{"course":{"slug":"ai-video","title":"Video with AI","url":"https://addaly.com/learn/ai-video","level":"intermediate","track":"create"},"lesson":{"slug":"draft-passes-and-quality-ladders","title":"Test the idea at a tenth of the price","url":"https://addaly.com/learn/ai-video/draft-passes-and-quality-ladders","position":62,"of":66,"minutes":8,"answer":"A draft generation at low resolution on a distilled model answers whether a motion idea works, which is a different question from whether the shot is good.","keyIdea":"A draft generation at low resolution on a distilled model answers whether a motion idea works, which is a different question from whether the shot is good, and answering it cheaply first is where most of the saving in a project lives."},"body":{"format":"markdown","text":"## Two different questions\n\nWhen you generate a shot you are asking one of two things:\n\n1. **Does this idea work at all?** Does the camera move suit the frame, does the subject motion read, is the shot even a shot?\n2. **Is this specific take good enough to ship?**\n\nThese cost wildly different amounts to answer, and most people pay the second price to get the first answer.\n\nA draft (lowest resolution, shortest duration, fastest or distilled model) answers question one for perhaps a tenth of the price. It tells you nine times as much per unit of money, because you can ask it nine times.\n\n## What a draft is for, and what it is not\n\n**A draft tells you:** whether the move suits the composition, whether the subject motion reads, roughly where the clip starts drifting, whether the shot has a purpose in the sequence.\n\n**A draft does not tell you:** the final quality, whether the face holds at full resolution, or what the finished grade will look like. Turbo models have reduced motion range and weaker prompt adherence, so a draft that looks calm may come back livelier at full quality, and one that ignored an instruction may obey it.\n\nBe clear which question you are answering. Treating a draft as a preview of the final is how people get surprised twice.\n\n## The ladder\n\nA typical project moves through three tiers:\n\n1. **Idea pass.** Distilled model, lowest resolution, shortest duration. Every shot on the list, once. Cheap enough to do the whole list.\n2. **Selection pass.** Full model, working resolution, on the shots that survived the assembly. Three or four takes each.\n3. **Final pass.** Full model, delivery resolution, same still, same prompt, same seed where the tool allows it, on the chosen take only.\n\nThe saving comes from the fact that tier one covers thirty shots and tier three covers the twenty that made the cut, at three or four takes rather than ten.\n\n> Figure: Three tiers, and where the saving actually is. Steps: 1. Idea pass, thirty shots (Distilled model, lowest resolution, shortest duration, once each. It answers whether the move suits the frame, which is a different question from whether the take is good enough to ship.); 2. Cut the drafts together (Put them on a timeline in order. On a thirty-shot project it is normal to lose six or seven here, and every one of those is a final generation you never pay for.); 3. Selection pass, three takes (Full model at working resolution, on the twenty-three shots that survived, and never beyond the kill number. On the fourth attempt you change the still, change the shot, or cut it.); 4. Final pass, chosen take only (Full model at delivery resolution, same still, same prompt, same seed where the tool allows it. Twenty-three generations, all of them for shots that are definitely in the piece.). One hundred and twenty-two generations in total, and the thirty cheapest are what decided the other ninety-two. Running the whole list at full quality from the start means paying final prices for the seven shots that were always going to be cut.\n\n## Edit the drafts\n\nThe step that makes the whole thing work. Put the draft clips on a timeline and cut them together.\n\nThis is a moving version of the piece at a small fraction of the cost, and it does the job the animatic did one level up: it tells you which shots are unnecessary, which are missing, and which are the wrong length. Every shot you delete at this stage is a final generation you never pay for.\n\nOn a thirty-shot project it is normal to cut six or seven shots at this point. That is a fifth of your generation budget recovered for the price of an afternoon.\n\n## The kill number\n\nSet it before you start: **three attempts per shot.**\n\nOn the fourth attempt you are not allowed to generate again. You must do one of three things:\n\n- Change the still.\n- Change the shot: a different size, a different move, a simpler idea.\n- Cut it from the list.\n\nWithout a kill number, one stubborn shot quietly eats the budget for eight others, and it is always a shot that did not matter. With one, the failure is bounded and the decision is made while you still have money.\n\nThe kill number is the single most effective budget control in this whole subject, and it is free.\n\n## Fix the still, not the prompt\n\nRestating it because this is where it costs money. Regenerating a still is on the order of one per cent of the cost of regenerating a clip. Almost every clip failure has a cheaper diagnosis one step upstream.\n\nWhen a generation disappoints, the first question is always \"is there something wrong with the frame?\" and the answer is usually yes.\n\n## The free path\n\nFree daily allowances are exactly right for the idea pass. Spend them on real shots from the real list at the cheapest settings rather than on tests you will throw away, and you learn far more per free credit.\n\nA phone plus two or three free tiers plus an afternoon is a complete idea pass for a thirty-shot project, and it costs nothing at all.