{"course":{"slug":"ai-video","title":"Video with AI","url":"https://addaly.com/learn/ai-video","level":"intermediate","track":"create"},"lesson":{"slug":"denoising-a-block-of-time","title":"The whole clip is made at once, not frame by frame","url":"https://addaly.com/learn/ai-video/denoising-a-block-of-time","position":3,"of":66,"minutes":10,"answer":"A video model denoises every frame of the clip simultaneously across a fixed number of steps, which is why nothing can be streamed out early, why step count hits diminishing returns fast.","keyIdea":"A video model denoises every frame of the clip simultaneously across a fixed number of steps, which is why nothing can be streamed out early, why step count hits diminishing returns fast, and why distilled turbo models trade motion range for speed."},"body":{"format":"markdown","text":"## Starting from noise, all at once\n\nGeneration begins with the entire latent block filled with random noise: all 30 latent frames of it. The model then makes a prediction about that noise, subtracts part of it, and repeats. After a fixed number of steps the block is clean and gets decoded.\n\nThe important word is *simultaneously*. At step 7, frame 1 and frame 120 are both 7 steps along. They are being refined together, seeing each other, agreeing with each other about where the wall is.\n\nThis is the opposite of how most people imagine it, and it explains several things that otherwise look arbitrary:\n\n- **You cannot watch the first second while the rest renders.** There is no first second yet. Progress bars that show a preview are decoding a partially denoised block, which is why early previews look like coloured fog.\n- **You cannot generate a bit more onto the end.** The block size was fixed before step 1. Extending means a new run conditioned on a decoded frame, with all the photocopy damage that implies.\n- **Cost is decided at the start.** Duration, resolution and step count are all fixed before any picture exists. There is no way to spend more only on the interesting part.\n\n## What a step actually buys\n\nEach step is one full forward pass of the transformer over the whole latent. Twenty steps means twenty passes; fifty means fifty. Cost is close to linear in steps.\n\nQuality is not. The curve is steep at the start and flat after a point that is lower than most people expect. On a typical modern video model:\n\n> Figure: Where the step count stops buying anything. Line chart of Quality reached, per cent of this model's best against Denoising steps: Full model (4, 52) (8, 72) (15, 88) (20, 93) (25, 96) (30, 97) (40, 99) (50, 100); Distilled turbo (4, 84) (8, 90) (15, 91) (20, 91) (25, 91) (30, 91) (40, 91) (50, 91). Cost is close to linear in steps and quality is not. The shape is what transfers, not the numbers: run the same shot at four step counts with the seed fixed and find your own model's elbow. Note also where the turbo curve stops: it arrives fast and it stops lower, which is why a draft answers a different question from a final.\n\n- 4 to 8 steps: shapes and colour are right, motion is mushy, fine structure is unresolved.\n- 15 to 25 steps: most of the quality. This is where the majority of production work sits.\n- 30 to 50 steps: small improvements in fine detail and edge stability.\n- Beyond 50: you are usually paying for a different random walk, not a better one.\n\nRun this yourself with a fixed seed. Generate the same shot at 10, 20, 30 and 50 steps and put the four side by side. Most people are surprised at how early the returns stop, and it is worth knowing your own model's number rather than a general one.\n\n## Flow matching, and why older advice does not transfer\n\nEarly diffusion models used a noise schedule with hundreds of possible steps and samplers with names like DDIM and Euler-a. Most current video models are trained with **flow matching** instead, which learns a straighter path from noise to data and converges in far fewer steps.\n\nThis matters practically because a lot of advice written for 2022-era image models is wrong here. Step counts of 100, high guidance values, and sampler names that no longer exist in the interface all come from that lineage. If a tutorial tells you to use 80 steps on a flow-matching video model, it was written for different machinery.\n\nFlow-matching models also expose a **shift** or **sigma shift** parameter, which redistributes where the steps are spent along the path. Higher shift spends more effort in the early, structural part of the denoise. Video models generally want a higher shift than image models, and the tool usually sets a sensible default. Change it last, if at all.