{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"slug":"crop-safe-composition","title":"Designing for the crop you will not control","url":"https://addaly.com/learn/ai-images/crop-safe-composition","position":14,"of":69,"minutes":8,"answer":"Plan the frame as concentric zones (a core that survives every crop, a middle that survives most.","keyIdea":"Plan the frame as concentric zones (a core that survives every crop, a middle that survives most, and an outer band that exists to be thrown away) because the image will be re-cropped by people and platforms who never see your original."},"body":{"format":"markdown","text":"## Your image will be cropped by strangers\n\nYou deliver a beautiful 4:5 frame. It appears as a 1:1 thumbnail in a feed, as a 16:9 card when shared to a messaging app, as a circular avatar somewhere, and with a 200-pixel strip of it behind a headline on the website. You are not consulted about any of this.\n\nThe answer used by broadcast and print for decades is **safe areas**, and it transfers directly.\n\n## Three zones\n\nThink of the frame as three nested regions.\n\n**The core**, the central square or thereabouts. Everything the image *means* lives here: the subject's face, the product, the gesture that carries the idea. This survives every crop anyone will apply.\n\n**The support band**, out to roughly 85% of the frame. Context that makes the image good but not essential: the table the product sits on, the second figure, the environment. Some crops keep it, some do not.\n\n**The sacrificial edge**, the outer 10 to 15%. Texture, gradient, background. Designed to be cut. Never put anything here that must be seen.\n\n> Figure: Three nested zones, from the middle outwards. Layers, top to bottom: The core: the central square, roughly (Face, product, the gesture. Survives every crop.); The support band: out to about 85% (Context that helps. Some crops keep it, some do not.); The sacrificial edge: the outer 10 to 15% (Texture and background. Designed to be thrown away.). Eyes forty pixels from the top edge are not bad luck. A 1:1 crop of a 4:5 frame takes them, and the composition was never crop-safe.\n\nIf your subject's eyes are 40 pixels from the top edge, a 1:1 crop of a 4:5 frame will cut them. That is not bad luck; it is a composition that was never crop-safe.\n\n## Getting a generator to leave you room\n\nGenerators pull toward filling the frame: the training data is full of well-composed photographs where the subject dominates. To get usable margin, ask for it explicitly, and in the language photography uses:\n\n- \"wide shot, subject small in frame, generous negative space above\"\n- \"centred composition, plain background, copy space at the top\"\n- \"full body with headroom, shot on a 35mm lens\"\n\nLens language does real work here. \"85mm portrait\" produces a tighter, more compressed frame; \"24mm wide\" gives you room and a more environmental feel. These are more reliable levers than the word \"space\", because the model has seen the lens described in thousands of captions alongside the framing it produces.\n\nThe other approach is to generate at a wider ratio than you need and treat the extra as margin by design. Render 1344 × 768 when you need 1200 × 630, and you have 144 pixels of horizontal slack and 138 of vertical to place the crop with, after the fact, where it actually works.\n\n## The type zone is part of the composition\n\nIf a headline goes across the top third, that third has two requirements: it must be quiet, and it must have enough contrast against white or black type to be legible. A busy sky with bright clouds fails the second even though it is \"empty\".\n\nGenerate for it. \"Plain overcast sky occupying the upper third\" gives you a place to set type. So does \"dark studio background, subject lit from the side, lower two thirds\". You are not decorating; you are building a surface that type can sit on, and the finishing block returns to the legibility arithmetic.\n\n## Check it before you deliver\n\nNinety seconds in any editor, including free ones:\n\n1. Open the image in GIMP, Krita or Photopea.\n2. Add guides at 1:1, 16:9 and 9:16 centred on the frame.\n3. Look at each. Is the subject intact? Is anything important cut? Does the 1:1 still read at 150 pixels?\n\nIf a crop fails, you have three options and they are not equal. Re-crop, which is free. Outpaint to add room, which takes a few minutes. Or regenerate, which costs the composition you had approved. Doing this check before the client sees anything means you only ever have the cheap version of the problem.\n\n## The limitation\n\nCrop-safe composition costs you the strongest version of the picture. A frame designed to survive five crops is, almost by definition, more centred and more conservative than one composed for a single format. That is a real trade, not a free technique.\n\nSo make it deliberately. A hero image that only ever appears in one place should be composed for that place and nothing else. Crop safety is for images that will travel: social assets, stock, anything a client's marketing team will re-use without asking. Applying it everywhere produces a portfolio of safe, centred, slightly dull pictures, and that is its own kind of failure."