{"course":{"slug":"ai-video","title":"Video with AI","url":"https://addaly.com/learn/ai-video","level":"intermediate","track":"create"},"lesson":{"slug":"shot-size-and-angle","title":"The two words that decide most of the frame","url":"https://addaly.com/learn/ai-video/shot-size-and-angle","position":22,"of":66,"minutes":9,"answer":"Shot size and camera angle were named explicitly in the captions the model trained on, so they steer more reliably than any other words, and they carry meaning an audience reads without noticing.","keyIdea":"Shot size and camera angle were named explicitly in the captions the model trained on, so they steer more reliably than any other words, and they carry meaning an audience reads without noticing."},"body":{"format":"markdown","text":"## The vocabulary, and why it works here\n\nAuto-captioners describe shot size and angle because both are visible facts about a frame. The words appeared millions of times, attached precisely to what they name. That makes them the most dependable instructions you have.\n\nThe standard sizes, from widest:\n\n- **Extreme wide.** The subject is small in a large space. Establishes where.\n- **Wide** or **long shot.** Full body, head to feet, with room around.\n- **Medium wide.** Knees up. The workhorse of dialogue scenes.\n- **Medium.** Waist up.\n- **Medium close-up.** Chest up. The interview shot.\n- **Close-up.** Shoulders and head.\n- **Extreme close-up.** Part of a face, or an object filling frame.\n\nAnd the angles:\n\n- **Eye level.** Neutral. The default and the one you should use unless you want something.\n- **Low angle**, looking up. Gives the subject stature.\n- **High angle**, looking down. Diminishes.\n- **Overhead** or top-down. Abstracts a scene into a pattern.\n- **Dutch** or canted, the horizon tilted. Unease. Use once per project at most.\n\n## What each size costs you in a generated clip\n\nThis is where film grammar and the machinery interact, and it is worth knowing before you choose.\n\n**Wide shots are cheap and safe.** The face is forty pixels across, so identity drift is invisible. Hands are too small to go wrong in a way anyone can see. Wides are the single most reliable shot in generative video and you should use more of them than you think.\n\n**Close-ups are expensive and risky.** Every pixel of the face is being scrutinised by the part of your viewer's brain that is best at faces. Skin texture, teeth and eyes are all near the autoencoder's resolution limit at that scale. Close-ups are where you spend attempts.\n\n**Extreme close-ups of objects are surprisingly good.** No face, no hands, large simple forms filling frame. A drop of water, a fabric weave, a hand-free detail. These are among the most reliable shots available and they are undervalued because they feel like filler. They are not filler; they are what a real editor cuts to.\n\n**Medium shots are the compromise that works.** Face large enough to read, small enough to survive. If a shot has to contain a person and has to hold, make it a medium.\n\n## Angle and the space behind the subject\n\nA low angle looks up and therefore shows ceiling or sky. A high angle shows floor. An overhead shows the ground plane and nothing else.\n\nEach of those is a region the still may not contain, and if the camera moves the model has to invent it. So angles interact with the composition rule from the previous module: pick the angle in the still, and if you then move the camera, know which new territory the move opens up.\n\nOverhead shots are worth a specific note. They are strong compositions, they hide faces, they hide hands, and the ground plane is usually simple. For a difficult subject, an overhead is often the shot that works when nothing else does.\n\n## What the size means to a viewer\n\nShot size is not decoration. It is how a film controls distance and attention.\n\nA sequence that stays wide keeps the audience outside, observing. A sequence that moves inward as it progresses pulls them in. Cutting from a wide directly to an extreme close-up is a jolt and can be exactly right. Cutting between two shots of nearly the same size on the same subject is a jump cut and reads as a mistake.\n\nThat last point has a direct consequence for how you plan generated shots. **Change the size by at least two steps between consecutive shots of the same subject**, or change the angle substantially. A wide to a medium wide is a stumble. A wide to a close-up is an edit.\n\nThis also happens to be the rule that hides identity drift, because a big size change gives the eye a new thing to read and less opportunity to compare faces. Grammar and machinery agree again.\n\n## Writing it\n\nPut it first in the instruction, because it is the strongest term:\n\n> Medium close-up, eye level, a woman at a kitchen table turns her head to look out of the window on the left. The camera does not move. Warm window light from camera left.\n\nSize, angle, subject verb, camera behaviour, light. Five clauses, no adjectives about mood, and every word of it is something an auto-captioner would have written.