\n\n## Distilled and turbo models\n\nA distilled model is trained to imitate the output of a many-step model in very few steps: often 4 to 8. They are useful and they are how draft passes become affordable.\n\nWhat they cost you is real and worth naming:\n\n- **Reduced motion range.** Distillation tends to pull toward the average trajectory. Turbo outputs are noticeably calmer, which sometimes flatters a shot and sometimes kills it.\n- **Weaker prompt adherence on unusual requests.** Common instructions survive distillation. Odd ones get smoothed toward the common case.\n- **Guidance behaves differently.** Most distilled models want a guidance scale near 1. Feeding them the value you use on the full model produces burnt contrast and frozen motion.\n\nThe workflow that follows from this: draft on the turbo variant to test whether a motion idea works at all, then do the final on the full model with the same seed and still. The draft is not a preview of the final: it is a cheap answer to a different question.\n\n## Why the same settings give different pictures\n\nTwo runs with everything identical except the seed produce different clips because they start from different noise. That is expected. But two runs that you believe are identical can also differ, because the effective inputs include model version, precision, batch size and the specific GPU kernels used. The next lesson deals with that properly.\n\n## The free path\n\nEvery point above can be tested on a free Hugging Face Space or a Colab session, because they are all about settings rather than about scale. Fix the still and the seed, move one number, compare. An afternoon of that teaches you more about your chosen model than any comparison video, and it costs a queue rather than a balance.\n\n---\n\nToday: fix a seed and a still, generate at four step counts, and find the point on your own model where the improvement stops being visible. That number is now your production default."},"check":{"question":"A studio wants to show a client the first two seconds of a ten-second generation while the rest finishes rendering, to save waiting time. Why is this not possible?","options":[{"text":"The API does not expose partial results, though the model itself produces frames sequentially and a self-hosted setup could stream them out as they finish.","correct":false,"tempting":"Blames the interface for an architectural property; a local run has the same block and the same problem."},{"text":"Early frames are rendered at lower quality first and only refined at the end, so a partial result would misrepresent the shot.","correct":false,"tempting":"Invents a progressive per-frame refinement order that flow-matching video diffusion does not use."},{"text":"The decoder can only run on the complete latent block, although the latent frames themselves finish at different times.","correct":false,"tempting":"Half right about the decoder but wrong about the latent: no group of latent frames finishes ahead of the others."},{"text":"Every frame is at the same stage of denoising throughout, so before the last step the first two seconds are as unfinished as the rest.","correct":true}]},"prev":{"slug":"latent-space-for-video","title":"Your clip is compressed fifty times before anything happens","url":"https://addaly.com/learn/ai-video/latent-space-for-video"},"next":{"slug":"attention-across-frames","title":"Why a chair stays a chair and a face does not","url":"https://addaly.com/learn/ai-video/attention-across-frames"},"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 whole clip is made at once, not frame by frame', in Video with AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-video/denoising-a-block-of-time (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 whole clip is made at once, not frame by frame","courseTitle":"Video with AI","module":{"index":1,"slug":"how-video-models-work","title":"What the machine is actually doing"},"answer":null,"keyIdea":"A video model denoises every frame of the clip simultaneously across a fixed number of steps, which is why nothing can be streamed out early, why step count hits diminishing returns fast, and why distilled turbo models trade motion range for speed.