{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"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","position":24,"of":69,"minutes":8,"answer":"Light is the single highest-leverage slot in a prompt, and leaving it unspecified is the main reason generated images look generically generated.","keyIdea":"Light is the single highest-leverage slot in a prompt, and leaving it unspecified is the main reason generated images look generically generated."},"body":{"format":"markdown","text":"## The tell you can fix in one line\n\nPut two generated images side by side: one with no lighting described, one with. The first has soft, even, directionless illumination with weak shadows, faces lit from everywhere, and a slight glow. It is the average of millions of photographs, and averaging light is what produces that look.\n\nPhotographers and cinematographers spend most of their working lives on light because it carries almost everything: mood, depth, time, weather, whether a scene feels safe or not. The same is true here, and it costs one slot in your prompt.\n\n## Vocabulary that lands\n\n**Direction**, which is the most important and most neglected:\n\n- `side light`, `lit from the left`: texture, form, drama\n- `backlit`, `rim light`, `contre-jour`: outline, separation, haze\n- `top light`: harsh, institutional, unflattering\n- `underlit`: unsettling; almost never accidental\n- `front light`: flat, documentary, the passport-photo look\n\n**Quality**, meaning hard or soft:\n\n- `hard sunlight`, `direct sun`, `harsh shadows`: sharp-edged shadows\n- `soft light`, `overcast`, `diffused`, `large softbox`: gradual shadows\n- `dappled light through leaves`: a specific, very effective texture\n\n**Time and weather**, which bundle direction, quality and colour into one phrase:\n\n- `golden hour`, `blue hour`, `midday sun`, `overcast afternoon`, `pre-dawn`\n- `foggy`, `after rain`, `humid haze`\n\n**Source**, which sets colour and mood:\n\n- `window light`, `candlelight`, `fluorescent office lighting`, `sodium street lamp`, `screen glow`, `neon`, `single bare bulb`\n\n**Ratio**, borrowed from the studio:\n\n- `high contrast, deep shadows`, `low key`, `high key`, `chiaroscuro`\n\n## Three combinations worth memorising\n\nFor a portrait that does not look generated: `window light from the left, soft, deep shadows on the right side of the face, overcast afternoon`. Directional but soft, and the shadow side does the work.\n\nFor a product that looks like it was shot properly: `single large softbox from above and slightly behind, black card on the left for a dark edge, dark background`. That dark edge along one side is what separates a real product shot from a floating object.\n\nFor a documentary feel: `hard late-afternoon sun, long shadows, lens flare, shot into the light`. Imperfect on purpose.\n\n## Why this fixes more than lighting\n\nDescribing light also fixes composition and mood as a side effect, because light and composition are not separable. \"Backlit in a doorway\" decides where the subject is, what the background does, and how the eye moves through the frame. That is three slots' worth of control from one phrase.\n\nIt also solves the most common complaint about generated images: that they look plastic. Plastic is not a texture problem. It is a light problem: no dominant source, no real shadow, no falloff. Give the scene one strong source and a direction, and the plastic mostly goes.\n\n## The limitation to expect\n\nThe model does not compute light. There is no source, no ray, no surface. It reproduces the *appearance* of a lighting setup, which means the light is right where the training data was dense and wrong where it was not.\n\nConcretely: shadows that do not agree on a direction, a reflection that does not correspond to anything in the room, a mirror showing a scene that is not the scene, a light source visible in frame that casts nothing. These get worse with more objects and worse again with reflective or transparent materials: glass, water, chrome, polished floors.\n\nYou catch them by looking, deliberately, at a checklist: every shadow should agree on a direction; every reflective surface should reflect something plausible; anything glowing in frame should be illuminating what is next to it. Two of the three will usually be fine and one will not, and that one is an inpainting job rather than a prompting one. Nobody's eye catches these automatically at first. It is a habit you build in about a fortnight, and after that you cannot stop seeing them.\n\n## Learning light without a studio\n\nThe fastest way to build this vocabulary costs nothing and does not involve a generator at all.\n\nTake a single object (an onion, a mug, a shoe) and photograph it eight times in one afternoon with a phone. By a window at 10am. By the same window at 4pm. With a white sheet of paper bouncing light back on the shadow side. With a dark cloth there instead. Under the overhead room light. With a phone torch held to one side, then above, then below.\n\nEight photographs of the same object, and the difference between them is enormous. Look at where the shadow falls, how sharp its edge is, how much of the form you can read, and what each one makes the object feel like.\n\nThis is a real afternoon and it is worth more than any amount of reading, because it attaches the words to something you have seen happen. Afterwards, `hard light from the left` is not a phrase you copied from a prompt list; it is a thing you have made and can predict. That is the difference between describing light and choosing it."