{"course":{"slug":"ai-video","title":"Video with AI","url":"https://addaly.com/learn/ai-video","level":"intermediate","track":"create"},"lesson":{"slug":"building-a-character-sheet","title":"The document you build before you generate anything","url":"https://addaly.com/learn/ai-video/building-a-character-sheet","position":31,"of":66,"minutes":9,"answer":"A character sheet is a fixed set of images of one person under one light in one wardrobe, and it works because every later first frame is built from it rather than described afresh.","keyIdea":"A character sheet is a fixed set of images of one person under one light in one wardrobe, and it works because every later first frame is built from it rather than described afresh."},"body":{"format":"markdown","text":"## What it is, and what it is not\n\nA character sheet is a small set of images of your character, made once, that every shot in the project is built from. Animation studios have used them for a century. They exist because a description is not a specification, and a drawing is.\n\nIt is not a mood board and not a set of nice pictures. It is a reference document with a job: any first frame you need later should be constructible from it, by generation with reference, by compositing, or by hand.\n\n## The minimum sheet\n\nSix images. Fewer and you will be describing rather than referencing.\n\n1. **Frontal head and shoulders**, neutral expression, eyes to camera.\n2. **Three-quarter**, the most useful angle for almost every shot.\n3. **Profile.**\n4. **Full body, front**, in the exact wardrobe, showing footwear.\n5. **Full body, back.** Needed for over-the-shoulder and walking-away shots, which are the ones you will lean on hardest.\n6. **Medium shot in context**: the character in the actual location, at the actual light.\n\nAll six at the same light direction, the same quality, the same colour temperature. That is the part people get wrong: six lovely images under six different lights is not a character sheet, it is six characters.\n\n## Building it\n\nMake image one properly. Spend the attempts. Repair it by hand. This is the face for the whole project and everything else derives from it.\n\nThen produce the others from it:\n\n- With a reference feature, if your image tool has one: supply image one and ask for the three-quarter.\n- By generating a batch and keeping the ones that match, if it does not.\n- By hand, in Photopea or GIMP, for the composites: the correct head from one generation on the correct body from another.\n\nThe hand route feels like defeat the first time. It is faster than the twelfth generation, and it is deterministic, which no amount of re-rolling is.\n\n> Figure: Six images, and the order they are made in. Steps: 1. Frontal head and shoulders (Made properly, with the attempts spent and the flaws repaired by hand. This is the face for the whole project and everything else is derived from it.); 2. Three-quarter, then profile (Generated with image one as a reference, or composited from a batch of near-misses. Same light direction, same colour temperature, same wardrobe.); 3. Full body, front and back (The back is the one people skip, and over-the-shoulder and walking-away shots are the ones you will lean on hardest to keep the face out of frame.); 4. Medium shot in the location (The character in the actual set under the actual light, which is what most later first frames are built from.); 5. The written half (Age, hair, wardrobe colour, light direction, and which wrist the bangle is on. Models mirror freely and nobody notices until two shots sit side by side.). Six lovely images under six different lights is not a character sheet, it is six characters. Put all of them on one canvas at the same size and fix every mismatch before you generate a second of video, because a mismatch here becomes a mismatch in every shot derived from it.\n\n## What goes in the written half\n\nThe sheet is images plus a short text file, because some things are easier to check as words than to read off a picture:\n\n```\nPRIYA\nage ~34, medium-brown skin, short black hair tucked behind left ear\nsmall mole above right eyebrow\nWARDROBE: olive green cotton kurta, sleeves to elbow,\n  dark grey trousers, brown leather sandals, thin silver bangle LEFT wrist\nLIGHT: soft key from camera left, ~45 deg, slightly above eye level,\n  warm (window), low contrast\nPROPS: canvas shoulder bag, mustard, worn on RIGHT shoulder\n```\n\nNote the sides. Left and right are where continuity dies, because a model will mirror things freely and you will not notice until two shots are next to each other.\n\nThis file goes into every still prompt for the character, in some compressed form. It is the thing you paste.