\n\n---\n\nToday: pick your kill number, write it at the top of your shot list, and hold to it on the next shot that fights you."},"check":{"question":"A draft generated on a distilled model at low resolution comes back calm and slightly stiff. The operator concludes the motion idea does not work and cuts the shot. What is the error?","options":[{"text":"Low resolution reduces the latent's capacity for motion, so any draft will understate movement regardless of which model produced it.","correct":false,"tempting":"Resolution affects spatial detail; motion range is a property of the model and its training, not of the latent's width."},{"text":"Distilled models have reduced motion range, so a calm draft does not establish that the full model would be calm.","correct":true},{"text":"Drafts are generated at a shorter duration, so the clip was cut before the motion had time to develop into what the shot needed.","correct":false,"tempting":"Shorter drafts do truncate motion, and the stiffness described is a character of the output rather than an absence of time."},{"text":"The draft used a different seed from the one the final pass would use, so the two are separate random draws and no conclusion transfers between them.","correct":false,"tempting":"Seed differences make individual takes incomparable, while systematic properties of a model still transfer across draws."}]},"prev":{"slug":"previsualisation-and-client-approval","title":"Get the yes before you spend the money","url":"https://addaly.com/learn/ai-video/previsualisation-and-client-approval"},"next":{"slug":"batching-and-apis","title":"Queue everything, then judge it together","url":"https://addaly.com/learn/ai-video/batching-and-apis"},"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) 'Test the idea at a tenth of the price', in Video with AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-video/draft-passes-and-quality-ladders (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":"Test the idea at a tenth of the price","courseTitle":"Video with AI","module":{"index":7,"slug":"running-the-job","title":"Money, clients and running the job"},"answer":null,"keyIdea":"A draft generation at low resolution on a distilled model answers whether a motion idea works, which is a different question from whether the shot is good, and answering it cheaply first is where most of the saving in a project lives.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"flow","steps":[{"label":"Idea pass, thirty shots","detail":"Distilled model, lowest resolution, shortest duration, once each. It answers whether the move suits the frame, which is a different question from whether the take is good enough to ship."},{"label":"Cut the drafts together","detail":"Put them on a timeline in order. On a thirty-shot project it is normal to lose six or seven here, and every one of those is a final generation you never pay for."},{"label":"Selection pass, three takes","detail":"Full model at working resolution, on the twenty-three shots that survived, and never beyond the kill number. On the fourth attempt you change the still, change the shot, or cut it."},{"label":"Final pass, chosen take only","detail":"Full model at delivery resolution, same still, same prompt, same seed where the tool allows it. Twenty-three generations, all of them for shots that are definitely in the piece."}],"title":"Three tiers, and where the saving actually is","caption":"One hundred and twenty-two generations in total, and the thirty cheapest are what decided the other ninety-two. Running the whole list at full quality from the start means paying final prices for the seven shots that were always going to be cut."},"alt":"Figure: Three tiers, and where the saving actually is. Steps: 1. Idea pass, thirty shots (Distilled model, lowest resolution, shortest duration, once each. It answers whether the move suits the frame, which is a different question from whether the take is good enough to ship.); 2. Cut the drafts together (Put them on a timeline in order. On a thirty-shot project it is normal to lose six or seven here, and every one of those is a final generation you never pay for.); 3. Selection pass, three takes (Full model at working resolution, on the twenty-three shots that survived, and never beyond the kill number. On the fourth attempt you change the still, change the shot, or cut it.); 4. Final pass, chosen take only (Full model at delivery resolution, same still, same prompt, same seed where the tool allows it. Twenty-three generations, all of them for shots that are definitely in the piece.). One hundred and twenty-two generations in total, and the thirty cheapest are what decided the other ninety-two. Running the whole list at full quality from the start means paying final prices for the seven shots that were always going to be cut."}},"blocks":[{"t":"h2","id":"two-different-questions","runs":[{"text":"Two different questions"}]},{"t":"p","runs":[{"text":"When you generate a shot you are asking one of two things:"}]},{"t":"ol","start":1,"items":[[{"text":"Does this idea work at all?","b":true},{"text":" Does the camera move suit the frame, does the subject motion read, is the shot even a shot?"}],[{"text":"Is this specific take good enough to ship?","b":true}]]},{"t":"p","runs":[{"text":"These cost wildly different amounts to answer, and most people pay the second price to get the first answer."}]},{"t":"p","runs":[{"text":"A draft (lowest resolution, shortest duration, fastest or distilled model) answers question one for perhaps a tenth of the price. It tells you nine times as much per unit of money, because you can ask it nine times."}]},{"t":"h2","id":"what-a-draft-is-for-and-what-it-is-not","runs":[{"text":"What a draft is for, and what it is not"}]},{"t":"p","runs":[{"text":"A draft tells you:","b":true},{"text":" whether the move suits the composition, whether the subject motion reads, roughly where the clip starts drifting, whether the shot has a purpose in the sequence."