{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"slug":"when-the-model-will-not-do-it","title":"Recognising the twenty minutes that will never converge","url":"https://addaly.com/learn/ai-images/when-the-model-will-not-do-it","position":29,"of":69,"minutes":7,"answer":"Some requests fail for structural reasons rather than for want of a better prompt, and the professional skill is recognising them inside three attempts and switching method instead of continuing.","keyIdea":"Some requests fail for structural reasons rather than for want of a better prompt, and the professional skill is recognising them inside three attempts and switching method instead of continuing."},"body":{"format":"markdown","text":"## The sunk-cost hour\n\nYou have been at it for forty minutes. Each attempt is nearly right. One more phrasing will do it. It will not, and you knew by attempt four.\n\nCertain requests do not respond to prompting at all, and the reason is always the same shape: the model is producing a plausible image, and plausibility does not include the constraint you care about. Recognising the category takes three attempts. Here are the categories.\n\n## The list\n\n**Exact counts above about four.** \"Seven birds\" gives you five, or nine. The model has no counting mechanism; number words shift a distribution over how many things appear. Below four it usually lands; above, it drifts. Fix: generate the scene without the count and composite, or accept whatever number appears.\n\n**Legible specific text.** Newer models can produce short phrases reasonably often, and even those fail on unusual words, small sizes and long strings. Fix: set the type yourself over a clean plate. This is not a workaround; it is how the industry works anyway, and the repair block covers it.\n\n**Precise spatial relations between three or more objects.** \"The red cup to the left of the blue book, behind the lamp\" binds unreliably. Fix: sketch it and condition on the sketch, which is the whole of the next block.\n\n**A specific real object.** Your client's actual bottle, with its actual label. The model has never seen it. Fix: photograph and composite.\n\n**A named living person's likeness.** Blocked on hosted services, unreliable locally, and a separate legal question covered later. Fix: do not, unless you have permission and a real reason.\n\n**Consistent identity across many images from prompts alone.** Fix: adapter or LoRA, which the consistency block covers.\n\n**Perfect symmetry or exact mechanical geometry.** Logos, technical diagrams, repeating grids, architectural plans. Fix: vector tools. Inkscape is free and does in five minutes what forty prompts cannot.\n\n**Anatomy in unusual configurations.** Hands interacting with objects, several people touching, a body from a rare angle. Fix: simpler pose, or repair, or pose conditioning.\n\n## The three-attempt rule\n\nAdopt a rule and hold yourself to it: **three attempts, then change method.**\n\nAttempt one, the straightforward prompt. Attempt two, rephrased with the failure addressed directly. Attempt three, one structural change: different model, much simpler framing. If all three fail in the same way, it is not a prompting problem. The identical failure repeating is the diagnostic: random failures suggest you are close, and a consistent one says the model cannot represent what you want.\n\n## What \"change method\" means\n\nThe alternatives are the rest of this course, and they are all stronger tools than a fourth prompt:\n\n- Condition on an image: sketch, depth, pose, edges.\n- Generate the parts separately and composite.\n- Photograph the difficult element.\n- Make it in a different kind of software entirely: Inkscape for vectors, Blender for geometry, a text tool for text.\n- Change the brief so the impossible thing is not in frame.\n\nThat last one deserves more respect than it gets. If the composition requires a hand holding a fork at an unusual angle, and nothing works, a version where the hand is out of frame is often a better picture anyway. Constraints have improved compositions since long before any of this existed.\n\n## The caveat\n\nThis list has a shelf life. Text rendering was impossible, then unreliable, and is now workable on several models for short phrases. Counting and spatial binding have both improved with newer architectures. Anything written today about what models cannot do will be partly wrong within a year, and you should treat the list as a starting point rather than a law.\n\nWhat does not change is the method: three attempts, watch for the *same* failure repeating, then switch. That rule stays correct regardless of which items are on the list this year, and it is the actual skill.\n\n## Telling the difference between close and stuck\n\nThe distinction that matters is between a *random* failure and a *systematic* one, and you can usually read it from three attempts.\n\n**Random failure**: each attempt fails differently. The hands are wrong in one, the light is flat in another, the composition is off in the third. This means the model can produce what you want and you are sampling around it. Keep going, or run a contact sheet.\n\n**Systematic failure**: each attempt fails the same way. Always the wrong number. Always the label illegible. Always the same anatomical impossibility at the same joint. This means the thing you want is outside what the model represents, and the twentieth attempt will look like the third.\n\nWrite down the failure after each attempt, in three words. Doing so converts a vague sense of frustration into a readable pattern, and the pattern answers the question.\n\n## Say it out loud on a shared job\n\nIf you are working with a client or a team, name it early: \"This element is not going to come out of the generator: I'll shoot it and composite.\" Said on day one, that is expertise. Said on day four after a silent struggle, it reads as an excuse, even though it is the same sentence and the same fact."},"check":{"question":"Three attempts at \"seven birds on a wire\" give five, nine and four birds. What does the consistency of the failure tell you?","options":[{"text":"That the prompt is ambiguous about arrangement, and specifying the spacing along the wire would allow the model to resolve the count correctly.","correct":false,"tempting":"Adds detail to a description that was already clear; the count is unreliable because nothing in the model enumerates objects, not because the arrangement was underspecified."