\n\n## Design the character to be reproducible\n\nA point that saves more time than any technique. Some characters are far easier to hold than others, and you usually get to choose.\n\nEasier: distinctive silhouette, strong single hair shape, one saturated wardrobe colour, one memorable accessory, no fine pattern, no small jewellery, no glasses.\n\nHarder: subtle features, elaborate patterned clothing, long loose hair, fine jewellery, glasses, anything with text on it.\n\nGlasses deserve a specific mention: thin frames sit near the autoencoder's resolution floor and flicker, and the reflections on the lenses drift. If a character can be written without glasses, write them without glasses.\n\nThis is not a compromise imposed by the tool. It is costume design, and costume designers have always chosen silhouettes that read at distance and colours that survive the grade.\n\n## Checking the sheet before you commit\n\nPut all six images on one canvas, at the same size, and look at them together. Specifically check:\n\n- Is the light coming from the same side in every one?\n- Is the wardrobe the same colour, not merely a similar colour?\n- Is the accessory on the same wrist, the same shoulder?\n- Does the hair part on the same side?\n\nAny mismatch here becomes a mismatch in every shot that derives from that image. Fixing it now is one regeneration; fixing it later is a scene.\n\n## The free path\n\nA browser image tool with a daily allowance for the generations, Photopea for the composites and repairs, and a text file. Total cost zero, total time a couple of hours for a character you will use for a whole project.\n\n---\n\nToday: build the six images for one character, put them on a single canvas side by side, and fix every mismatch before you generate one second of video."},"check":{"question":"An artist produces six beautiful reference images of a character, each with excellent lighting suited to its own composition. Why is this not yet a usable character sheet?","options":[{"text":"Six images is too few to condition a reference feature reliably; fifteen to thirty are needed for the model to generalise the identity.","correct":false,"tempting":"Quotes the dataset size for training a character model, which is a different technique with different requirements."},{"text":"Reference features need consistent framing rather than consistent lighting, so the images should all be shot at the same distance instead.","correct":false,"tempting":"Picks a genuine but secondary property; varied framing across the sheet is useful, varied lighting is not."},{"text":"The differing light means each image specifies a different key direction, so shots derived from them will not cut together.","correct":true},{"text":"Images generated separately carry different seeds, and without a shared seed the underlying facial geometry will differ between them in ways that accumulate across a sequence.","correct":false,"tempting":"Attributes identity to the seed; a shared seed with a different prompt produces a different face, and a different seed with a strong reference can produce the same one."}]},"prev":{"slug":"consistency-across-shots","title":"A cut breaks on the light before it breaks on the face","url":"https://addaly.com/learn/ai-video/consistency-across-shots"},"next":{"slug":"reference-and-element-conditioning","title":"What a reference image is actually doing","url":"https://addaly.com/learn/ai-video/reference-and-element-conditioning"},"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 document you build before you generate anything', in Video with AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-video/building-a-character-sheet (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 document you build before you generate anything","courseTitle":"Video with AI","module":{"index":4,"slug":"continuity-and-character","title":"Making separate shots belong to each other"},"answer":null,"keyIdea":"A character sheet is a fixed set of images of one person under one light in one wardrobe, and it works because every later first frame is built from it rather than described afresh.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"flow","steps":[{"label":"Frontal head and shoulders","detail":"Made properly, with the attempts spent and the flaws repaired by hand. This is the face for the whole project and everything else is derived from it."},{"label":"Three-quarter, then profile","detail":"Generated with image one as a reference, or composited from a batch of near-misses. Same light direction, same colour temperature, same wardrobe."},{"label":"Full body, front and back","detail":"The back is the one people skip, and over-the-shoulder and walking-away shots are the ones you will lean on hardest to keep the face out of frame."