{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"slug":"lens-and-camera-words","title":"The camera vocabulary that works, and the words that do nothing","url":"https://addaly.com/learn/ai-images/lens-and-camera-words","position":23,"of":69,"minutes":8,"answer":"Focal length, aperture, camera height and shot size measurably change composition because they appeared in the captions of millions of training photographs, while quality words like \"8k\" and \"masterpiece\" mostly do nothing on modern models.","keyIdea":"Focal length, aperture, camera height and shot size measurably change composition because they appeared in the captions of millions of training photographs, while quality words like \"8k\" and \"masterpiece\" mostly do nothing on modern models."},"body":{"format":"markdown","text":"## Why photographic words work\n\nA model learned from images with captions, and a very large number of those captions came from photography sites where the metadata was attached: focal length, aperture, shot type. So the model has seen tens of thousands of examples of what \"85mm f/1.8\" looks like (compressed perspective, shallow depth of field, a particular subject distance) and the phrase reliably produces that.\n\nThis makes photographic vocabulary the most efficient lever available. One term replaces a paragraph of description.\n\n## The terms that reliably change the image\n\n**Focal length** changes perspective, not just zoom:\n\n- `24mm` or `wide angle`: broad view, exaggerated depth, distortion near edges. Good for environments and for leaving headline room.\n- `35mm`: the documentary standard. Natural, slightly wide, subject in context.\n- `50mm`: close to human vision. Neutral.\n- `85mm`: portrait compression. Flattering, background falls away, tighter frame.\n- `135mm` or `200mm`: heavy compression, background collapses toward the subject.\n- `macro`: very close, extremely shallow focus.\n\n**Aperture** controls the background:\n\n- `f/1.4`, `f/1.8`: background dissolved, one plane sharp.\n- `f/5.6`: subject and near context sharp.\n- `f/16`: everything sharp, deep landscape look.\n\n**Camera height and angle**, which is the most under-used control of all:\n\n- `low angle`, `from below`: the subject gains stature.\n- `eye level`: neutral, the default.\n- `high angle`, `from above`: the subject diminishes.\n- `overhead`, `top-down`: flat-lay, for objects and food.\n- `waist-level`: the documentary look, and the one that stops portraits feeling posed.\n\n**Shot size**, borrowed from film and instantly understood:\n\n- `extreme wide shot`, `wide shot`, `medium shot`, `close-up`, `extreme close-up`.\n\n## The words that do nothing\n\nOn SD 1.5, quality tags had measurable effect because community fine-tunes had been trained with them in the captions. On modern models they are mostly inert or actively harmful:\n\n- `8k`, `4k`, `ultra HD`: resolution is set by the image size, not by words. On some checkpoints they nudge toward an over-sharpened stock-photo look.\n- `masterpiece`, `best quality`, `award winning`: largely inert on newer models; on some anime-lineage checkpoints they still do something, which is why the habit persists.\n- `highly detailed`, `intricate`: tends to fill the frame with ornament, which is usually not what you meant.\n- `trending on ArtStation`: a phrase with a specific historical effect on an old model that people still copy into models where it means nothing.\n\nTest them rather than trusting either me or the forum: fix the seed, generate with and without, and look. On most current checkpoints you will see almost no difference, and the sixty characters are better spent on light.\n\n## The failure this vocabulary causes\n\nThere is a real cost to leaning on it, and it is worth naming because it is the most common way competent prompts produce boring pictures.\n\nPhotographic terms pull toward the *centre* of what that term looked like in the training data. Ask for `85mm portrait, f/1.8, soft light` and you will get an extremely competent, extremely average portrait: the statistical middle of a million flattering headshots. The vocabulary is precise, which means it is precisely conventional.\n\nThe way out is to use the technical slot for control and put the interest somewhere else: an unusual subject, an odd moment, a specific place, an awkward gesture. Or to deliberately combine terms that rarely co-occur (a macro lens on a landscape subject, a low angle on a domestic scene) which pushes the model off the well-worn path.