slow push-in, and are disappointed. The push-in cropped fifteen per cent off every edge, so the composition they liked is gone by second three and what remains is a tighter, worse version of it."}]},{"t":"p","runs":[{"text":"Compose for the end of the move. If the shot pushes in, start looser than feels right. If it pulls out, start tighter. If it tilts up, leave the headroom you will need at the top. This is ordinary camera-operator thinking and it transfers exactly."}]},{"t":"p","runs":[{"text":"A practical way to do it: decide the move first, then open the still in any image editor and draw a rectangle where the frame will end up. If the composition inside that rectangle does not work, the still is wrong, not the prompt."}]},{"t":"h2","id":"headroom-lead-room-and-why-models-get-them-wrong","runs":[{"text":"Headroom, lead room, and why models get them wrong"}]},{"t":"p","runs":[{"text":"Two classical rules the model has no opinion about:"}]},{"t":"ul","items":[[{"text":"Headroom.","b":true},{"text":" The gap above the head. Too much and the subject sinks; too little and they feel cramped. Models trained on centred stock footage default to too much."}],[{"text":"Lead room.","b":true},{"text":" A person looking or moving to the left needs space on the left to look into. Without it the frame feels like it is pushing them into the edge."}]]},{"t":"p","runs":[{"text":"The model will not fix either. It animates the composition it is given. Fix them in the still, where a crop costs nothing."}]},{"t":"h2","id":"every-edge-is-a-place-things-come-from","runs":[{"text":"Every edge is a place things come from"}]},{"t":"p","runs":[{"text":"This is the composition rule specific to generative video, and it does not exist in still photography."}]},{"t":"p","runs":[{"text":"When the camera moves, new picture has to enter at the leading edge. The model invents it. It has no idea what was outside your frame, because nothing was: the still ends where it ends."}]},{"t":"p","runs":[{"text":"So a pan to the right invents everything on the right. A pull-out invents a ring of new scene on all four sides. A tilt up invents a ceiling or a sky."}]},{"t":"p","runs":[{"text":"What follows:"}]},{"t":"ul","items":[[{"text":"Put the difficult content in the middle.","b":true},{"text":" Anything with structure (a building edge, a bookshelf, a patterned floor) should not be sitting half in and half out of frame, because the half that is outside will be invented and will not match."}],[{"text":"Prefer moves that reveal simple things.","b":true},{"text":" Tilting up into plain sky works. Tilting up into an ornate ceiling produces melted ornament."}],[{"text":"A push-in invents nothing.","b":true},{"text":" It only crops. That is why it is the safest move in the category and why models default to it."}]]},{"t":"h2","id":"orbits-and-the-flat-backdrop-problem","runs":[{"text":"Orbits and the flat-backdrop problem"}]},{"t":"p","runs":[{"text":"An orbit or a parallax move asks the model to show the scene from an angle it was never given. If the still is a subject standing in front of an out-of-focus wall, there is no space behind them to reveal, so the model invents one as it goes and the shot reads as a cardboard cut-out sliding across a painted flat."}]},{"t":"p","runs":[{"text":"The fix is in the still: give the space depth before you ask the camera to move through it. Foreground element, mid-ground subject, background with real recession. A doorway, a table edge, a lamp in front, a corridor behind. A still with three planes orbits convincingly. A still with one does not, regardless of the model."}]},{"t":"h2","id":"aspect-ratio-is-a-composition-decision-not-an-export-setting","runs":[{"text":"Aspect ratio is a composition decision, not an export setting"}]},{"t":"p","runs":[{"text":"Match the model’s native ratio when you make the still: this was covered before and is worth repeating because it is the most common waste. But there is a second point."}]},{"t":"p","runs":[{"text":"If you need both a 16:9 and a 9:16 version of a piece, you cannot crop one from the other and keep the composition. A vertical reframe of a landscape shot either loses the sides or scales everything up until it is soft."}]},{"t":"p","runs":[{"text":"Two honest approaches. Either compose the landscape still with a vertical safe area in mind, keeping everything essential inside a centre column, which costs you the sides as usable composition, or generate the shot twice, once in each ratio, from two stills. The second costs more money and looks considerably better, and for anything with a subject it is usually the right call."}]},{"t":"h2","id":"safe-areas-still","runs":[{"text":"Safe areas, still"}]},{"t":"p","runs":[{"text":"If a platform overlays a caption, a profile picture and three buttons on the lower third of a vertical video, then the lower third of your composition is not yours. Check where the interface sits on the platform you are delivering to, and keep faces and text out of it. This is unglamorous and it is the difference between a shot that works in the app and one that works only in your editor."}]},{"t":"h2","id":"the-free-path","runs":[{"text":"The free path"}],"role":"free"},{"t":"p","runs":[{"text":"Everything here is a crop, a guide and a rectangle. Photopea in a browser does all of it and opens PSDs. GIMP and Krita if you can install. On a phone, any editor with a crop tool and the ability to draw a box. The thinking is the expensive part and it costs nothing."}]},{"t":"rule"},{"t":"p","runs":[{"text":"Today: take a still you intend to push in on. Draw the end frame on it. Decide whether the shot you actually want is the one inside that box."}]}],"check":{"question":[{"text":"A generated orbit around a seated figure looks like a flat cut-out sliding across a painted background. What was wrong with the still?"}],"options":[{"text":[{"text":"The still had only one plane of depth, so the move asked the model to reveal a space the frame never contained."}],"correct":true,"tempting":null},{"text":[{"text":"The orbit was too fast for the model’s motion range, and a slower rotation would have preserved the geometry of the scene properly."}],"correct":false,"tempting":[{"text":"Treats a missing-information problem as a speed problem; a slow orbit reveals the same absent space more gradually."}]},{"text":[{"text":"The subject was not centred, so the orbit rotated around the wrong point and the parallax was computed against the wrong axis."}],"correct":false,"tempting":[{"text":"Assumes an explicit rotation axis exists; the model has no geometric pivot to get right or wrong."}]},{"text":[{"text":"The background was out of focus, and blurred regions carry too little signal for the attention layers to keep them consistent across the move."}],"correct":false,"tempting":[{"text":"Blames defocus, but a sharp single-plane backdrop orbits just as badly: the problem is the absence of recession, not of detail."}]}],"order":[2,0,1,3],"why":null},"prev":{"slug":"start-from-a-still","title":"The still frame is where the decisions belong","minutes":9},"next":{"slug":"lighting-the-still","title":"One light direction per scene, written down","minutes":9}}}