},{"label":"Medium shot in the location","detail":"The character in the actual set under the actual light, which is what most later first frames are built from."},{"label":"The written half","detail":"Age, hair, wardrobe colour, light direction, and which wrist the bangle is on. Models mirror freely and nobody notices until two shots sit side by side."}],"title":"Six images, and the order they are made in","caption":"Six lovely images under six different lights is not a character sheet, it is six characters. Put all of them on one canvas at the same size and fix every mismatch before you generate a second of video, because a mismatch here becomes a mismatch in every shot derived from it."},"alt":"Figure: Six images, and the order they are made in. Steps: 1. Frontal head and shoulders (Made properly, with the attempts spent and the flaws repaired by hand. This is the face for the whole project and everything else is derived from it.); 2. Three-quarter, then profile (Generated with image one as a reference, or composited from a batch of near-misses. Same light direction, same colour temperature, same wardrobe.); 3. Full body, front and back (The back is the one people skip, and over-the-shoulder and walking-away shots are the ones you will lean on hardest to keep the face out of frame.); 4. Medium shot in the location (The character in the actual set under the actual light, which is what most later first frames are built from.); 5. The written half (Age, hair, wardrobe colour, light direction, and which wrist the bangle is on. Models mirror freely and nobody notices until two shots sit side by side.). Six lovely images under six different lights is not a character sheet, it is six characters. Put all of them on one canvas at the same size and fix every mismatch before you generate a second of video, because a mismatch here becomes a mismatch in every shot derived from it."}},"blocks":[{"t":"h2","id":"what-it-is-and-what-it-is-not","runs":[{"text":"What it is, and what it is not"}]},{"t":"p","runs":[{"text":"A character sheet is a small set of images of your character, made once, that every shot in the project is built from. Animation studios have used them for a century. They exist because a description is not a specification, and a drawing is."}]},{"t":"p","runs":[{"text":"It is not a mood board and not a set of nice pictures. It is a reference document with a job: any first frame you need later should be constructible from it, by generation with reference, by compositing, or by hand."}]},{"t":"h2","id":"the-minimum-sheet","runs":[{"text":"The minimum sheet"}]},{"t":"p","runs":[{"text":"Six images. Fewer and you will be describing rather than referencing."}]},{"t":"ol","start":1,"items":[[{"text":"Frontal head and shoulders","b":true},{"text":", neutral expression, eyes to camera."}],[{"text":"Three-quarter","b":true},{"text":", the most useful angle for almost every shot."}],[{"text":"Profile.","b":true}],[{"text":"Full body, front","b":true},{"text":", in the exact wardrobe, showing footwear."}],[{"text":"Full body, back.","b":true},{"text":" Needed for over-the-shoulder and walking-away shots, which are the ones you will lean on hardest."}],[{"text":"Medium shot in context","b":true},{"text":": the character in the actual location, at the actual light."}]]},{"t":"p","runs":[{"text":"All six at the same light direction, the same quality, the same colour temperature. That is the part people get wrong: six lovely images under six different lights is not a character sheet, it is six characters."}]},{"t":"h2","id":"building-it","runs":[{"text":"Building it"}]},{"t":"p","runs":[{"text":"Make image one properly. Spend the attempts. Repair it by hand. This is the face for the whole project and everything else derives from it."}]},{"t":"p","runs":[{"text":"Then produce the others from it:"}]},{"t":"ul","items":[[{"text":"With a reference feature, if your image tool has one: supply image one and ask for the three-quarter."}],[{"text":"By generating a batch and keeping the ones that match, if it does not."}],[{"text":"By hand, in Photopea or GIMP, for the composites: the correct head from one generation on the correct body from another."}]]},{"t":"p","runs":[{"text":"The hand route feels like defeat the first time. It is faster than the twelfth generation, and it is deterministic, which no amount of re-rolling is."