},"check":{"question":"Generated portraits look plastic despite a detailed subject description. What is the most likely cause?","options":[{"text":"The checkpoint is a community merge whose averaging has pushed its output toward an over-smoothed skin rendering that no prompt can fully counteract.","correct":false,"tempting":"Merges do drift toward a glossy house look and this is worth checking, but the same checkpoint produces convincing skin the moment a directional source is named."},{"text":"No light is specified, so the model produces its soft directionless average, which reads as plastic.","correct":true},{"text":"The prompt lacks skin-texture vocabulary such as pores and fine lines, which the model needs in order to render a surface as anything other than uniformly smooth.","correct":false,"tempting":"Texture words help a little at the pixel level, but a face with real pores under flat directionless light still reads as plastic, because the problem is form rather than surface."},{"text":"The guidance scale is set too high, which smooths fine detail as the sampler is pushed harder toward the prompt on every step.","correct":false,"tempting":"High guidance does flatten and over-contrast an image, but it produces a burnt look rather than the specific weightless quality being described."}]},"prev":{"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"},"next":{"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"},"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) 'Name the light or accept the model's average', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/light-is-the-whole-look (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":"Name the light or accept the model’s average","courseTitle":"Image Generation, in Practice","module":{"index":3,"slug":"first-frames","title":"Getting to a frame worth keeping"},"answer":null,"keyIdea":"Light is the single highest-leverage slot in a prompt, and leaving it unspecified is the main reason generated images look generically generated.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"the-tell-you-can-fix-in-one-line","runs":[{"text":"The tell you can fix in one line"}]},{"t":"p","runs":[{"text":"Put two generated images side by side: one with no lighting described, one with. The first has soft, even, directionless illumination with weak shadows, faces lit from everywhere, and a slight glow. It is the average of millions of photographs, and averaging light is what produces that look."}]},{"t":"p","runs":[{"text":"Photographers and cinematographers spend most of their working lives on light because it carries almost everything: mood, depth, time, weather, whether a scene feels safe or not. The same is true here, and it costs one slot in your prompt."}]},{"t":"h2","id":"vocabulary-that-lands","runs":[{"text":"Vocabulary that lands"}]},{"t":"p","runs":[{"text":"Direction","b":true},{"text":", which is the most important and most neglected:"}]},{"t":"ul","items":[[{"text":"side light","code":true},{"text":", "},{"text":"lit from the left","code":true},{"text":": texture, form, drama"}],[{"text":"backlit","code":true},{"text":", "},{"text":"rim light","code":true},{"text":", "},{"text":"contre-jour","code":true},{"text":": outline, separation, haze"}],[{"text":"top light","code":true},{"text":": harsh, institutional, unflattering"}],[{"text":"underlit","code":true},{"text":": unsettling; almost never accidental"}],[{"text":"front light","code":true},{"text":": flat, documentary, the passport-photo look"}]]},{"t":"p","runs":[{"text":"Quality","b":true},{"text":", meaning hard or soft:"}]},{"t":"ul","items":[[{"text":"hard sunlight","code":true},{"text":", "},{"text":"direct sun","code":true},{"text":", "},{"text":"harsh shadows","code":true},{"text":": sharp-edged shadows"}],[{"text":"soft light","code":true},{"text":", "},{"text":"overcast","code":true},{"text":", "},{"text":"diffused","code":true},{"text":", "},{"text":"large softbox","code":true},{"text":": gradual shadows"}],[{"text":"dappled light through leaves","code":true},{"text":": a specific, very effective texture"}]]},{"t":"p","runs":[{"text":"Time and weather","b":true},{"text":", which bundle direction, quality and colour into one phrase:"}]},{"t":"ul","items":[[{"text":"golden hour","code":true},{"text":", "},{"text":"blue hour","code":true},{"text":", "},{"text":"midday sun","code":true},{"text":", "},{"text":"overcast afternoon","code":true},{"text":", "},{"text":"pre-dawn","code":true}],[{"text":"foggy","code":true},{"text":", "},{"text":"after rain","code":true},{"text":", "},{"text":"humid haze","code":true}]]},{"t":"p","runs":[{"text":"Source","b":true},{"text":", which sets colour and mood:"}]},{"t":"ul","items":[[{"text":"window light","code":true},{"text":", "},{"text":"candlelight","code":true},{"text":", "},{"text":"fluorescent office lighting","code":true},{"text":", "},{"text":"sodium street lamp","code":true},{"text":", "},{"text":"screen glow","code":true},{"text":", "},{"text":"neon","code":true},{"text":", "},{"text":"single bare bulb","code":true}]]},{"t":"p","runs":[{"text":"Ratio","b":true},{"text":", borrowed from the studio:"}]},{"t":"ul","items":[[{"text":"high contrast, deep shadows","code":true},{"text":", "},{"text":"low key","code":true},{"text":", "},{"text":"high key","code":true},{"text":", "},{"text":"chiaroscuro","code":true}]]},{"t":"h2","id":"three-combinations-worth-memorising","runs":[{"text":"Three combinations worth memorising"}]},{"t":"p","runs":[{"text":"For a portrait that does not look generated: "},{"text":"window light from the left, soft, deep shadows on the right side of the face, overcast afternoon","code":true},{"text":". Directional but soft, and the shadow side does the work."