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"line","title":"Where the step count stops buying anything","series":[{"label":"Full model","points":[[4,52],[8,72],[15,88],[20,93],[25,96],[30,97],[40,99],[50,100]]},{"label":"Distilled turbo","points":[[4,84],[8,90],[15,91],[20,91],[25,91],[30,91],[40,91],[50,91]]}],"xLabel":"Denoising steps","yLabel":"Quality reached, per cent of this model’s best","caption":"Cost is close to linear in steps and quality is not. The shape is what transfers, not the numbers: run the same shot at four step counts with the seed fixed and find your own model’s elbow. Note also where the turbo curve stops: it arrives fast and it stops lower, which is why a draft answers a different question from a final."},"alt":"Figure: Where the step count stops buying anything. Line chart of Quality reached, per cent of this model’s best against Denoising steps: Full model (4, 52) (8, 72) (15, 88) (20, 93) (25, 96) (30, 97) (40, 99) (50, 100); Distilled turbo (4, 84) (8, 90) (15, 91) (20, 91) (25, 91) (30, 91) (40, 91) (50, 91). Cost is close to linear in steps and quality is not. The shape is what transfers, not the numbers: run the same shot at four step counts with the seed fixed and find your own model’s elbow. Note also where the turbo curve stops: it arrives fast and it stops lower, which is why a draft answers a different question from a final."}},"blocks":[{"t":"h2","id":"starting-from-noise-all-at-once","runs":[{"text":"Starting from noise, all at once"}]},{"t":"p","runs":[{"text":"Generation begins with the entire latent block filled with random noise: all 30 latent frames of it. The model then makes a prediction about that noise, subtracts part of it, and repeats. After a fixed number of steps the block is clean and gets decoded."}]},{"t":"p","runs":[{"text":"The important word is "},{"text":"simultaneously","i":true},{"text":". At step 7, frame 1 and frame 120 are both 7 steps along. They are being refined together, seeing each other, agreeing with each other about where the wall is."}]},{"t":"p","runs":[{"text":"This is the opposite of how most people imagine it, and it explains several things that otherwise look arbitrary:"}]},{"t":"ul","items":[[{"text":"You cannot watch the first second while the rest renders.","b":true},{"text":" There is no first second yet. Progress bars that show a preview are decoding a partially denoised block, which is why early previews look like coloured fog."}],[{"text":"You cannot generate a bit more onto the end.","b":true},{"text":" The block size was fixed before step 1. Extending means a new run conditioned on a decoded frame, with all the photocopy damage that implies."}],[{"text":"Cost is decided at the start.","b":true},{"text":" Duration, resolution and step count are all fixed before any picture exists. There is no way to spend more only on the interesting part."}]]},{"t":"h2","id":"what-a-step-actually-buys","runs":[{"text":"What a step actually buys"}]},{"t":"p","runs":[{"text":"Each step is one full forward pass of the transformer over the whole latent. Twenty steps means twenty passes; fifty means fifty. Cost is close to linear in steps."}]},{"t":"p","runs":[{"text":"Quality is not. The curve is steep at the start and flat after a point that is lower than most people expect. On a typical modern video model:"}]},{"t":"ul","items":[[{"text":"4 to 8 steps: shapes and colour are right, motion is mushy, fine structure is unresolved."}],[{"text":"15 to 25 steps: most of the quality. This is where the majority of production work sits."}],[{"text":"30 to 50 steps: small improvements in fine detail and edge stability."}],[{"text":"Beyond 50: you are usually paying for a different random walk, not a better one."}]]},{"t":"p","runs":[{"text":"Run this yourself with a fixed seed. Generate the same shot at 10, 20, 30 and 50 steps and put the four side by side. Most people are surprised at how early the returns stop, and it is worth knowing your own model’s number rather than a general one."}]},{"t":"h2","id":"flow-matching-and-why-older-advice-does-not-transfer","runs":[{"text":"Flow matching, and why older advice does not transfer"}]},{"t":"p","runs":[{"text":"Early diffusion models used a noise schedule with hundreds of possible steps and samplers with names like DDIM and Euler-a. Most current video models are trained with "},{"text":"flow matching","b":true},{"text":" instead, which learns a straighter path from noise to data and converges in far fewer steps."}]},{"t":"p","runs":[{"text":"This matters practically because a lot of advice written for 2022-era image models is wrong here. Step counts of 100, high guidance values, and sampler names that no longer exist in the interface all come from that lineage. If a tutorial tells you to use 80 steps on a flow-matching video model, it was written for different machinery."