}]},{"t":"h2","id":"what-goes-in-the-written-half","runs":[{"text":"What goes in the written half"}]},{"t":"p","runs":[{"text":"The sheet is images plus a short text file, because some things are easier to check as words than to read off a picture:"}]},{"t":"code","lang":"","label":"Text","text":"PRIYA\nage ~34, medium-brown skin, short black hair tucked behind left ear\nsmall mole above right eyebrow\nWARDROBE: olive green cotton kurta, sleeves to elbow,\n  dark grey trousers, brown leather sandals, thin silver bangle LEFT wrist\nLIGHT: soft key from camera left, ~45 deg, slightly above eye level,\n  warm (window), low contrast\nPROPS: canvas shoulder bag, mustard, worn on RIGHT shoulder"},{"t":"p","runs":[{"text":"Note the sides. Left and right are where continuity dies, because a model will mirror things freely and you will not notice until two shots are next to each other."}]},{"t":"p","runs":[{"text":"This file goes into every still prompt for the character, in some compressed form. It is the thing you paste."}]},{"t":"h2","id":"design-the-character-to-be-reproducible","runs":[{"text":"Design the character to be reproducible"}]},{"t":"p","runs":[{"text":"A point that saves more time than any technique. Some characters are far easier to hold than others, and you usually get to choose."}]},{"t":"p","runs":[{"text":"Easier: distinctive silhouette, strong single hair shape, one saturated wardrobe colour, one memorable accessory, no fine pattern, no small jewellery, no glasses."}]},{"t":"p","runs":[{"text":"Harder: subtle features, elaborate patterned clothing, long loose hair, fine jewellery, glasses, anything with text on it."}]},{"t":"p","runs":[{"text":"Glasses deserve a specific mention: thin frames sit near the autoencoder’s resolution floor and flicker, and the reflections on the lenses drift. If a character can be written without glasses, write them without glasses."}]},{"t":"p","runs":[{"text":"This is not a compromise imposed by the tool. It is costume design, and costume designers have always chosen silhouettes that read at distance and colours that survive the grade."}]},{"t":"h2","id":"checking-the-sheet-before-you-commit","runs":[{"text":"Checking the sheet before you commit"}]},{"t":"p","runs":[{"text":"Put all six images on one canvas, at the same size, and look at them together. Specifically check:"}]},{"t":"ul","items":[[{"text":"Is the light coming from the same side in every one?"}],[{"text":"Is the wardrobe the same colour, not merely a similar colour?"}],[{"text":"Is the accessory on the same wrist, the same shoulder?"}],[{"text":"Does the hair part on the same side?"}]]},{"t":"p","runs":[{"text":"Any mismatch here becomes a mismatch in every shot that derives from that image. Fixing it now is one regeneration; fixing it later is a scene."}]},{"t":"h2","id":"the-free-path","runs":[{"text":"The free path"}],"role":"free"},{"t":"p","runs":[{"text":"A browser image tool with a daily allowance for the generations, Photopea for the composites and repairs, and a text file. Total cost zero, total time a couple of hours for a character you will use for a whole project."}]},{"t":"rule"},{"t":"p","runs":[{"text":"Today: build the six images for one character, put them on a single canvas side by side, and fix every mismatch before you generate one second of video."}]}],"check":{"question":[{"text":"An artist produces six beautiful reference images of a character, each with excellent lighting suited to its own composition. Why is this not yet a usable character sheet?"}],"options":[{"text":[{"text":"Six images is too few to condition a reference feature reliably; fifteen to thirty are needed for the model to generalise the identity."}],"correct":false,"tempting":[{"text":"Quotes the dataset size for training a character model, which is a different technique with different requirements."}]},{"text":[{"text":"Reference features need consistent framing rather than consistent lighting, so the images should all be shot at the same distance instead."}],"correct":false,"tempting":[{"text":"Picks a genuine but secondary property; varied framing across the sheet is useful, varied lighting is not."}]},{"text":[{"text":"The differing light means each image specifies a different key direction, so shots derived from them will not cut together."}],"correct":true,"tempting":null},{"text":[{"text":"Images generated separately carry different seeds, and without a shared seed the underlying facial geometry will differ between them in ways that accumulate across a sequence."}],"correct":false,"tempting":[{"text":"Attributes identity to the seed; a shared seed with a different prompt produces a different face, and a different seed with a strong reference can produce the same one."}]}],"order":[3,0,1,2],"why":null},"prev":{"slug":"consistency-across-shots","title":"A cut breaks on the light before it breaks on the face","minutes":10},"next":{"slug":"reference-and-element-conditioning","title":"What a reference image is actually doing","minutes":9}}}