},"check":{"question":"A 4:5 social image keeps losing the subject's face when platforms crop it to a square thumbnail. What is the underlying composition error?","options":[{"text":"The image was generated at a non-native aspect ratio, so the model placed the subject inconsistently within the frame.","correct":false,"tempting":"Bucket mismatch produces duplicated or distorted anatomy rather than a subject positioned too near an edge; the framing here is a compositional choice, not an artefact."},{"text":"Important content was placed in the band that only survives the original ratio, instead of the central core.","correct":true},{"text":"The resolution is too low, so the platform's automatic crop cannot identify the subject and defaults to a centred rectangle.","correct":false,"tempting":"Invents a mechanism: platform crops are usually fixed or face-detected, and neither behaves differently because an image has more pixels in it."},{"text":"The delivered file should have been a square from the beginning, because any image reused across several placements ought to be generated in the most neutral format available.","correct":false,"tempting":"Solves it by giving up the 4:5 frame entirely; the point of safe areas is to serve several crops from one strong composition rather than defaulting to the weakest one."}]},"prev":{"slug":"aspect-ratio-and-the-native-bucket","title":"The ratios the model was trained on, and what happens outside them","url":"https://addaly.com/learn/ai-images/aspect-ratio-and-the-native-bucket"},"next":{"slug":"the-reference-pack","title":"Six references, each with a job","url":"https://addaly.com/learn/ai-images/the-reference-pack"},"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) 'Designing for the crop you will not control', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/crop-safe-composition (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":"Designing for the crop you will not control","courseTitle":"Image Generation, in Practice","module":{"index":2,"slug":"the-brief","title":"From a request to a specification you can generate against"},"answer":null,"keyIdea":"Plan the frame as concentric zones (a core that survives every crop, a middle that survives most, and an outer band that exists to be thrown away) because the image will be re-cropped by people and platforms who never see your original.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"stack","title":"Three nested zones, from the middle outwards","layers":[{"label":"The core: the central square, roughly","detail":"Face, product, the gesture. Survives every crop."},{"label":"The support band: out to about 85%","detail":"Context that helps. Some crops keep it, some do not."},{"label":"The sacrificial edge: the outer 10 to 15%","detail":"Texture and background. Designed to be thrown away."}],"caption":"Eyes forty pixels from the top edge are not bad luck. A 1:1 crop of a 4:5 frame takes them, and the composition was never crop-safe."},"alt":"Figure: Three nested zones, from the middle outwards. Layers, top to bottom: The core: the central square, roughly (Face, product, the gesture. Survives every crop.); The support band: out to about 85% (Context that helps. Some crops keep it, some do not.); The sacrificial edge: the outer 10 to 15% (Texture and background. Designed to be thrown away.). Eyes forty pixels from the top edge are not bad luck. A 1:1 crop of a 4:5 frame takes them, and the composition was never crop-safe."}},"blocks":[{"t":"h2","id":"your-image-will-be-cropped-by-strangers","runs":[{"text":"Your image will be cropped by strangers"}]},{"t":"p","runs":[{"text":"You deliver a beautiful 4:5 frame. It appears as a 1:1 thumbnail in a feed, as a 16:9 card when shared to a messaging app, as a circular avatar somewhere, and with a 200-pixel strip of it behind a headline on the website. You are not consulted about any of this."}]},{"t":"p","runs":[{"text":"The answer used by broadcast and print for decades is "},{"text":"safe areas","b":true},{"text":", and it transfers directly."}]},{"t":"h2","id":"three-zones","runs":[{"text":"Three zones"}]},{"t":"p","runs":[{"text":"Think of the frame as three nested regions."}]},{"t":"p","runs":[{"text":"The core","b":true},{"text":", the central square or thereabouts. Everything the image "},{"text":"means","i":true},{"text":" lives here: the subject’s face, the product, the gesture that carries the idea. This survives every crop anyone will apply."}]},{"t":"p","runs":[{"text":"The support band","b":true},{"text":", out to roughly 85% of the frame. Context that makes the image good but not essential: the table the product sits on, the second figure, the environment. Some crops keep it, some do not."}]},{"t":"p","runs":[{"text":"The sacrificial edge","b":true},{"text":", the outer 10 to 15%. Texture, gradient, background. Designed to be cut. Never put anything here that must be seen."}]},{"t":"p","runs":[{"text":"If your subject’s eyes are 40 pixels from the top edge, a 1:1 crop of a 4:5 frame will cut them. That is not bad luck; it is a composition that was never crop-safe."