\n\nOne more practical note: these terms describe an appearance, not an optical simulation. There is no camera. `f/1.4` produces the *look* of shallow depth of field, applied by a model that has no depth information, which is why the blur sometimes falls in the wrong place: a hand at the same distance as the face going soft, or a background object mysteriously sharp. When that happens, the fix is a real depth-of-field pass in an editor, not a different aperture number.\n\n## Film and video terms, which also work\n\nThe same logic extends beyond stills, because film stills and production photographs were captioned too.\n\n- `anamorphic`: wide frame, horizontal flares, oval highlights in the background blur.\n- `handheld`: slight imperfection in framing, a documentary immediacy.\n- `dutch angle`: tilted horizon. Use sparingly; it is a strong signal.\n- `shot on 16mm`: grainier, contrastier, a particular period feel.\n- `establishing shot`: pulls the camera back and includes the environment.\n\nAnd two terms from the studio that do a great deal of work for product and still life: `seamless backdrop` gives you the infinite white or grey sweep behind a product, and `tabletop, hero angle` gives the slightly-above three-quarter view that almost every product photograph in the world uses.\n\n## A caution about stacking\n\nEvery one of these terms narrows the distribution, and four or five of them together narrow it hard. `85mm, f/1.4, golden hour, shallow depth of field, bokeh` is five constraints that all point at the same well-trodden place, and the result will be the most predictable image the model can make.\n\nUse two or three. If you find yourself adding a fifth camera term, the problem is almost certainly in a different slot."},"check":{"question":"Why do focal-length terms like \"35mm\" and \"85mm\" change an image more reliably than \"8k, masterpiece\"?","options":[{"text":"Because numeric terms are tokenised more precisely than adjectives, which gives the text encoder a sharper signal to condition the denoising process on.","correct":false,"tempting":"Tokenisation is not the mechanism; \"35mm\" works because of what it was attached to in captions, not because digits encode more crisply than letters."},{"text":"Because focal length physically determines perspective, and the model has learned an internal optical simulation of how lenses project a scene.","correct":false,"tempting":"The model has no geometry of lenses; it reproduces the appearance that accompanied those captions, which is why the resulting blur sometimes lands at the wrong depth."},{"text":"Because focal lengths appeared in enormous numbers of training captions attached to a consistent look, whereas quality tags did not.","correct":true},{"text":"Because quality tags are usually placed at the end of a prompt, where the encoder weights them less heavily than terms appearing near the beginning.","correct":false,"tempting":"Position has some effect on CLIP-based models, but moving \"masterpiece\" to the front does not make it work, which is the test that rules this out."}]},"prev":{"slug":"the-prompt-in-slots","title":"Build the prompt in slots, not as a sentence you keep rewriting","url":"https://addaly.com/learn/ai-images/the-prompt-in-slots"},"next":{"slug":"light-is-the-whole-look","title":"Name the light or accept the model's average","url":"https://addaly.com/learn/ai-images/light-is-the-whole-look"},"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 camera vocabulary that works, and the words that do nothing', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/lens-and-camera-words (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 camera vocabulary that works, and the words that do nothing","courseTitle":"Image Generation, in Practice","module":{"index":3,"slug":"first-frames","title":"Getting to a frame worth keeping"},"answer":null,"keyIdea":"Focal length, aperture, camera height and shot size measurably change composition because they appeared in the captions of millions of training photographs, while quality words like “8k” and “masterpiece” mostly do nothing on modern models.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"why-photographic-words-work","runs":[{"text":"Why photographic words work"}]},{"t":"p","runs":[{"text":"A model learned from images with captions, and a very large number of those captions came from photography sites where the metadata was attached: focal length, aperture, shot type. So the model has seen tens of thousands of examples of what “85mm f/1.8” looks like (compressed perspective, shallow depth of field, a particular subject distance) and the phrase reliably produces that."