}]},{"t":"p","runs":[{"text":"For a product that looks like it was shot properly: "},{"text":"single large softbox from above and slightly behind, black card on the left for a dark edge, dark background","code":true},{"text":". That dark edge along one side is what separates a real product shot from a floating object."}]},{"t":"p","runs":[{"text":"For a documentary feel: "},{"text":"hard late-afternoon sun, long shadows, lens flare, shot into the light","code":true},{"text":". Imperfect on purpose."}]},{"t":"h2","id":"why-this-fixes-more-than-lighting","runs":[{"text":"Why this fixes more than lighting"}]},{"t":"p","runs":[{"text":"Describing light also fixes composition and mood as a side effect, because light and composition are not separable. “Backlit in a doorway” decides where the subject is, what the background does, and how the eye moves through the frame. That is three slots’ worth of control from one phrase."}]},{"t":"p","runs":[{"text":"It also solves the most common complaint about generated images: that they look plastic. Plastic is not a texture problem. It is a light problem: no dominant source, no real shadow, no falloff. Give the scene one strong source and a direction, and the plastic mostly goes."}]},{"t":"h2","id":"the-limitation-to-expect","runs":[{"text":"The limitation to expect"}],"role":"limit"},{"t":"p","runs":[{"text":"The model does not compute light. There is no source, no ray, no surface. It reproduces the "},{"text":"appearance","i":true},{"text":" of a lighting setup, which means the light is right where the training data was dense and wrong where it was not."}]},{"t":"p","runs":[{"text":"Concretely: shadows that do not agree on a direction, a reflection that does not correspond to anything in the room, a mirror showing a scene that is not the scene, a light source visible in frame that casts nothing. These get worse with more objects and worse again with reflective or transparent materials: glass, water, chrome, polished floors."}]},{"t":"p","runs":[{"text":"You catch them by looking, deliberately, at a checklist: every shadow should agree on a direction; every reflective surface should reflect something plausible; anything glowing in frame should be illuminating what is next to it. Two of the three will usually be fine and one will not, and that one is an inpainting job rather than a prompting one. Nobody’s eye catches these automatically at first. It is a habit you build in about a fortnight, and after that you cannot stop seeing them."}]},{"t":"h2","id":"learning-light-without-a-studio","runs":[{"text":"Learning light without a studio"}]},{"t":"p","runs":[{"text":"The fastest way to build this vocabulary costs nothing and does not involve a generator at all."}]},{"t":"p","runs":[{"text":"Take a single object (an onion, a mug, a shoe) and photograph it eight times in one afternoon with a phone. By a window at 10am. By the same window at 4pm. With a white sheet of paper bouncing light back on the shadow side. With a dark cloth there instead. Under the overhead room light. With a phone torch held to one side, then above, then below."}]},{"t":"p","runs":[{"text":"Eight photographs of the same object, and the difference between them is enormous. Look at where the shadow falls, how sharp its edge is, how much of the form you can read, and what each one makes the object feel like."}]},{"t":"p","runs":[{"text":"This is a real afternoon and it is worth more than any amount of reading, because it attaches the words to something you have seen happen. Afterwards, "},{"text":"hard light from the left","code":true},{"text":" is not a phrase you copied from a prompt list; it is a thing you have made and can predict. That is the difference between describing light and choosing it."}]}],"check":{"question":[{"text":"Generated portraits look plastic despite a detailed subject description. What is the most likely cause?"}],"options":[{"text":[{"text":"The checkpoint is a community merge whose averaging has pushed its output toward an over-smoothed skin rendering that no prompt can fully counteract."}],"correct":false,"tempting":[{"text":"Merges do drift toward a glossy house look and this is worth checking, but the same checkpoint produces convincing skin the moment a directional source is named."}]},{"text":[{"text":"No light is specified, so the model produces its soft directionless average, which reads as plastic."}],"correct":true,"tempting":null},{"text":[{"text":"The prompt lacks skin-texture vocabulary such as pores and fine lines, which the model needs in order to render a surface as anything other than uniformly smooth."}],"correct":false,"tempting":[{"text":"Texture words help a little at the pixel level, but a face with real pores under flat directionless light still reads as plastic, because the problem is form rather than surface."}]},{"text":[{"text":"The guidance scale is set too high, which smooths fine detail as the sampler is pushed harder toward the prompt on every step."}],"correct":false,"tempting":[{"text":"High guidance does flatten and over-contrast an image, but it produces a burnt look rather than the specific weightless quality being described."}]}],"order":[0,3,1,2],"why":null},"prev":{"slug":"lens-and-camera-words","title":"The camera vocabulary that works, and the words that do nothing","minutes":8},"next":{"slug":"style-without-a-living-artist","title":"Describing a look without naming a living artist","minutes":8}}}