}]},{"t":"h2","id":"getting-a-generator-to-leave-you-room","runs":[{"text":"Getting a generator to leave you room"}]},{"t":"p","runs":[{"text":"Generators pull toward filling the frame: the training data is full of well-composed photographs where the subject dominates. To get usable margin, ask for it explicitly, and in the language photography uses:"}]},{"t":"ul","items":[[{"text":"“wide shot, subject small in frame, generous negative space above”"}],[{"text":"“centred composition, plain background, copy space at the top”"}],[{"text":"“full body with headroom, shot on a 35mm lens”"}]]},{"t":"p","runs":[{"text":"Lens language does real work here. “85mm portrait” produces a tighter, more compressed frame; “24mm wide” gives you room and a more environmental feel. These are more reliable levers than the word “space”, because the model has seen the lens described in thousands of captions alongside the framing it produces."}]},{"t":"p","runs":[{"text":"The other approach is to generate at a wider ratio than you need and treat the extra as margin by design. Render 1344 × 768 when you need 1200 × 630, and you have 144 pixels of horizontal slack and 138 of vertical to place the crop with, after the fact, where it actually works."}]},{"t":"h2","id":"the-type-zone-is-part-of-the-composition","runs":[{"text":"The type zone is part of the composition"}]},{"t":"p","runs":[{"text":"If a headline goes across the top third, that third has two requirements: it must be quiet, and it must have enough contrast against white or black type to be legible. A busy sky with bright clouds fails the second even though it is “empty”."}]},{"t":"p","runs":[{"text":"Generate for it. “Plain overcast sky occupying the upper third” gives you a place to set type. So does “dark studio background, subject lit from the side, lower two thirds”. You are not decorating; you are building a surface that type can sit on, and the finishing block returns to the legibility arithmetic."}]},{"t":"h2","id":"check-it-before-you-deliver","runs":[{"text":"Check it before you deliver"}]},{"t":"p","runs":[{"text":"Ninety seconds in any editor, including free ones:"}]},{"t":"ol","start":1,"items":[[{"text":"Open the image in GIMP, Krita or Photopea."}],[{"text":"Add guides at 1:1, 16:9 and 9:16 centred on the frame."}],[{"text":"Look at each. Is the subject intact? Is anything important cut? Does the 1:1 still read at 150 pixels?"}]]},{"t":"p","runs":[{"text":"If a crop fails, you have three options and they are not equal. Re-crop, which is free. Outpaint to add room, which takes a few minutes. Or regenerate, which costs the composition you had approved. Doing this check before the client sees anything means you only ever have the cheap version of the problem."}]},{"t":"h2","id":"the-limitation","runs":[{"text":"The limitation"}],"role":"limit"},{"t":"p","runs":[{"text":"Crop-safe composition costs you the strongest version of the picture. A frame designed to survive five crops is, almost by definition, more centred and more conservative than one composed for a single format. That is a real trade, not a free technique."}]},{"t":"p","runs":[{"text":"So make it deliberately. A hero image that only ever appears in one place should be composed for that place and nothing else. Crop safety is for images that will travel: social assets, stock, anything a client’s marketing team will re-use without asking. Applying it everywhere produces a portfolio of safe, centred, slightly dull pictures, and that is its own kind of failure."}]}],"check":{"question":[{"text":"A 4:5 social image keeps losing the subject’s face when platforms crop it to a square thumbnail. What is the underlying composition error?"}],"options":[{"text":[{"text":"The image was generated at a non-native aspect ratio, so the model placed the subject inconsistently within the frame."}],"correct":false,"tempting":[{"text":"Bucket mismatch produces duplicated or distorted anatomy rather than a subject positioned too near an edge; the framing here is a compositional choice, not an artefact."}]},{"text":[{"text":"Important content was placed in the band that only survives the original ratio, instead of the central core."}],"correct":true,"tempting":null},{"text":[{"text":"The resolution is too low, so the platform’s automatic crop cannot identify the subject and defaults to a centred rectangle."}],"correct":false,"tempting":[{"text":"Invents a mechanism: platform crops are usually fixed or face-detected, and neither behaves differently because an image has more pixels in it."}]},{"text":[{"text":"The delivered file should have been a square from the beginning, because any image reused across several placements ought to be generated in the most neutral format available."}],"correct":false,"tempting":[{"text":"Solves it by giving up the 4:5 frame entirely; the point of safe areas is to serve several crops from one strong composition rather than defaulting to the weakest one."}]}],"order":[3,0,2,1],"why":null},"prev":{"slug":"aspect-ratio-and-the-native-bucket","title":"The ratios the model was trained on, and what happens outside them","minutes":8},"next":{"slug":"the-reference-pack","title":"Six references, each with a job","minutes":7}}}