}]},{"t":"p","runs":[{"text":"Flow-matching models also expose a "},{"text":"shift","b":true},{"text":" or "},{"text":"sigma shift","b":true},{"text":" parameter, which redistributes where the steps are spent along the path. Higher shift spends more effort in the early, structural part of the denoise. Video models generally want a higher shift than image models, and the tool usually sets a sensible default. Change it last, if at all."}]},{"t":"h2","id":"distilled-and-turbo-models","runs":[{"text":"Distilled and turbo models"}]},{"t":"p","runs":[{"text":"A distilled model is trained to imitate the output of a many-step model in very few steps: often 4 to 8. They are useful and they are how draft passes become affordable."}]},{"t":"p","runs":[{"text":"What they cost you is real and worth naming:"}]},{"t":"ul","items":[[{"text":"Reduced motion range.","b":true},{"text":" Distillation tends to pull toward the average trajectory. Turbo outputs are noticeably calmer, which sometimes flatters a shot and sometimes kills it."}],[{"text":"Weaker prompt adherence on unusual requests.","b":true},{"text":" Common instructions survive distillation. Odd ones get smoothed toward the common case."}],[{"text":"Guidance behaves differently.","b":true},{"text":" Most distilled models want a guidance scale near 1. Feeding them the value you use on the full model produces burnt contrast and frozen motion."}]]},{"t":"p","runs":[{"text":"The workflow that follows from this: draft on the turbo variant to test whether a motion idea works at all, then do the final on the full model with the same seed and still. The draft is not a preview of the final: it is a cheap answer to a different question."}]},{"t":"h2","id":"why-the-same-settings-give-different-pictures","runs":[{"text":"Why the same settings give different pictures"}]},{"t":"p","runs":[{"text":"Two runs with everything identical except the seed produce different clips because they start from different noise. That is expected. But two runs that you believe are identical can also differ, because the effective inputs include model version, precision, batch size and the specific GPU kernels used. The next lesson deals with that properly."}]},{"t":"h2","id":"the-free-path","runs":[{"text":"The free path"}],"role":"free"},{"t":"p","runs":[{"text":"Every point above can be tested on a free Hugging Face Space or a Colab session, because they are all about settings rather than about scale. Fix the still and the seed, move one number, compare. An afternoon of that teaches you more about your chosen model than any comparison video, and it costs a queue rather than a balance."}]},{"t":"rule"},{"t":"p","runs":[{"text":"Today: fix a seed and a still, generate at four step counts, and find the point on your own model where the improvement stops being visible. That number is now your production default."}]}],"check":{"question":[{"text":"A studio wants to show a client the first two seconds of a ten-second generation while the rest finishes rendering, to save waiting time. Why is this not possible?"}],"options":[{"text":[{"text":"The API does not expose partial results, though the model itself produces frames sequentially and a self-hosted setup could stream them out as they finish."}],"correct":false,"tempting":[{"text":"Blames the interface for an architectural property; a local run has the same block and the same problem."}]},{"text":[{"text":"Early frames are rendered at lower quality first and only refined at the end, so a partial result would misrepresent the shot."}],"correct":false,"tempting":[{"text":"Invents a progressive per-frame refinement order that flow-matching video diffusion does not use."}]},{"text":[{"text":"The decoder can only run on the complete latent block, although the latent frames themselves finish at different times."}],"correct":false,"tempting":[{"text":"Half right about the decoder but wrong about the latent: no group of latent frames finishes ahead of the others."}]},{"text":[{"text":"Every frame is at the same stage of denoising throughout, so before the last step the first two seconds are as unfinished as the rest."}],"correct":true,"tempting":null}],"order":[2,1,3,0],"why":null},"prev":{"slug":"latent-space-for-video","title":"Your clip is compressed fifty times before anything happens","minutes":10},"next":{"slug":"attention-across-frames","title":"Why a chair stays a chair and a face does not","minutes":10}}}