{"course":{"slug":"ai-video","title":"Video with AI","url":"https://addaly.com/learn/ai-video","level":"intermediate","track":"create"},"lesson":{"slug":"steps-guidance-and-samplers","title":"The four numbers, and which of them you should touch","url":"https://addaly.com/learn/ai-video/steps-guidance-and-samplers","position":7,"of":66,"minutes":9,"answer":"Guidance scale controls how hard the model is pushed toward your prompt and away from its unconditional prediction, so raising it past a narrow band burns contrast and freezes motion rather than improving adherence.","keyIdea":"Guidance scale controls how hard the model is pushed toward your prompt and away from its unconditional prediction, so raising it past a narrow band burns contrast and freezes motion rather than improving adherence."},"body":{"format":"markdown","text":"## The controls, in order of how often they should move\n\nMost interfaces expose four things: steps, guidance, sampler, and shift. In practice you should touch them in almost the reverse of the order people usually do.\n\n- **Steps.** Set once per model from your own step-count test. Then leave it.\n- **Guidance.** The one worth understanding and occasionally adjusting.\n- **Sampler.** Pick the documented default. Differences are small on video.\n- **Shift.** Leave alone unless the model's own guide tells you otherwise.\n\nEverything else (prompt, still, motion strength, duration) matters more than all four combined.\n\n## What guidance actually does\n\nAt each step the model makes two predictions: one conditioned on your prompt, one unconditioned. Classifier-free guidance takes the difference and extrapolates along it:\n\n```\nprediction = uncond + scale * (cond - uncond)\n```\n\nAt scale 1 you get the conditioned prediction unchanged. Above 1 you are pushing *past* it: deliberately overshooting in the direction the prompt pulls.\n\nThat overshoot is why guidance works and why it breaks. Push harder and the prompt is obeyed more literally. Push too hard and you leave the region where the model's predictions are valid. The symptoms are specific and recognisable:\n\n> Figure: Raising guidance does not buy adherence, it buys stillness. Bars (movement in the clip, per cent of the amount at guidance 3): Guidance 1 78; Guidance 3 100; Guidance 5 94; Guidance 7 71; Guidance 9 38. High guidance drives every frame toward the strongest single interpretation of the prompt, and the strongest interpretation does not change over time, so the clip stiffens toward a photograph while the colours cook. A full model usually wants 3 to 7. A distilled one wants about 1, and feeding it 7 produces exactly the burnt, frozen output on the right.\n\n- Contrast and saturation climb until highlights clip to white.\n- Edges get a hard, posterised outline.\n- **Motion freezes.** This one is particular to video and catches people out. High guidance drives every frame toward the strongest interpretation of the prompt, and the strongest interpretation does not change over time, so the clip stiffens into a near-still.\n\nIf your motion looks locked and your colours look cooked, lower the guidance before you touch anything else.\n\n## The usable ranges\n\nVideo models generally want lower guidance than image models. Typical bands, which you should verify on your own model rather than trust here:\n\n- Full (non-distilled) video models: roughly 3 to 7, with 5 a common default.\n- Distilled or turbo variants: near 1. These were trained to need no guidance, and feeding them 7 produces exactly the burnt, frozen output described above.\n- Some models expose separate guidance for the image condition and the text condition. Raising image guidance tightens adherence to your first frame at the cost of motion; lowering it lets the model drift from your still.\n\nThe diagnostic: generate the same shot at guidance 2, 5 and 9 with a fixed seed. You will see the motion die and the colour cook, and you will know your model's ceiling for the rest of your career with it.\n\n## Samplers, briefly and honestly\n\nOn image models sampler choice was a genuine subject. On flow-matching video models it is mostly not. The differences between Euler, DPM variants and the rest are small compared with the differences between prompts, stills and guidance values.\n\nUse whatever the model card recommends. If you want to test it, test it once, with a fixed seed, and then stop thinking about it. Time spent on samplers is time not spent on the shot list, and the shot list is where the quality is.