{"course":{"slug":"ai-video","title":"Video with AI","url":"https://addaly.com/learn/ai-video","level":"intermediate","track":"create"},"lesson":{"slug":"negative-prompts-and-their-limits","title":"What a negative prompt can and cannot remove","url":"https://addaly.com/learn/ai-video/negative-prompts-and-their-limits","position":27,"of":66,"minutes":8,"answer":"A negative prompt is a second conditioning the model is pushed away from, which works well for qualities of rendering and badly for objects.","keyIdea":"A negative prompt is a second conditioning the model is pushed away from, which works well for qualities of rendering and badly for objects, because naming an object pulls it into the guided direction before the negation subtracts it."},"body":{"format":"markdown","text":"## The mechanism, in one paragraph\n\nClassifier-free guidance compares a conditioned prediction against an unconditional one. A negative prompt replaces \"unconditional\" with \"conditioned on this other text\", and the model is pushed away from it:\n\n```\nprediction = neg + scale * (pos - neg)\n```\n\nSo a negative prompt does not delete anything. It moves the target away from a region. Whether that helps depends entirely on whether the thing you named is a direction in that space or an object in the picture.\n\n## What works: qualities of rendering\n\nTerms describing *how* the image looks rather than *what is in it*:\n\n- `blurry, low quality, compression artefacts, jpeg artifacts`\n- `warping, morphing, distorted, deformed`\n- `extra limbs, extra fingers, fused fingers`\n- `watermark, text overlay, logo, subtitles, timestamp`\n- `oversaturated, washed out, flat lighting`\n- `flickering, strobing, stuttering`\n\nThese work because they name a broad region of bad output that the model has seen a great deal of. Pushing away from that region improves things across the frame.\n\nThe video-specific ones (warping, morphing, flickering) are worth having in a default negative, because they describe exactly the failures video models produce.\n\n## What fails: objects\n\nPut `no dog` or `dog` in a negative prompt and you will often get a dog.\n\nThe reason is worth understanding. Text conditioning does not carry negation well; the token for the concept activates the concept. In the positive prompt, saying \"a street with no dogs\" reliably produces dogs, for the same reason telling somebody not to think of an elephant does not work. In the negative prompt the subtraction is real but weak, because one noun against a whole conditioned prediction is a small vector.\n\nWhat happens in practice is that the concept gets activated and only partly subtracted, so you get a dog-shaped smear, or a dog in the background, or a dog-adjacent animal.\n\n**The reliable way to remove an object from a frame is to remove it from the frame.** Repair the still. Clone it out. This is the point of the repair lesson and it applies with full force here.\n\n## Negatives can suppress motion you want\n\nA trap specific to video. A negative list containing `motion blur` and `movement` and `camera shake` will produce a stiffer clip than you intended, because you have pushed away from the region of the space that contains movement generally.\n\nIf your clips come back oddly frozen, read your negative prompt before you touch anything else. This is a common and quietly expensive mistake.\n\nKeep the negative list short and about rendering failures. Four to eight terms. Long copied-and-pasted negative lists from image-model forums are a liability here.\n\n## Distilled models often ignore it entirely\n\nA model trained to work at guidance 1 is not doing the extrapolation the formula above describes, so the negative branch may have little or no effect. Some interfaces still show the box.\n\nTest it once: generate with an absurd negative prompt (`blue, sky, person`) and see whether anything changes. If nothing does, stop typing in that box and save yourself the time.\n\n## Prompt weighting, briefly\n\nSome tools accept emphasis syntax such as `(wet pavement:1.3)` to push a term harder. Where it exists it is worth knowing, and it obeys the same rule as everything else: it steers emphasis, it does not supply information the model lacks. Weighting `(logo:1.5)` does not make the model know your logo.\n\nOveruse produces the same burnt, frozen output as excessive guidance, because it is the same operation applied to one token.