{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"slug":"the-contact-sheet","title":"Sixteen cheap frames beat four expensive ones","url":"https://addaly.com/learn/ai-images/the-contact-sheet","position":26,"of":69,"minutes":7,"answer":"Judging a grid of many low-cost renders at once is faster and more accurate than judging full-quality images one at a time, because composition is visible at thumbnail size and commitment bias is not.","keyIdea":"Judging a grid of many low-cost renders at once is faster and more accurate than judging full-quality images one at a time, because composition is visible at thumbnail size and commitment bias is not."},"body":{"format":"markdown","text":"## Borrowed from film, and still correct\n\nA photographer shooting film did not examine each frame. They printed a contact sheet (every frame from the roll on one sheet, thumbnail-sized) and marked the ones worth enlarging. The reason it worked is that **composition is legible at thumbnail size and detail is not**, so looking at small images forces you to judge the thing that actually decides whether a frame is any good.\n\nThe same method, applied here, is the biggest single improvement most people can make to how they work.\n\n## The method\n\n1. Pick a fast, cheap configuration: a distilled model at 6 steps, or your normal model at 12 to 15 steps. Quality does not matter. Composition does.\n2. Generate 16 images at the native ratio with random seeds, batched.\n3. Assemble them as a 4 × 4 grid. Most interfaces have a grid output option; otherwise it is one line of ImageMagick, which is free.\n4. Look at the grid **at thumbnail size**, not full screen.\n5. Mark two or three. Discard the rest without sentiment.\n6. Take the marked seeds forward at full quality.\n\n```bash\nmontage *.png -tile 4x4 -geometry 256x256+4+4 sheet.jpg\n```\n\nOn a mid-range card with a distilled model, sixteen frames is under a minute. On a hosted service at $0.01 a small image, it is sixteen cents. Either way it is cheap enough to do four times with different prompts, which is the point.\n\n> Figure: What one exploration frame costs, by how you render it. Bars (seconds per frame on an RTX 3060): Distilled model, 6 steps 3; Your own checkpoint, 12 steps 6; Full quality, 30 steps 13. Sixteen frames on the top row is under a minute; sixteen on the bottom is three and a half. Composition is legible at 250 pixels, so the cheap sheet answers the only question you are asking at this stage.\n\n## The three questions\n\nLooking at the grid, ask only these:\n\n1. **Does the eye land in the right place?** If you cannot tell what the picture is about at 250 pixels, no amount of finishing will fix it.\n2. **Is the light doing something?** Flat frames stay flat.\n3. **Is the frame structurally sound**: subject not cut awkwardly, not duplicated, room where the type goes?\n\nNotice what is not on the list: hands, faces, small text, fine detail. All of those are repairable and none should influence the choice. The most expensive mistake in this work is choosing a frame with beautiful hands and mediocre composition, because you can fix the hands.\n\n## Why it beats iterating one at a time\n\nThree reasons, and the third is the one people underrate.\n\n**Cost per decision.** Sixteen cheap renders cost what two full-quality ones do, and give you eight times the information about composition.\n\n**Comparison rather than evaluation.** Asked \"is this good?\", people say yes. Asked \"which of these sixteen is best?\", people answer accurately. Judgement is far more reliable when it is relative.\n\n**No commitment bias.** After waiting ninety seconds for a full-quality render, you want it to be good, and you will start negotiating with yourself about its composition. A thumbnail you waited three seconds for gets thrown away honestly.\n\n## The cost of the cheap render\n\nBeing straight about the trade: a distilled model at 6 steps does not compose identically to the full model at 30. It has less variety, and it pulls toward the most probable interpretation of the prompt, so your sheet under-represents the unusual frames the full model would have produced.\n\nTwo things follow. First, if the sheet is dull, suspect the model before the prompt: run four at full quality to check. Second, use the *same* model at low steps when the difference matters; 12 steps on your real checkpoint is still four times cheaper than 30 and composes the way the final render will.\n\nThe more subtle failure is that a grid rewards frames that read instantly. That is exactly right for a thumbnail, a feed or a poster seen in passing, and it quietly penalises the quieter image that rewards a long look. If the deliverable is something people will sit with, mark one frame from the sheet that you suspect is better than it looks small, and render it properly before deciding.\n\n## Sheets of variations, not just of seeds\n\nOnce the habit is established, the sheet becomes the tool for every question, not only for seed selection.\n\n- **Four prompts × four seeds** on one sheet tells you which phrasing is actually stronger, rather than which seed was lucky.\n- **One prompt × four checkpoints × four seeds** tells you which model suits the brief, which is a question people otherwise answer from forum reputation.\n- **One seed × a sweep of CFG values** shows you the burn point in a single glance.\n\nThe rule that makes all of these readable is to vary one thing along each axis and label it. A sheet where four things changed at once is a pretty picture and tells you nothing, which is the subject of a later lesson in this block.