\n\n## Shift, and why it exists for video\n\nFlow-matching models spend their steps along a path from noise to data. **Shift** biases where along that path the effort goes. Higher shift concentrates steps early, where large-scale structure and motion are decided; lower shift concentrates them late, where fine detail resolves.\n\nVideo wants more early effort than images do, because getting the motion trajectory right matters more than getting a texture crisp. That is why video models ship with a higher default shift, and why raising resolution sometimes calls for raising shift too: a bigger latent has more structure to settle.\n\nIt is a real control and a rarely useful one. If motion is consistently mushy at good guidance and adequate steps, try raising it. Otherwise leave it.\n\n## What people reach for instead, and should not\n\nThree habits that waste money:\n\n1. **Raising steps to fix a bad shot.** Steps resolve detail. They do not change what the model decided to depict. A shot that is wrong at 20 steps is the same shot, sharper, at 50.\n2. **Raising guidance to force an ignored instruction.** If the model ignored \"she turns left\", guidance 9 gives you a stiffer clip in which she still does not turn left. The fix is the still or the phrasing.\n3. **Changing four things at once and keeping the result.** You now have a setting you cannot explain and cannot repeat.\n\n## The free path\n\nAll of this is testable for nothing on a Hugging Face Space or a Colab notebook, because none of it needs scale: it needs one still, one seed, and patience with a queue. Build a small grid: three guidance values across two step counts, six generations, one page of notes. That page is worth more than any settings guide, including this one, because it is about your model.\n\n---\n\nToday: generate one shot at guidance 2, 5 and 9 with the seed fixed, and watch the motion disappear as the colour cooks."},"check":{"question":"A shot generated at guidance 9 comes back with clipped highlights and almost no movement, although the prompt asked for a walking figure. What is happening mechanically?","options":[{"text":"Guidance extrapolates past the conditioned prediction, and at a high scale every frame is pushed toward the same fixed interpretation.","correct":true},{"text":"The prompt is being followed so precisely that the model has no freedom left to invent motion between the frames.","correct":false,"tempting":"Frames the failure as adherence succeeding; the figure is not walking, so adherence is worse, not better."},{"text":"High guidance requires proportionally more steps, and at the default step count the sampler has not converged yet.","correct":false,"tempting":"Confuses an out-of-distribution extrapolation with an under-converged one; more steps sharpen the same frozen clip."},{"text":"The unconditional branch is skipped above a threshold scale, so the model loses the motion prior that this branch normally carries into each step.","correct":false,"tempting":"Invents a threshold behaviour; the unconditional prediction is used at every scale, and it is the extrapolation away from it that causes the damage."}]},"prev":{"slug":"seeds-and-determinism","title":"Same settings, different clip, and why that is not a bug","url":"https://addaly.com/learn/ai-video/seeds-and-determinism"},"next":{"slug":"frame-rate-and-interpolation","title":"Generated at sixteen, delivered at twenty-four","url":"https://addaly.com/learn/ai-video/frame-rate-and-interpolation"},"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 four numbers, and which of them you should touch', in Video with AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-video/steps-guidance-and-samplers (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 four numbers, and which of them you should touch","courseTitle":"Video with AI","module":{"index":1,"slug":"how-video-models-work","title":"What the machine is actually doing"},"answer":null,"keyIdea":"Guidance scale controls how hard the model is pushed toward your prompt and away from its unconditional prediction, so raising it past a narrow band burns contrast and freezes motion rather than improving adherence.