\n\n## The summary\n\nA negative prompt is a small, free improvement for rendering quality and a poor tool for content control. The hierarchy of what actually controls a generated clip, strongest first:\n\n> Figure: What actually controls a generated clip. Layers, top to bottom: The first frame (A dense, complete specification of the picture. When the image and the text disagree, the image wins, every time.); A conditioning signal (Depth, pose or optical flow from a reference clip: a per-frame structural target, which a sentence cannot be.); Camera and motion controls (A strength slider, a painted region, a drawn path. A number is unambiguous and a sentence is not.); The positive prompt (A few hundred text tokens steering a quarter of a million video tokens. Real, and light.); The negative prompt (A small free improvement to rendering quality, and a poor tool for deciding what is in the frame.). Most people spend their attention in the reverse order, rewriting a prompt to fix something the first frame decided. Spending it in this order is most of what separates a production workflow from an afternoon of tinkering.\n\n1. The first frame.\n2. The conditioning signal, if you supplied one.\n3. Camera and motion controls.\n4. The positive prompt.\n5. The negative prompt.\n\nPeople spend their time in the reverse order. Spending it in this order is most of what separates a production workflow from an afternoon of prompt tinkering.\n\n---\n\nToday: shorten your negative prompt to six terms about rendering quality, remove anything that names an object or a kind of movement, and see whether your clips get looser and better."},"check":{"question":"An operator adds \"dog\" to a negative prompt to keep a dog out of a street scene, and dogs keep appearing in the background. What is the reliable fix?","options":[{"text":"Increase the guidance scale so the subtraction of the negative branch is applied more strongly to the conditioned prediction.","correct":false,"tempting":"Amplifies the whole extrapolation, which burns contrast and freezes motion long before it removes the animal."},{"text":"Add more related terms (dog, puppy, canine, animal) so the negative vector covers the concept more completely.","correct":false,"tempting":"Broadens a weak subtraction and risks removing wanted animals and fur textures without reliably removing the dog."},{"text":"Remove the dog from the first frame instead, because the still governs content far more strongly than any text branch.","correct":true},{"text":"Move the term into the positive prompt as \"a street with no dogs\", since positive conditioning has a stronger effect on the output than negative conditioning does.","correct":false,"tempting":"Correct that the positive branch is stronger, and wrong about negation: naming the concept positively activates it more reliably still."}]},"prev":{"slug":"conditioned-motion-controls","title":"Depth, pose and flow as instructions","url":"https://addaly.com/learn/ai-video/conditioned-motion-controls"},"next":{"slug":"one-variable-at-a-time","title":"An experiment protocol that fits on a card","url":"https://addaly.com/learn/ai-video/one-variable-at-a-time"},"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) 'What a negative prompt can and cannot remove', in Video with AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-video/negative-prompts-and-their-limits (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":"What a negative prompt can and cannot remove","courseTitle":"Video with AI","module":{"index":3,"slug":"directing-the-shot","title":"Directing, not describing"},"answer":null,"keyIdea":"A negative prompt is a second conditioning the model is pushed away from, which works well for qualities of rendering and badly for objects, because naming an object pulls it into the guided direction before the negation subtracts it.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"stack","title":"What actually controls a generated clip","layers":[{"label":"The first frame","detail":"A dense, complete specification of the picture. When the image and the text disagree, the image wins, every time."},{"label":"A conditioning signal","detail":"Depth, pose or optical flow from a reference clip: a per-frame structural target, which a sentence cannot be."},{"label":"Camera and motion controls","detail":"A strength slider, a painted region, a drawn path. A number is unambiguous and a sentence is not."