{"course":{"slug":"ai-video","title":"Video with AI","url":"https://addaly.com/learn/ai-video","level":"intermediate","track":"create"},"lesson":{"slug":"wardrobe-props-and-colour-continuity","title":"The blue that is not quite the same blue","url":"https://addaly.com/learn/ai-video/wardrobe-props-and-colour-continuity","position":35,"of":66,"minutes":8,"answer":"Audiences detect wardrobe and prop discontinuity faster than facial drift because clothing is a large area of uniform colour.","keyIdea":"Audiences detect wardrobe and prop discontinuity faster than facial drift because clothing is a large area of uniform colour, and colour differences of a few per cent across a large area are exactly what human vision is built to notice."},"body":{"format":"markdown","text":"## Why clothes break cuts\n\nA face is small, complex and read by specialised machinery. A shirt is large, flat and read by the general colour system, which is extremely good at comparing two large uniform areas.\n\nSo a shirt that shifts from a slightly greenish blue to a slightly purplish blue between shots is noticed immediately, even by somebody who could not tell you what changed. \"Something feels off\" is the usual report, and the face gets blamed when the face was fine.\n\nThis is the most under-rated continuity risk in generative video, and it is also one of the most fixable.\n\n> Figure: Which wardrobe survives a cut. A two by two grid, rows One strong flat colour and Fine pattern or thin metal, columns Inside one shot and Across a cut: One strong flat colour and Inside one shot: Holds; One strong flat colour and Across a cut: Any mismatch is catchable; Fine pattern or thin metal and Inside one shot: Churns and changes count; Fine pattern or thin metal and Across a cut: Reads as a different garment. A shirt is a large area of uniform colour read by the general colour system, which is extremely good at comparing two large flat areas. A face is small and complex. That is why wardrobe drift is noticed first and why the face gets blamed for it.\n\n## Name the colour, do not describe it\n\n\"A blue shirt\" is a family. The model picks a member of it each time.\n\nWrite the colour as narrowly as language allows: \"a deep indigo cotton shirt, almost navy\", \"a mustard-yellow kurta\", \"a faded brick-red jacket\". Narrow descriptions land in a smaller region and vary less between samples.\n\nBetter still, do not rely on language. Put the garment in the still by compositing or by deriving each first frame from the character sheet, where the colour is already fixed as pixels rather than as a request.\n\n## Design the wardrobe to be holdable\n\nThe same logic as character design, and it is legitimate costume work:\n\n- **One strong colour per character.** It reads at distance, it survives the grade, and it makes mismatches obvious enough that you catch them.\n- **Avoid fine patterns.** Stripes, checks and small florals are below the encoder's resolution and will churn, warp and change count. A large bold pattern survives; a small busy one does not.\n- **Avoid text on clothing.** Same reason as text anywhere.\n- **Keep accessories few and large.** A single wide bangle holds. A stack of thin ones becomes a different stack every shot.\n- **Decide the side.** Bag on which shoulder, watch on which wrist, hair parted which way. Write it down, because models mirror freely.\n\n## Props that have to persist\n\nAny object a character carries through several shots is a continuity liability, for the reason established earlier: there is no object memory, and an item that leaves frame and returns has been re-sampled.\n\nThree responses:\n\n- **Reduce the number of persistent props to as close to one as the story allows.**\n- **Keep the prop in frame** rather than letting it exit and return.\n- **Composite it.** If the prop is specific (a particular book, a particular phone, the client's product) it goes in by hand, which was the conclusion of the compositing lesson and applies here unchanged.\n\n## Catching mismatches before you generate the clip\n\nThe check takes two minutes and it is the whole defence.\n\nPut every first frame for a scene on one canvas, at the same size, side by side. Then look at:\n\n- The garment colour as a block. Squint, or apply a heavy blur to all of them at once: blurring removes the detail and leaves the colour, which is exactly what you are comparing.\n- Which side the accessory is on.\n- The hair shape.\n- The prop's presence and position.\n\nDoing this in Photopea takes as long as opening the files. Doing it after generating costs you the generations.\n\n## What the grade can and cannot rescue\n\nA warm-cool mismatch across the whole frame is fixable: a shot-match in Resolve's free tier pulls one shot toward another in a minute, and a single grade across the timeline unifies the rest.\n\nA garment that is a different colour while the rest of the frame matches is not fixable that way, because a global correction would break everything else. It needs a secondary (isolating that colour range and shifting it alone) which Resolve's free tier does support and which is fiddly, imprecise on moving subjects, and sometimes the only thing standing between you and a reshoot.\n\nThe better answer is not to need it.\n\n## The one that always catches people\n\nMirroring. A model will happily flip a composition, and the character's bag moves to the other shoulder while everything else looks fine. It is invisible in a single shot and glaring across a cut.\n\nCheck sides explicitly, every time. It is the single highest-yield item on the list.