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"bars","unit":"movement in the clip, per cent of the amount at guidance 3","title":"Raising guidance does not buy adherence, it buys stillness","series":[{"label":"Guidance 1","value":78},{"label":"Guidance 3","value":100},{"label":"Guidance 5","value":94},{"label":"Guidance 7","value":71},{"label":"Guidance 9","value":38,"highlight":true}],"caption":"High guidance drives every frame toward the strongest single interpretation of the prompt, and the strongest interpretation does not change over time, so the clip stiffens toward a photograph while the colours cook. A full model usually wants 3 to 7. A distilled one wants about 1, and feeding it 7 produces exactly the burnt, frozen output on the right."},"alt":"Figure: Raising guidance does not buy adherence, it buys stillness. Bars (movement in the clip, per cent of the amount at guidance 3): Guidance 1 78; Guidance 3 100; Guidance 5 94; Guidance 7 71; Guidance 9 38. High guidance drives every frame toward the strongest single interpretation of the prompt, and the strongest interpretation does not change over time, so the clip stiffens toward a photograph while the colours cook. A full model usually wants 3 to 7. A distilled one wants about 1, and feeding it 7 produces exactly the burnt, frozen output on the right."}},"blocks":[{"t":"h2","id":"the-controls-in-order-of-how-often-they-should-move","runs":[{"text":"The controls, in order of how often they should move"}]},{"t":"p","runs":[{"text":"Most interfaces expose four things: steps, guidance, sampler, and shift. In practice you should touch them in almost the reverse of the order people usually do."}]},{"t":"ul","items":[[{"text":"Steps.","b":true},{"text":" Set once per model from your own step-count test. Then leave it."}],[{"text":"Guidance.","b":true},{"text":" The one worth understanding and occasionally adjusting."}],[{"text":"Sampler.","b":true},{"text":" Pick the documented default. Differences are small on video."}],[{"text":"Shift.","b":true},{"text":" Leave alone unless the model’s own guide tells you otherwise."}]]},{"t":"p","runs":[{"text":"Everything else (prompt, still, motion strength, duration) matters more than all four combined."}]},{"t":"h2","id":"what-guidance-actually-does","runs":[{"text":"What guidance actually does"}]},{"t":"p","runs":[{"text":"At each step the model makes two predictions: one conditioned on your prompt, one unconditioned. Classifier-free guidance takes the difference and extrapolates along it:"}]},{"t":"code","lang":"","label":"Text","text":"prediction = uncond + scale * (cond - uncond)"},{"t":"p","runs":[{"text":"At scale 1 you get the conditioned prediction unchanged. Above 1 you are pushing "},{"text":"past","i":true},{"text":" it: deliberately overshooting in the direction the prompt pulls."}]},{"t":"p","runs":[{"text":"That overshoot is why guidance works and why it breaks. Push harder and the prompt is obeyed more literally. Push too hard and you leave the region where the model’s predictions are valid. The symptoms are specific and recognisable:"}]},{"t":"ul","items":[[{"text":"Contrast and saturation climb until highlights clip to white."}],[{"text":"Edges get a hard, posterised outline."}],[{"text":"Motion freezes.","b":true},{"text":" This one is particular to video and catches people out. High guidance drives every frame toward the strongest interpretation of the prompt, and the strongest interpretation does not change over time, so the clip stiffens into a near-still."}]]},{"t":"p","runs":[{"text":"If your motion looks locked and your colours look cooked, lower the guidance before you touch anything else."}]},{"t":"h2","id":"the-usable-ranges","runs":[{"text":"The usable ranges"}]},{"t":"p","runs":[{"text":"Video models generally want lower guidance than image models. Typical bands, which you should verify on your own model rather than trust here:"}]},{"t":"ul","items":[[{"text":"Full (non-distilled) video models: roughly 3 to 7, with 5 a common default."}],[{"text":"Distilled or turbo variants: near 1. These were trained to need no guidance, and feeding them 7 produces exactly the burnt, frozen output described above."}],[{"text":"Some models expose separate guidance for the image condition and the text condition. Raising image guidance tightens adherence to your first frame at the cost of motion; lowering it lets the model drift from your still."}]]},{"t":"p","runs":[{"text":"The diagnostic: generate the same shot at guidance 2, 5 and 9 with a fixed seed. You will see the motion die and the colour cook, and you will know your model’s ceiling for the rest of your career with it."