},{"label":"The positive prompt","detail":"A few hundred text tokens steering a quarter of a million video tokens. Real, and light."},{"label":"The negative prompt","detail":"A small free improvement to rendering quality, and a poor tool for deciding what is in the frame."}],"caption":"Most people spend their attention in the reverse order, rewriting a prompt to fix something the first frame decided. Spending it in this order is most of what separates a production workflow from an afternoon of tinkering."},"alt":"Figure: What actually controls a generated clip. Layers, top to bottom: The first frame (A dense, complete specification of the picture. When the image and the text disagree, the image wins, every time.); A conditioning signal (Depth, pose or optical flow from a reference clip: a per-frame structural target, which a sentence cannot be.); Camera and motion controls (A strength slider, a painted region, a drawn path. A number is unambiguous and a sentence is not.); The positive prompt (A few hundred text tokens steering a quarter of a million video tokens. Real, and light.); The negative prompt (A small free improvement to rendering quality, and a poor tool for deciding what is in the frame.). Most people spend their attention in the reverse order, rewriting a prompt to fix something the first frame decided. Spending it in this order is most of what separates a production workflow from an afternoon of tinkering."}},"blocks":[{"t":"h2","id":"the-mechanism-in-one-paragraph","runs":[{"text":"The mechanism, in one paragraph"}]},{"t":"p","runs":[{"text":"Classifier-free guidance compares a conditioned prediction against an unconditional one. A negative prompt replaces “unconditional” with “conditioned on this other text”, and the model is pushed away from it:"}]},{"t":"code","lang":"","label":"Text","text":"prediction = neg + scale * (pos - neg)"},{"t":"p","runs":[{"text":"So a negative prompt does not delete anything. It moves the target away from a region. Whether that helps depends entirely on whether the thing you named is a direction in that space or an object in the picture."}]},{"t":"h2","id":"what-works-qualities-of-rendering","runs":[{"text":"What works: qualities of rendering"}]},{"t":"p","runs":[{"text":"Terms describing "},{"text":"how","i":true},{"text":" the image looks rather than "},{"text":"what is in it","i":true},{"text":":"}]},{"t":"ul","items":[[{"text":"blurry, low quality, compression artefacts, jpeg artifacts","code":true}],[{"text":"warping, morphing, distorted, deformed","code":true}],[{"text":"extra limbs, extra fingers, fused fingers","code":true}],[{"text":"watermark, text overlay, logo, subtitles, timestamp","code":true}],[{"text":"oversaturated, washed out, flat lighting","code":true}],[{"text":"flickering, strobing, stuttering","code":true}]]},{"t":"p","runs":[{"text":"These work because they name a broad region of bad output that the model has seen a great deal of. Pushing away from that region improves things across the frame."}]},{"t":"p","runs":[{"text":"The video-specific ones (warping, morphing, flickering) are worth having in a default negative, because they describe exactly the failures video models produce."}]},{"t":"h2","id":"what-fails-objects","runs":[{"text":"What fails: objects"}]},{"t":"p","runs":[{"text":"Put "},{"text":"no dog","code":true},{"text":" or "},{"text":"dog","code":true},{"text":" in a negative prompt and you will often get a dog."}]},{"t":"p","runs":[{"text":"The reason is worth understanding. Text conditioning does not carry negation well; the token for the concept activates the concept. In the positive prompt, saying “a street with no dogs” reliably produces dogs, for the same reason telling somebody not to think of an elephant does not work. In the negative prompt the subtraction is real but weak, because one noun against a whole conditioned prediction is a small vector."}]},{"t":"p","runs":[{"text":"What happens in practice is that the concept gets activated and only partly subtracted, so you get a dog-shaped smear, or a dog in the background, or a dog-adjacent animal."}]},{"t":"p","runs":[{"text":"The reliable way to remove an object from a frame is to remove it from the frame.","b":true},{"text":" Repair the still. Clone it out. This is the point of the repair lesson and it applies with full force here."