}]},{"t":"p","runs":[{"text":"This makes photographic vocabulary the most efficient lever available. One term replaces a paragraph of description."}]},{"t":"h2","id":"the-terms-that-reliably-change-the-image","runs":[{"text":"The terms that reliably change the image"}]},{"t":"p","runs":[{"text":"Focal length","b":true},{"text":" changes perspective, not just zoom:"}]},{"t":"ul","items":[[{"text":"24mm","code":true},{"text":" or "},{"text":"wide angle","code":true},{"text":": broad view, exaggerated depth, distortion near edges. Good for environments and for leaving headline room."}],[{"text":"35mm","code":true},{"text":": the documentary standard. Natural, slightly wide, subject in context."}],[{"text":"50mm","code":true},{"text":": close to human vision. Neutral."}],[{"text":"85mm","code":true},{"text":": portrait compression. Flattering, background falls away, tighter frame."}],[{"text":"135mm","code":true},{"text":" or "},{"text":"200mm","code":true},{"text":": heavy compression, background collapses toward the subject."}],[{"text":"macro","code":true},{"text":": very close, extremely shallow focus."}]]},{"t":"p","runs":[{"text":"Aperture","b":true},{"text":" controls the background:"}]},{"t":"ul","items":[[{"text":"f/1.4","code":true},{"text":", "},{"text":"f/1.8","code":true},{"text":": background dissolved, one plane sharp."}],[{"text":"f/5.6","code":true},{"text":": subject and near context sharp."}],[{"text":"f/16","code":true},{"text":": everything sharp, deep landscape look."}]]},{"t":"p","runs":[{"text":"Camera height and angle","b":true},{"text":", which is the most under-used control of all:"}]},{"t":"ul","items":[[{"text":"low angle","code":true},{"text":", "},{"text":"from below","code":true},{"text":": the subject gains stature."}],[{"text":"eye level","code":true},{"text":": neutral, the default."}],[{"text":"high angle","code":true},{"text":", "},{"text":"from above","code":true},{"text":": the subject diminishes."}],[{"text":"overhead","code":true},{"text":", "},{"text":"top-down","code":true},{"text":": flat-lay, for objects and food."}],[{"text":"waist-level","code":true},{"text":": the documentary look, and the one that stops portraits feeling posed."}]]},{"t":"p","runs":[{"text":"Shot size","b":true},{"text":", borrowed from film and instantly understood:"}]},{"t":"ul","items":[[{"text":"extreme wide shot","code":true},{"text":", "},{"text":"wide shot","code":true},{"text":", "},{"text":"medium shot","code":true},{"text":", "},{"text":"close-up","code":true},{"text":", "},{"text":"extreme close-up","code":true},{"text":"."}]]},{"t":"h2","id":"the-words-that-do-nothing","runs":[{"text":"The words that do nothing"}]},{"t":"p","runs":[{"text":"On SD 1.5, quality tags had measurable effect because community fine-tunes had been trained with them in the captions. On modern models they are mostly inert or actively harmful:"}]},{"t":"ul","items":[[{"text":"8k","code":true},{"text":", "},{"text":"4k","code":true},{"text":", "},{"text":"ultra HD","code":true},{"text":": resolution is set by the image size, not by words. On some checkpoints they nudge toward an over-sharpened stock-photo look."}],[{"text":"masterpiece","code":true},{"text":", "},{"text":"best quality","code":true},{"text":", "},{"text":"award winning","code":true},{"text":": largely inert on newer models; on some anime-lineage checkpoints they still do something, which is why the habit persists."}],[{"text":"highly detailed","code":true},{"text":", "},{"text":"intricate","code":true},{"text":": tends to fill the frame with ornament, which is usually not what you meant."}],[{"text":"trending on ArtStation","code":true},{"text":": a phrase with a specific historical effect on an old model that people still copy into models where it means nothing."}]]},{"t":"p","runs":[{"text":"Test them rather than trusting either me or the forum: fix the seed, generate with and without, and look. On most current checkpoints you will see almost no difference, and the sixty characters are better spent on light."}]},{"t":"h2","id":"the-failure-this-vocabulary-causes","runs":[{"text":"The failure this vocabulary causes"}]},{"t":"p","runs":[{"text":"There is a real cost to leaning on it, and it is worth naming because it is the most common way competent prompts produce boring pictures."}]},{"t":"p","runs":[{"text":"Photographic terms pull toward the "},{"text":"centre","i":true},{"text":" of what that term looked like in the training data. Ask for "},{"text":"85mm portrait, f/1.8, soft light","code":true},{"text":" and you will get an extremely competent, extremely average portrait: the statistical middle of a million flattering headshots. The vocabulary is precise, which means it is precisely conventional."