\n\n---\n\nToday: put every first frame for one scene side by side, apply a heavy blur to all of them, and compare the wardrobe colours as blocks."},"check":{"question":"Why does a slightly mismatched shirt colour between two shots get noticed faster than a slightly mismatched face?","options":[{"text":"Clothing occupies a large uniform area, and comparing two large flat colours is something ordinary colour vision does extremely well.","correct":true},{"text":"Faces are processed by specialised machinery that tolerates small variation as natural change in expression and lighting.","correct":false,"tempting":"Inverts the known asymmetry: face-processing machinery is unusually intolerant of small geometric change, not tolerant of it."},{"text":"Colour drift between generations is larger in magnitude than geometric drift, so the shirt simply changes more than the face does.","correct":false,"tempting":"Assumes the drift magnitudes differ; both drift by similar proportions, and it is detection rather than magnitude that separates them."},{"text":"The autoencoder preserves large flat regions more faithfully than complex ones, so any colour change in the garment is rendered with full fidelity while facial change is smoothed away by the decoder.","correct":false,"tempting":"Correct that flat regions encode well, but the decoder does not smooth facial drift away: it reproduces it, and viewers still notice it more slowly."}]},"prev":{"slug":"location-and-set-continuity","title":"Building a room that stays the same room","url":"https://addaly.com/learn/ai-video/location-and-set-continuity"},"next":{"slug":"post-hoc-identity-repair","title":"Fixing a face after the clip exists","url":"https://addaly.com/learn/ai-video/post-hoc-identity-repair"},"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 blue that is not quite the same blue', in Video with AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-video/wardrobe-props-and-colour-continuity (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 blue that is not quite the same blue","courseTitle":"Video with AI","module":{"index":4,"slug":"continuity-and-character","title":"Making separate shots belong to each other"},"answer":null,"keyIdea":"Audiences detect wardrobe and prop discontinuity faster than facial drift because clothing is a large area of uniform colour, and colour differences of a few per cent across a large area are exactly what human vision is built to notice.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"cols":["Inside one shot","Across a cut"],"rows":["One strong flat colour","Fine pattern or thin metal"],"type":"matrix","cells":[["Holds","Any mismatch is catchable"],["Churns and changes count","Reads as a different garment"]],"title":"Which wardrobe survives a cut","caption":"A shirt is a large area of uniform colour read by the general colour system, which is extremely good at comparing two large flat areas. A face is small and complex. That is why wardrobe drift is noticed first and why the face gets blamed for it.","cellLabels":[["a mustard kurta stays a mustard kurta","you see it the moment two stills sit side by side"],["stripes warp; a stack of thin bangles is a new stack","blame lands on the face, and the face was fine"]]},"alt":"Figure: Which wardrobe survives a cut. A two by two grid, rows One strong flat colour and Fine pattern or thin metal, columns Inside one shot and Across a cut: One strong flat colour and Inside one shot: Holds; One strong flat colour and Across a cut: Any mismatch is catchable; Fine pattern or thin metal and Inside one shot: Churns and changes count; Fine pattern or thin metal and Across a cut: Reads as a different garment. A shirt is a large area of uniform colour read by the general colour system, which is extremely good at comparing two large flat areas. A face is small and complex. That is why wardrobe drift is noticed first and why the face gets blamed for it."}},"blocks":[{"t":"h2","id":"why-clothes-break-cuts","runs":[{"text":"Why clothes break cuts"}]},{"t":"p","runs":[{"text":"A face is small, complex and read by specialised machinery. A shirt is large, flat and read by the general colour system, which is extremely good at comparing two large uniform areas."}]},{"t":"p","runs":[{"text":"So a shirt that shifts from a slightly greenish blue to a slightly purplish blue between shots is noticed immediately, even by somebody who could not tell you what changed. “Something feels off” is the usual report, and the face gets blamed when the face was fine."}]},{"t":"p","runs":[{"text":"This is the most under-rated continuity risk in generative video, and it is also one of the most fixable."}]},{"t":"h2","id":"name-the-colour-do-not-describe-it","runs":[{"text":"Name the colour, do not describe it"}]},{"t":"p","runs":[{"text":"“A blue shirt” is a family. The model picks a member of it each time."}]},{"t":"p","runs":[{"text":"Write the colour as narrowly as language allows: “a deep indigo cotton shirt, almost navy”, “a mustard-yellow kurta”, “a faded brick-red jacket”. Narrow descriptions land in a smaller region and vary less between samples."}]},{"t":"p","runs":[{"text":"Better still, do not rely on language. Put the garment in the still by compositing or by deriving each first frame from the character sheet, where the colour is already fixed as pixels rather than as a request."