}]},{"t":"p","runs":[{"text":"A draft does not tell you:","b":true},{"text":" the final quality, whether the face holds at full resolution, or what the finished grade will look like. Turbo models have reduced motion range and weaker prompt adherence, so a draft that looks calm may come back livelier at full quality, and one that ignored an instruction may obey it."}]},{"t":"p","runs":[{"text":"Be clear which question you are answering. Treating a draft as a preview of the final is how people get surprised twice."}]},{"t":"h2","id":"the-ladder","runs":[{"text":"The ladder"}]},{"t":"p","runs":[{"text":"A typical project moves through three tiers:"}]},{"t":"ol","start":1,"items":[[{"text":"Idea pass.","b":true},{"text":" Distilled model, lowest resolution, shortest duration. Every shot on the list, once. Cheap enough to do the whole list."}],[{"text":"Selection pass.","b":true},{"text":" Full model, working resolution, on the shots that survived the assembly. Three or four takes each."}],[{"text":"Final pass.","b":true},{"text":" Full model, delivery resolution, same still, same prompt, same seed where the tool allows it, on the chosen take only."}]]},{"t":"p","runs":[{"text":"The saving comes from the fact that tier one covers thirty shots and tier three covers the twenty that made the cut, at three or four takes rather than ten."}]},{"t":"h2","id":"edit-the-drafts","runs":[{"text":"Edit the drafts"}]},{"t":"p","runs":[{"text":"The step that makes the whole thing work. Put the draft clips on a timeline and cut them together."}]},{"t":"p","runs":[{"text":"This is a moving version of the piece at a small fraction of the cost, and it does the job the animatic did one level up: it tells you which shots are unnecessary, which are missing, and which are the wrong length. Every shot you delete at this stage is a final generation you never pay for."}]},{"t":"p","runs":[{"text":"On a thirty-shot project it is normal to cut six or seven shots at this point. That is a fifth of your generation budget recovered for the price of an afternoon."}]},{"t":"h2","id":"the-kill-number","runs":[{"text":"The kill number"}]},{"t":"p","runs":[{"text":"Set it before you start: "},{"text":"three attempts per shot.","b":true}]},{"t":"p","runs":[{"text":"On the fourth attempt you are not allowed to generate again. You must do one of three things:"}]},{"t":"ul","items":[[{"text":"Change the still."}],[{"text":"Change the shot: a different size, a different move, a simpler idea."}],[{"text":"Cut it from the list."}]]},{"t":"p","runs":[{"text":"Without a kill number, one stubborn shot quietly eats the budget for eight others, and it is always a shot that did not matter. With one, the failure is bounded and the decision is made while you still have money."}]},{"t":"p","runs":[{"text":"The kill number is the single most effective budget control in this whole subject, and it is free."}]},{"t":"h2","id":"fix-the-still-not-the-prompt","runs":[{"text":"Fix the still, not the prompt"}]},{"t":"p","runs":[{"text":"Restating it because this is where it costs money. Regenerating a still is on the order of one per cent of the cost of regenerating a clip. Almost every clip failure has a cheaper diagnosis one step upstream."}]},{"t":"p","runs":[{"text":"When a generation disappoints, the first question is always “is there something wrong with the frame?” and the answer is usually yes."}]},{"t":"h2","id":"the-free-path","runs":[{"text":"The free path"}],"role":"free"},{"t":"p","runs":[{"text":"Free daily allowances are exactly right for the idea pass. Spend them on real shots from the real list at the cheapest settings rather than on tests you will throw away, and you learn far more per free credit."}]},{"t":"p","runs":[{"text":"A phone plus two or three free tiers plus an afternoon is a complete idea pass for a thirty-shot project, and it costs nothing at all."}]},{"t":"rule"},{"t":"p","runs":[{"text":"Today: pick your kill number, write it at the top of your shot list, and hold to it on the next shot that fights you."}]}],"check":{"question":[{"text":"A draft generated on a distilled model at low resolution comes back calm and slightly stiff. The operator concludes the motion idea does not work and cuts the shot. What is the error?"}],"options":[{"text":[{"text":"Low resolution reduces the latent’s capacity for motion, so any draft will understate movement regardless of which model produced it."}],"correct":false,"tempting":[{"text":"Resolution affects spatial detail; motion range is a property of the model and its training, not of the latent’s width."}]},{"text":[{"text":"Distilled models have reduced motion range, so a calm draft does not establish that the full model would be calm."}],"correct":true,"tempting":null},{"text":[{"text":"Drafts are generated at a shorter duration, so the clip was cut before the motion had time to develop into what the shot needed."}],"correct":false,"tempting":[{"text":"Shorter drafts do truncate motion, and the stiffness described is a character of the output rather than an absence of time."}]},{"text":[{"text":"The draft used a different seed from the one the final pass would use, so the two are separate random draws and no conclusion transfers between them."}],"correct":false,"tempting":[{"text":"Seed differences make individual takes incomparable, while systematic properties of a model still transfer across draws."}]}],"order":[0,1,3,2],"why":null},"prev":{"slug":"previsualisation-and-client-approval","title":"Get the yes before you spend the money","minutes":9},"next":{"slug":"batching-and-apis","title":"Queue everything, then judge it together","minutes":9}}}