\n\n## Show the sheet to the client\n\nAn unexpected use, and a good one. Handing over a contact sheet with three frames circled turns an approval into a choice, and choices are easy for people who find open-ended feedback hard.\n\nTwo cautions. Only include frames you would be content to deliver, because somebody will pick the one you were least keen on. And say plainly that these are rough exploration renders, or you will get notes about hands."},"check":{"question":"Why judge exploration frames as thumbnails on a grid rather than at full size one at a time?","options":[{"text":"Because thumbnails load faster, which allows many more candidates to be reviewed within the same working session.","correct":false,"tempting":"Review speed is a minor benefit; the reason the method works is what small size makes invisible, not how quickly the files open."},{"text":"Because small size hides repairable detail and forces judgement onto composition, and relative comparison is more reliable than absolute.","correct":true},{"text":"Because low-step renders are too unfinished to be judged at full size, so viewing them small avoids drawing conclusions from artefacts that the final render will not contain.","correct":false,"tempting":"Sensible-sounding but backwards: the artefacts are ignored because they are irrelevant to the decision, not because the images are unfinished."},{"text":"Because a grid layout makes it easier to spot duplicated subjects and other structural failures that appear across several frames at once.","correct":false,"tempting":"Structural faults are one of the three questions, not the purpose; a single full-size frame shows a duplicated torso perfectly well."}]},"prev":{"slug":"style-without-a-living-artist","title":"Describing a look without naming a living artist","url":"https://addaly.com/learn/ai-images/style-without-a-living-artist"},"next":{"slug":"seed-discipline","title":"Fix the seed or learn nothing","url":"https://addaly.com/learn/ai-images/seed-discipline"},"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) 'Sixteen cheap frames beat four expensive ones', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/the-contact-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":"Sixteen cheap frames beat four expensive ones","courseTitle":"Image Generation, in Practice","module":{"index":3,"slug":"first-frames","title":"Getting to a frame worth keeping"},"answer":null,"keyIdea":"Judging a grid of many low-cost renders at once is faster and more accurate than judging full-quality images one at a time, because composition is visible at thumbnail size and commitment bias is not.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"bars","unit":"seconds per frame on an RTX 3060","title":"What one exploration frame costs, by how you render it","series":[{"label":"Distilled model, 6 steps","value":3,"highlight":true},{"label":"Your own checkpoint, 12 steps","value":6,"highlight":true},{"label":"Full quality, 30 steps","value":13}],"caption":"Sixteen frames on the top row is under a minute; sixteen on the bottom is three and a half. Composition is legible at 250 pixels, so the cheap sheet answers the only question you are asking at this stage."},"alt":"Figure: What one exploration frame costs, by how you render it. Bars (seconds per frame on an RTX 3060): Distilled model, 6 steps 3; Your own checkpoint, 12 steps 6; Full quality, 30 steps 13. Sixteen frames on the top row is under a minute; sixteen on the bottom is three and a half. Composition is legible at 250 pixels, so the cheap sheet answers the only question you are asking at this stage."}},"blocks":[{"t":"h2","id":"borrowed-from-film-and-still-correct","runs":[{"text":"Borrowed from film, and still correct"}]},{"t":"p","runs":[{"text":"A photographer shooting film did not examine each frame. They printed a contact sheet (every frame from the roll on one sheet, thumbnail-sized) and marked the ones worth enlarging. The reason it worked is that "},{"text":"composition is legible at thumbnail size and detail is not","b":true},{"text":", so looking at small images forces you to judge the thing that actually decides whether a frame is any good."}]},{"t":"p","runs":[{"text":"The same method, applied here, is the biggest single improvement most people can make to how they work."}]},{"t":"h2","id":"the-method","runs":[{"text":"The method"}]},{"t":"ol","start":1,"items":[[{"text":"Pick a fast, cheap configuration: a distilled model at 6 steps, or your normal model at 12 to 15 steps. Quality does not matter. Composition does."}],[{"text":"Generate 16 images at the native ratio with random seeds, batched."}],[{"text":"Assemble them as a 4 × 4 grid. Most interfaces have a grid output option; otherwise it is one line of ImageMagick, which is free."}],[{"text":"Look at the grid "},{"text":"at thumbnail size","b":true},{"text":", not full screen."}],[{"text":"Mark two or three. Discard the rest without sentiment."}],[{"text":"Take the marked seeds forward at full quality."}]]},{"t":"code","lang":"bash","label":"Bash","text":"montage *.png -tile 4x4 -geometry 256x256+4+4 sheet.jpg"},{"t":"p","runs":[{"text":"On a mid-range card with a distilled model, sixteen frames is under a minute. On a hosted service at $0.01 a small image, it is sixteen cents. Either way it is cheap enough to do four times with different prompts, which is the point."}]},{"t":"h2","id":"the-three-questions","runs":[{"text":"The three questions"}]},{"t":"p","runs":[{"text":"Looking at the grid, ask only these:"}]},{"t":"ol","start":1,"items":[[{"text":"Does the eye land in the right place?","b":true},{"text":" If you cannot tell what the picture is about at 250 pixels, no amount of finishing will fix it."}],[{"text":"Is the light doing something?","b":true},{"text":" Flat frames stay flat."