}]},{"t":"h2","id":"design-the-wardrobe-to-be-holdable","runs":[{"text":"Design the wardrobe to be holdable"}]},{"t":"p","runs":[{"text":"The same logic as character design, and it is legitimate costume work:"}]},{"t":"ul","items":[[{"text":"One strong colour per character.","b":true},{"text":" It reads at distance, it survives the grade, and it makes mismatches obvious enough that you catch them."}],[{"text":"Avoid fine patterns.","b":true},{"text":" Stripes, checks and small florals are below the encoder’s resolution and will churn, warp and change count. A large bold pattern survives; a small busy one does not."}],[{"text":"Avoid text on clothing.","b":true},{"text":" Same reason as text anywhere."}],[{"text":"Keep accessories few and large.","b":true},{"text":" A single wide bangle holds. A stack of thin ones becomes a different stack every shot."}],[{"text":"Decide the side.","b":true},{"text":" Bag on which shoulder, watch on which wrist, hair parted which way. Write it down, because models mirror freely."}]]},{"t":"h2","id":"props-that-have-to-persist","runs":[{"text":"Props that have to persist"}]},{"t":"p","runs":[{"text":"Any object a character carries through several shots is a continuity liability, for the reason established earlier: there is no object memory, and an item that leaves frame and returns has been re-sampled."}]},{"t":"p","runs":[{"text":"Three responses:"}]},{"t":"ul","items":[[{"text":"Reduce the number of persistent props to as close to one as the story allows.","b":true}],[{"text":"Keep the prop in frame","b":true},{"text":" rather than letting it exit and return."}],[{"text":"Composite it.","b":true},{"text":" If the prop is specific (a particular book, a particular phone, the client’s product) it goes in by hand, which was the conclusion of the compositing lesson and applies here unchanged."}]]},{"t":"h2","id":"catching-mismatches-before-you-generate-the-clip","runs":[{"text":"Catching mismatches before you generate the clip"}]},{"t":"p","runs":[{"text":"The check takes two minutes and it is the whole defence."}]},{"t":"p","runs":[{"text":"Put every first frame for a scene on one canvas, at the same size, side by side. Then look at:"}]},{"t":"ul","items":[[{"text":"The garment colour as a block. Squint, or apply a heavy blur to all of them at once: blurring removes the detail and leaves the colour, which is exactly what you are comparing."}],[{"text":"Which side the accessory is on."}],[{"text":"The hair shape."}],[{"text":"The prop’s presence and position."}]]},{"t":"p","runs":[{"text":"Doing this in Photopea takes as long as opening the files. Doing it after generating costs you the generations."}]},{"t":"h2","id":"what-the-grade-can-and-cannot-rescue","runs":[{"text":"What the grade can and cannot rescue"}]},{"t":"p","runs":[{"text":"A warm-cool mismatch across the whole frame is fixable: a shot-match in Resolve’s free tier pulls one shot toward another in a minute, and a single grade across the timeline unifies the rest."}]},{"t":"p","runs":[{"text":"A garment that is a different colour while the rest of the frame matches is not fixable that way, because a global correction would break everything else. It needs a secondary (isolating that colour range and shifting it alone) which Resolve’s free tier does support and which is fiddly, imprecise on moving subjects, and sometimes the only thing standing between you and a reshoot."}]},{"t":"p","runs":[{"text":"The better answer is not to need it."}]},{"t":"h2","id":"the-one-that-always-catches-people","runs":[{"text":"The one that always catches people"}]},{"t":"p","runs":[{"text":"Mirroring. A model will happily flip a composition, and the character’s bag moves to the other shoulder while everything else looks fine. It is invisible in a single shot and glaring across a cut."}]},{"t":"p","runs":[{"text":"Check sides explicitly, every time. It is the single highest-yield item on the list."}]},{"t":"rule"},{"t":"p","runs":[{"text":"Today: put every first frame for one scene side by side, apply a heavy blur to all of them, and compare the wardrobe colours as blocks."}]}],"check":{"question":[{"text":"Why does a slightly mismatched shirt colour between two shots get noticed faster than a slightly mismatched face?"}],"options":[{"text":[{"text":"Clothing occupies a large uniform area, and comparing two large flat colours is something ordinary colour vision does extremely well."}],"correct":true,"tempting":null},{"text":[{"text":"Faces are processed by specialised machinery that tolerates small variation as natural change in expression and lighting."}],"correct":false,"tempting":[{"text":"Inverts the known asymmetry: face-processing machinery is unusually intolerant of small geometric change, not tolerant of it."}]},{"text":[{"text":"Colour drift between generations is larger in magnitude than geometric drift, so the shirt simply changes more than the face does."}],"correct":false,"tempting":[{"text":"Assumes the drift magnitudes differ; both drift by similar proportions, and it is detection rather than magnitude that separates them."}]},{"text":[{"text":"The autoencoder preserves large flat regions more faithfully than complex ones, so any colour change in the garment is rendered with full fidelity while facial change is smoothed away by the decoder."}],"correct":false,"tempting":[{"text":"Correct that flat regions encode well, but the decoder does not smooth facial drift away: it reproduces it, and viewers still notice it more slowly."}]}],"order":[2,3,1,0],"why":null},"prev":{"slug":"location-and-set-continuity","title":"Building a room that stays the same room","minutes":9},"next":{"slug":"post-hoc-identity-repair","title":"Fixing a face after the clip exists","minutes":9}}}