\n\n## The free path\n\nThis costs nothing. It is vocabulary and a decision. The most useful free exercise: take one still, generate it four times at four different described sizes, and confirm that your model responds to them. Some models reframe from the supplied still more willingly than others, and knowing which you have changes whether you specify size in the prompt or bake it into the frame.\n\n---\n\nToday: look at your shot list and mark the size of every shot. If two consecutive shots of the same subject are within one step of each other, change one."},"check":{"question":"A sequence keeps losing continuity on a character's face between shots. The shot list has: wide, medium wide, medium, medium close-up, close-up: a smooth progression inward. What change would help most?","options":[{"text":"Reverse the order so the sequence moves outward, since identity drift is less noticeable when the face is getting smaller rather than larger.","correct":false,"tempting":"Invents a directional asymmetry; drift is equally visible whichever way the progression runs."},{"text":"Make the size jumps larger, at least two steps between shots, so the eye gets new information and less chance to compare faces.","correct":true},{"text":"Hold every shot at the same size and let the angle carry the variation, so the face is rendered at a consistent scale throughout.","correct":false,"tempting":"Produces a sequence of jump cuts, and a constant scale makes the face easiest of all to compare between shots."},{"text":"Generate all five shots from a single still at successively tighter crops, so the face is literally identical in each one.","correct":false,"tempting":"Solves identity at the cost of every shot sharing one pose and one moment, which reads as a zoom rather than a scene."}]},"prev":{"slug":"directing-the-camera","title":"Cinematic, 8k, epic gets you the average of everything","url":"https://addaly.com/learn/ai-video/directing-the-camera"},"next":{"slug":"lens-and-focus-language","title":"Focal length, depth of field, and the words that carry them","url":"https://addaly.com/learn/ai-video/lens-and-focus-language"},"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) 'The two words that decide most of the frame', in Video with AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-video/shot-size-and-angle (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":"The two words that decide most of the frame","courseTitle":"Video with AI","module":{"index":3,"slug":"directing-the-shot","title":"Directing, not describing"},"answer":null,"keyIdea":"Shot size and camera angle were named explicitly in the captions the model trained on, so they steer more reliably than any other words, and they carry meaning an audience reads without noticing.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"the-vocabulary-and-why-it-works-here","runs":[{"text":"The vocabulary, and why it works here"}]},{"t":"p","runs":[{"text":"Auto-captioners describe shot size and angle because both are visible facts about a frame. The words appeared millions of times, attached precisely to what they name. That makes them the most dependable instructions you have."}]},{"t":"p","runs":[{"text":"The standard sizes, from widest:"}]},{"t":"ul","items":[[{"text":"Extreme wide.","b":true},{"text":" The subject is small in a large space. Establishes where."}],[{"text":"Wide","b":true},{"text":" or "},{"text":"long shot.","b":true},{"text":" Full body, head to feet, with room around."}],[{"text":"Medium wide.","b":true},{"text":" Knees up. The workhorse of dialogue scenes."}],[{"text":"Medium.","b":true},{"text":" Waist up."}],[{"text":"Medium close-up.","b":true},{"text":" Chest up. The interview shot."}],[{"text":"Close-up.","b":true},{"text":" Shoulders and head."}],[{"text":"Extreme close-up.","b":true},{"text":" Part of a face, or an object filling frame."}]]},{"t":"p","runs":[{"text":"And the angles:"}]},{"t":"ul","items":[[{"text":"Eye level.","b":true},{"text":" Neutral. The default and the one you should use unless you want something."}],[{"text":"Low angle","b":true},{"text":", looking up. Gives the subject stature."}],[{"text":"High angle","b":true},{"text":", looking down. Diminishes."}],[{"text":"Overhead","b":true},{"text":" or top-down. Abstracts a scene into a pattern."}],[{"text":"Dutch","b":true},{"text":" or canted, the horizon tilted. Unease. Use once per project at most."}]]},{"t":"h2","id":"what-each-size-costs-you-in-a-generated-clip","runs":[{"text":"What each size costs you in a generated clip"}]},{"t":"p","runs":[{"text":"This is where film grammar and the machinery interact, and it is worth knowing before you choose."}]},{"t":"p","runs":[{"text":"Wide shots are cheap and safe.","b":true},{"text":" The face is forty pixels across, so identity drift is invisible. Hands are too small to go wrong in a way anyone can see. Wides are the single most reliable shot in generative video and you should use more of them than you think."}]},{"t":"p","runs":[{"text":"Close-ups are expensive and risky.","b":true},{"text":" Every pixel of the face is being scrutinised by the part of your viewer’s brain that is best at faces. Skin texture, teeth and eyes are all near the autoencoder’s resolution limit at that scale. Close-ups are where you spend attempts."