}],[{"text":"Is the frame structurally sound","b":true},{"text":": subject not cut awkwardly, not duplicated, room where the type goes?"}]]},{"t":"p","runs":[{"text":"Notice what is not on the list: hands, faces, small text, fine detail. All of those are repairable and none should influence the choice. The most expensive mistake in this work is choosing a frame with beautiful hands and mediocre composition, because you can fix the hands."}]},{"t":"h2","id":"why-it-beats-iterating-one-at-a-time","runs":[{"text":"Why it beats iterating one at a time"}]},{"t":"p","runs":[{"text":"Three reasons, and the third is the one people underrate."}]},{"t":"p","runs":[{"text":"Cost per decision.","b":true},{"text":" Sixteen cheap renders cost what two full-quality ones do, and give you eight times the information about composition."}]},{"t":"p","runs":[{"text":"Comparison rather than evaluation.","b":true},{"text":" Asked “is this good?”, people say yes. Asked “which of these sixteen is best?”, people answer accurately. Judgement is far more reliable when it is relative."}]},{"t":"p","runs":[{"text":"No commitment bias.","b":true},{"text":" After waiting ninety seconds for a full-quality render, you want it to be good, and you will start negotiating with yourself about its composition. A thumbnail you waited three seconds for gets thrown away honestly."}]},{"t":"h2","id":"the-cost-of-the-cheap-render","runs":[{"text":"The cost of the cheap render"}]},{"t":"p","runs":[{"text":"Being straight about the trade: a distilled model at 6 steps does not compose identically to the full model at 30. It has less variety, and it pulls toward the most probable interpretation of the prompt, so your sheet under-represents the unusual frames the full model would have produced."}]},{"t":"p","runs":[{"text":"Two things follow. First, if the sheet is dull, suspect the model before the prompt: run four at full quality to check. Second, use the "},{"text":"same","i":true},{"text":" model at low steps when the difference matters; 12 steps on your real checkpoint is still four times cheaper than 30 and composes the way the final render will."}]},{"t":"p","runs":[{"text":"The more subtle failure is that a grid rewards frames that read instantly. That is exactly right for a thumbnail, a feed or a poster seen in passing, and it quietly penalises the quieter image that rewards a long look. If the deliverable is something people will sit with, mark one frame from the sheet that you suspect is better than it looks small, and render it properly before deciding."}]},{"t":"h2","id":"sheets-of-variations-not-just-of-seeds","runs":[{"text":"Sheets of variations, not just of seeds"}]},{"t":"p","runs":[{"text":"Once the habit is established, the sheet becomes the tool for every question, not only for seed selection."}]},{"t":"ul","items":[[{"text":"Four prompts × four seeds","b":true},{"text":" on one sheet tells you which phrasing is actually stronger, rather than which seed was lucky."}],[{"text":"One prompt × four checkpoints × four seeds","b":true},{"text":" tells you which model suits the brief, which is a question people otherwise answer from forum reputation."}],[{"text":"One seed × a sweep of CFG values","b":true},{"text":" shows you the burn point in a single glance."}]]},{"t":"p","runs":[{"text":"The rule that makes all of these readable is to vary one thing along each axis and label it. A sheet where four things changed at once is a pretty picture and tells you nothing, which is the subject of a later lesson in this block."}]},{"t":"h2","id":"show-the-sheet-to-the-client","runs":[{"text":"Show the sheet to the client"}]},{"t":"p","runs":[{"text":"An unexpected use, and a good one. Handing over a contact sheet with three frames circled turns an approval into a choice, and choices are easy for people who find open-ended feedback hard."}]},{"t":"p","runs":[{"text":"Two cautions. Only include frames you would be content to deliver, because somebody will pick the one you were least keen on. And say plainly that these are rough exploration renders, or you will get notes about hands."}]}],"check":{"question":[{"text":"Why judge exploration frames as thumbnails on a grid rather than at full size one at a time?"}],"options":[{"text":[{"text":"Because thumbnails load faster, which allows many more candidates to be reviewed within the same working session."}],"correct":false,"tempting":[{"text":"Review speed is a minor benefit; the reason the method works is what small size makes invisible, not how quickly the files open."}]},{"text":[{"text":"Because small size hides repairable detail and forces judgement onto composition, and relative comparison is more reliable than absolute."}],"correct":true,"tempting":null},{"text":[{"text":"Because low-step renders are too unfinished to be judged at full size, so viewing them small avoids drawing conclusions from artefacts that the final render will not contain."}],"correct":false,"tempting":[{"text":"Sensible-sounding but backwards: the artefacts are ignored because they are irrelevant to the decision, not because the images are unfinished."}]},{"text":[{"text":"Because a grid layout makes it easier to spot duplicated subjects and other structural failures that appear across several frames at once."}],"correct":false,"tempting":[{"text":"Structural faults are one of the three questions, not the purpose; a single full-size frame shows a duplicated torso perfectly well."}]}],"order":[3,0,2,1],"why":null},"prev":{"slug":"style-without-a-living-artist","title":"Describing a look without naming a living artist","minutes":8},"next":{"slug":"seed-discipline","title":"Fix the seed or learn nothing","minutes":7}}}