}]},{"t":"p","runs":[{"text":"The way out is to use the technical slot for control and put the interest somewhere else: an unusual subject, an odd moment, a specific place, an awkward gesture. Or to deliberately combine terms that rarely co-occur (a macro lens on a landscape subject, a low angle on a domestic scene) which pushes the model off the well-worn path."}]},{"t":"p","runs":[{"text":"One more practical note: these terms describe an appearance, not an optical simulation. There is no camera. "},{"text":"f/1.4","code":true},{"text":" produces the "},{"text":"look","i":true},{"text":" of shallow depth of field, applied by a model that has no depth information, which is why the blur sometimes falls in the wrong place: a hand at the same distance as the face going soft, or a background object mysteriously sharp. When that happens, the fix is a real depth-of-field pass in an editor, not a different aperture number."}]},{"t":"h2","id":"film-and-video-terms-which-also-work","runs":[{"text":"Film and video terms, which also work"}]},{"t":"p","runs":[{"text":"The same logic extends beyond stills, because film stills and production photographs were captioned too."}]},{"t":"ul","items":[[{"text":"anamorphic","code":true},{"text":": wide frame, horizontal flares, oval highlights in the background blur."}],[{"text":"handheld","code":true},{"text":": slight imperfection in framing, a documentary immediacy."}],[{"text":"dutch angle","code":true},{"text":": tilted horizon. Use sparingly; it is a strong signal."}],[{"text":"shot on 16mm","code":true},{"text":": grainier, contrastier, a particular period feel."}],[{"text":"establishing shot","code":true},{"text":": pulls the camera back and includes the environment."}]]},{"t":"p","runs":[{"text":"And two terms from the studio that do a great deal of work for product and still life: "},{"text":"seamless backdrop","code":true},{"text":" gives you the infinite white or grey sweep behind a product, and "},{"text":"tabletop, hero angle","code":true},{"text":" gives the slightly-above three-quarter view that almost every product photograph in the world uses."}]},{"t":"h2","id":"a-caution-about-stacking","runs":[{"text":"A caution about stacking"}]},{"t":"p","runs":[{"text":"Every one of these terms narrows the distribution, and four or five of them together narrow it hard. "},{"text":"85mm, f/1.4, golden hour, shallow depth of field, bokeh","code":true},{"text":" is five constraints that all point at the same well-trodden place, and the result will be the most predictable image the model can make."}]},{"t":"p","runs":[{"text":"Use two or three. If you find yourself adding a fifth camera term, the problem is almost certainly in a different slot."}]}],"check":{"question":[{"text":"Why do focal-length terms like “35mm” and “85mm” change an image more reliably than “8k, masterpiece”?"}],"options":[{"text":[{"text":"Because numeric terms are tokenised more precisely than adjectives, which gives the text encoder a sharper signal to condition the denoising process on."}],"correct":false,"tempting":[{"text":"Tokenisation is not the mechanism; “35mm” works because of what it was attached to in captions, not because digits encode more crisply than letters."}]},{"text":[{"text":"Because focal length physically determines perspective, and the model has learned an internal optical simulation of how lenses project a scene."}],"correct":false,"tempting":[{"text":"The model has no geometry of lenses; it reproduces the appearance that accompanied those captions, which is why the resulting blur sometimes lands at the wrong depth."}]},{"text":[{"text":"Because focal lengths appeared in enormous numbers of training captions attached to a consistent look, whereas quality tags did not."}],"correct":true,"tempting":null},{"text":[{"text":"Because quality tags are usually placed at the end of a prompt, where the encoder weights them less heavily than terms appearing near the beginning."}],"correct":false,"tempting":[{"text":"Position has some effect on CLIP-based models, but moving “masterpiece” to the front does not make it work, which is the test that rules this out."}]}],"order":[1,0,3,2],"why":null},"prev":{"slug":"the-prompt-in-slots","title":"Build the prompt in slots, not as a sentence you keep rewriting","minutes":8},"next":{"slug":"light-is-the-whole-look","title":"Name the light or accept the model’s average","minutes":8}}}