}]},{"t":"h2","id":"negatives-can-suppress-motion-you-want","runs":[{"text":"Negatives can suppress motion you want"}]},{"t":"p","runs":[{"text":"A trap specific to video. A negative list containing "},{"text":"motion blur","code":true},{"text":" and "},{"text":"movement","code":true},{"text":" and "},{"text":"camera shake","code":true},{"text":" will produce a stiffer clip than you intended, because you have pushed away from the region of the space that contains movement generally."}]},{"t":"p","runs":[{"text":"If your clips come back oddly frozen, read your negative prompt before you touch anything else. This is a common and quietly expensive mistake."}]},{"t":"p","runs":[{"text":"Keep the negative list short and about rendering failures. Four to eight terms. Long copied-and-pasted negative lists from image-model forums are a liability here."}]},{"t":"h2","id":"distilled-models-often-ignore-it-entirely","runs":[{"text":"Distilled models often ignore it entirely"}]},{"t":"p","runs":[{"text":"A model trained to work at guidance 1 is not doing the extrapolation the formula above describes, so the negative branch may have little or no effect. Some interfaces still show the box."}]},{"t":"p","runs":[{"text":"Test it once: generate with an absurd negative prompt ("},{"text":"blue, sky, person","code":true},{"text":") and see whether anything changes. If nothing does, stop typing in that box and save yourself the time."}]},{"t":"h2","id":"prompt-weighting-briefly","runs":[{"text":"Prompt weighting, briefly"}]},{"t":"p","runs":[{"text":"Some tools accept emphasis syntax such as "},{"text":"(wet pavement:1.3)","code":true},{"text":" to push a term harder. Where it exists it is worth knowing, and it obeys the same rule as everything else: it steers emphasis, it does not supply information the model lacks. Weighting "},{"text":"(logo:1.5)","code":true},{"text":" does not make the model know your logo."}]},{"t":"p","runs":[{"text":"Overuse produces the same burnt, frozen output as excessive guidance, because it is the same operation applied to one token."}]},{"t":"h2","id":"the-summary","runs":[{"text":"The summary"}]},{"t":"p","runs":[{"text":"A negative prompt is a small, free improvement for rendering quality and a poor tool for content control. The hierarchy of what actually controls a generated clip, strongest first:"}]},{"t":"ol","start":1,"items":[[{"text":"The first frame."}],[{"text":"The conditioning signal, if you supplied one."}],[{"text":"Camera and motion controls."}],[{"text":"The positive prompt."}],[{"text":"The negative prompt."}]]},{"t":"p","runs":[{"text":"People spend their time in the reverse order. Spending it in this order is most of what separates a production workflow from an afternoon of prompt tinkering."}]},{"t":"rule"},{"t":"p","runs":[{"text":"Today: shorten your negative prompt to six terms about rendering quality, remove anything that names an object or a kind of movement, and see whether your clips get looser and better."}]}],"check":{"question":[{"text":"An operator adds “dog” to a negative prompt to keep a dog out of a street scene, and dogs keep appearing in the background. What is the reliable fix?"}],"options":[{"text":[{"text":"Increase the guidance scale so the subtraction of the negative branch is applied more strongly to the conditioned prediction."}],"correct":false,"tempting":[{"text":"Amplifies the whole extrapolation, which burns contrast and freezes motion long before it removes the animal."}]},{"text":[{"text":"Add more related terms (dog, puppy, canine, animal) so the negative vector covers the concept more completely."}],"correct":false,"tempting":[{"text":"Broadens a weak subtraction and risks removing wanted animals and fur textures without reliably removing the dog."}]},{"text":[{"text":"Remove the dog from the first frame instead, because the still governs content far more strongly than any text branch."}],"correct":true,"tempting":null},{"text":[{"text":"Move the term into the positive prompt as “a street with no dogs”, since positive conditioning has a stronger effect on the output than negative conditioning does."}],"correct":false,"tempting":[{"text":"Correct that the positive branch is stronger, and wrong about negation: naming the concept positively activates it more reliably still."}]}],"order":[3,2,0,1],"why":null},"prev":{"slug":"conditioned-motion-controls","title":"Depth, pose and flow as instructions","minutes":10},"next":{"slug":"one-variable-at-a-time","title":"An experiment protocol that fits on a card","minutes":8}}}