}]},{"t":"h2","id":"samplers-briefly-and-honestly","runs":[{"text":"Samplers, briefly and honestly"}]},{"t":"p","runs":[{"text":"On image models sampler choice was a genuine subject. On flow-matching video models it is mostly not. The differences between Euler, DPM variants and the rest are small compared with the differences between prompts, stills and guidance values."}]},{"t":"p","runs":[{"text":"Use whatever the model card recommends. If you want to test it, test it once, with a fixed seed, and then stop thinking about it. Time spent on samplers is time not spent on the shot list, and the shot list is where the quality is."}]},{"t":"h2","id":"shift-and-why-it-exists-for-video","runs":[{"text":"Shift, and why it exists for video"}]},{"t":"p","runs":[{"text":"Flow-matching models spend their steps along a path from noise to data. "},{"text":"Shift","b":true},{"text":" biases where along that path the effort goes. Higher shift concentrates steps early, where large-scale structure and motion are decided; lower shift concentrates them late, where fine detail resolves."}]},{"t":"p","runs":[{"text":"Video wants more early effort than images do, because getting the motion trajectory right matters more than getting a texture crisp. That is why video models ship with a higher default shift, and why raising resolution sometimes calls for raising shift too: a bigger latent has more structure to settle."}]},{"t":"p","runs":[{"text":"It is a real control and a rarely useful one. If motion is consistently mushy at good guidance and adequate steps, try raising it. Otherwise leave it."}]},{"t":"h2","id":"what-people-reach-for-instead-and-should-not","runs":[{"text":"What people reach for instead, and should not"}]},{"t":"p","runs":[{"text":"Three habits that waste money:"}]},{"t":"ol","start":1,"items":[[{"text":"Raising steps to fix a bad shot.","b":true},{"text":" Steps resolve detail. They do not change what the model decided to depict. A shot that is wrong at 20 steps is the same shot, sharper, at 50."}],[{"text":"Raising guidance to force an ignored instruction.","b":true},{"text":" If the model ignored “she turns left”, guidance 9 gives you a stiffer clip in which she still does not turn left. The fix is the still or the phrasing."}],[{"text":"Changing four things at once and keeping the result.","b":true},{"text":" You now have a setting you cannot explain and cannot repeat."}]]},{"t":"h2","id":"the-free-path","runs":[{"text":"The free path"}],"role":"free"},{"t":"p","runs":[{"text":"All of this is testable for nothing on a Hugging Face Space or a Colab notebook, because none of it needs scale: it needs one still, one seed, and patience with a queue. Build a small grid: three guidance values across two step counts, six generations, one page of notes. That page is worth more than any settings guide, including this one, because it is about your model."}]},{"t":"rule"},{"t":"p","runs":[{"text":"Today: generate one shot at guidance 2, 5 and 9 with the seed fixed, and watch the motion disappear as the colour cooks."}]}],"check":{"question":[{"text":"A shot generated at guidance 9 comes back with clipped highlights and almost no movement, although the prompt asked for a walking figure. What is happening mechanically?"}],"options":[{"text":[{"text":"Guidance extrapolates past the conditioned prediction, and at a high scale every frame is pushed toward the same fixed interpretation."}],"correct":true,"tempting":null},{"text":[{"text":"The prompt is being followed so precisely that the model has no freedom left to invent motion between the frames."}],"correct":false,"tempting":[{"text":"Frames the failure as adherence succeeding; the figure is not walking, so adherence is worse, not better."}]},{"text":[{"text":"High guidance requires proportionally more steps, and at the default step count the sampler has not converged yet."}],"correct":false,"tempting":[{"text":"Confuses an out-of-distribution extrapolation with an under-converged one; more steps sharpen the same frozen clip."}]},{"text":[{"text":"The unconditional branch is skipped above a threshold scale, so the model loses the motion prior that this branch normally carries into each step."}],"correct":false,"tempting":[{"text":"Invents a threshold behaviour; the unconditional prediction is used at every scale, and it is the extrapolation away from it that causes the damage."}]}],"order":[2,1,0,3],"why":null},"prev":{"slug":"seeds-and-determinism","title":"Same settings, different clip, and why that is not a bug","minutes":8},"next":{"slug":"frame-rate-and-interpolation","title":"Generated at sixteen, delivered at twenty-four","minutes":9}}}