}]},{"t":"p","runs":[{"text":"Extreme close-ups of objects are surprisingly good.","b":true},{"text":" No face, no hands, large simple forms filling frame. A drop of water, a fabric weave, a hand-free detail. These are among the most reliable shots available and they are undervalued because they feel like filler. They are not filler; they are what a real editor cuts to."}]},{"t":"p","runs":[{"text":"Medium shots are the compromise that works.","b":true},{"text":" Face large enough to read, small enough to survive. If a shot has to contain a person and has to hold, make it a medium."}]},{"t":"h2","id":"angle-and-the-space-behind-the-subject","runs":[{"text":"Angle and the space behind the subject"}]},{"t":"p","runs":[{"text":"A low angle looks up and therefore shows ceiling or sky. A high angle shows floor. An overhead shows the ground plane and nothing else."}]},{"t":"p","runs":[{"text":"Each of those is a region the still may not contain, and if the camera moves the model has to invent it. So angles interact with the composition rule from the previous module: pick the angle in the still, and if you then move the camera, know which new territory the move opens up."}]},{"t":"p","runs":[{"text":"Overhead shots are worth a specific note. They are strong compositions, they hide faces, they hide hands, and the ground plane is usually simple. For a difficult subject, an overhead is often the shot that works when nothing else does."}]},{"t":"h2","id":"what-the-size-means-to-a-viewer","runs":[{"text":"What the size means to a viewer"}]},{"t":"p","runs":[{"text":"Shot size is not decoration. It is how a film controls distance and attention."}]},{"t":"p","runs":[{"text":"A sequence that stays wide keeps the audience outside, observing. A sequence that moves inward as it progresses pulls them in. Cutting from a wide directly to an extreme close-up is a jolt and can be exactly right. Cutting between two shots of nearly the same size on the same subject is a jump cut and reads as a mistake."}]},{"t":"p","runs":[{"text":"That last point has a direct consequence for how you plan generated shots. "},{"text":"Change the size by at least two steps between consecutive shots of the same subject","b":true},{"text":", or change the angle substantially. A wide to a medium wide is a stumble. A wide to a close-up is an edit."}]},{"t":"p","runs":[{"text":"This also happens to be the rule that hides identity drift, because a big size change gives the eye a new thing to read and less opportunity to compare faces. Grammar and machinery agree again."}]},{"t":"h2","id":"writing-it","runs":[{"text":"Writing it"}]},{"t":"p","runs":[{"text":"Put it first in the instruction, because it is the strongest term:"}]},{"t":"quote","runs":[{"text":"Medium close-up, eye level, a woman at a kitchen table turns her head to look out of the window on the left. The camera does not move. Warm window light from camera left."}]},{"t":"p","runs":[{"text":"Size, angle, subject verb, camera behaviour, light. Five clauses, no adjectives about mood, and every word of it is something an auto-captioner would have written."}]},{"t":"h2","id":"the-free-path","runs":[{"text":"The free path"}],"role":"free"},{"t":"p","runs":[{"text":"This costs nothing. It is vocabulary and a decision. The most useful free exercise: take one still, generate it four times at four different described sizes, and confirm that your model responds to them. Some models reframe from the supplied still more willingly than others, and knowing which you have changes whether you specify size in the prompt or bake it into the frame."}]},{"t":"rule"},{"t":"p","runs":[{"text":"Today: look at your shot list and mark the size of every shot. If two consecutive shots of the same subject are within one step of each other, change one."}]}],"check":{"question":[{"text":"A sequence keeps losing continuity on a character’s face between shots. The shot list has: wide, medium wide, medium, medium close-up, close-up: a smooth progression inward. What change would help most?"}],"options":[{"text":[{"text":"Reverse the order so the sequence moves outward, since identity drift is less noticeable when the face is getting smaller rather than larger."}],"correct":false,"tempting":[{"text":"Invents a directional asymmetry; drift is equally visible whichever way the progression runs."}]},{"text":[{"text":"Make the size jumps larger, at least two steps between shots, so the eye gets new information and less chance to compare faces."}],"correct":true,"tempting":null},{"text":[{"text":"Hold every shot at the same size and let the angle carry the variation, so the face is rendered at a consistent scale throughout."}],"correct":false,"tempting":[{"text":"Produces a sequence of jump cuts, and a constant scale makes the face easiest of all to compare between shots."}]},{"text":[{"text":"Generate all five shots from a single still at successively tighter crops, so the face is literally identical in each one."}],"correct":false,"tempting":[{"text":"Solves identity at the cost of every shot sharing one pose and one moment, which reads as a zoom rather than a scene."}]}],"order":[3,2,0,1],"why":null},"prev":{"slug":"directing-the-camera","title":"Cinematic, 8k, epic gets you the average of everything","minutes":9},"next":{"slug":"lens-and-focus-language","title":"Focal length, depth of field, and the words that carry them","minutes":9}}}