},{"text":"That the seed is unlucky, and sampling considerably more seeds will eventually produce a frame containing exactly seven of them.","correct":false,"tempting":"Occasionally true by chance, and it is the reasoning that produces the sunk-cost hour; hunting a distribution for an exact value is not a method."},{"text":"That quantity is represented as a distribution rather than counted, so no rephrasing fixes it and the number has to come from compositing.","correct":true},{"text":"That the guidance scale is too low for the numeric part of the prompt to be enforced against the rest of the description around it.","correct":false,"tempting":"Guidance amplifies the whole conditioning signal, so raising it produces a more confident wrong number rather than a correct one."}]},"prev":{"slug":"one-change-at-a-time","title":"Change one thing, or you have learned nothing","url":"https://addaly.com/learn/ai-images/one-change-at-a-time"},"next":{"slug":"prompt-from-an-image","title":"Interrogating a picture for vocabulary","url":"https://addaly.com/learn/ai-images/prompt-from-an-image"},"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) 'Recognising the twenty minutes that will never converge', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/when-the-model-will-not-do-it (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":"Recognising the twenty minutes that will never converge","courseTitle":"Image Generation, in Practice","module":{"index":3,"slug":"first-frames","title":"Getting to a frame worth keeping"},"answer":null,"keyIdea":"Some requests fail for structural reasons rather than for want of a better prompt, and the professional skill is recognising them inside three attempts and switching method instead of continuing.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"the-sunk-cost-hour","runs":[{"text":"The sunk-cost hour"}]},{"t":"p","runs":[{"text":"You have been at it for forty minutes. Each attempt is nearly right. One more phrasing will do it. It will not, and you knew by attempt four."}]},{"t":"p","runs":[{"text":"Certain requests do not respond to prompting at all, and the reason is always the same shape: the model is producing a plausible image, and plausibility does not include the constraint you care about. Recognising the category takes three attempts. Here are the categories."}]},{"t":"h2","id":"the-list","runs":[{"text":"The list"}]},{"t":"p","runs":[{"text":"Exact counts above about four.","b":true},{"text":" “Seven birds” gives you five, or nine. The model has no counting mechanism; number words shift a distribution over how many things appear. Below four it usually lands; above, it drifts. Fix: generate the scene without the count and composite, or accept whatever number appears."}]},{"t":"p","runs":[{"text":"Legible specific text.","b":true},{"text":" Newer models can produce short phrases reasonably often, and even those fail on unusual words, small sizes and long strings. Fix: set the type yourself over a clean plate. This is not a workaround; it is how the industry works anyway, and the repair block covers it."}]},{"t":"p","runs":[{"text":"Precise spatial relations between three or more objects.","b":true},{"text":" “The red cup to the left of the blue book, behind the lamp” binds unreliably. Fix: sketch it and condition on the sketch, which is the whole of the next block."}]},{"t":"p","runs":[{"text":"A specific real object.","b":true},{"text":" Your client’s actual bottle, with its actual label. The model has never seen it. Fix: photograph and composite."}]},{"t":"p","runs":[{"text":"A named living person’s likeness.","b":true},{"text":" Blocked on hosted services, unreliable locally, and a separate legal question covered later. Fix: do not, unless you have permission and a real reason."}]},{"t":"p","runs":[{"text":"Consistent identity across many images from prompts alone.","b":true},{"text":" Fix: adapter or LoRA, which the consistency block covers."}]},{"t":"p","runs":[{"text":"Perfect symmetry or exact mechanical geometry.","b":true},{"text":" Logos, technical diagrams, repeating grids, architectural plans. Fix: vector tools. Inkscape is free and does in five minutes what forty prompts cannot."}]},{"t":"p","runs":[{"text":"Anatomy in unusual configurations.","b":true},{"text":" Hands interacting with objects, several people touching, a body from a rare angle. Fix: simpler pose, or repair, or pose conditioning."}]},{"t":"h2","id":"the-three-attempt-rule","runs":[{"text":"The three-attempt rule"}]},{"t":"p","runs":[{"text":"Adopt a rule and hold yourself to it: "},{"text":"three attempts, then change method.","b":true}]},{"t":"p","runs":[{"text":"Attempt one, the straightforward prompt. Attempt two, rephrased with the failure addressed directly. Attempt three, one structural change: different model, much simpler framing. If all three fail in the same way, it is not a prompting problem. The identical failure repeating is the diagnostic: random failures suggest you are close, and a consistent one says the model cannot represent what you want."}]},{"t":"h2","id":"what-change-method-means","runs":[{"text":"What “change method” means"}]},{"t":"p","runs":[{"text":"The alternatives are the rest of this course, and they are all stronger tools than a fourth prompt:"}]},{"t":"ul","items":[[{"text":"Condition on an image: sketch, depth, pose, edges."}],[{"text":"Generate the parts separately and composite."}],[{"text":"Photograph the difficult element."}],[{"text":"Make it in a different kind of software entirely: Inkscape for vectors, Blender for geometry, a text tool for text."}],[{"text":"Change the brief so the impossible thing is not in frame."}]]},{"t":"p","runs":[{"text":"That last one deserves more respect than it gets. If the composition requires a hand holding a fork at an unusual angle, and nothing works, a version where the hand is out of frame is often a better picture anyway. Constraints have improved compositions since long before any of this existed."}]},{"t":"h2","id":"the-caveat","runs":[{"text":"The caveat"}]},{"t":"p","runs":[{"text":"This list has a shelf life. Text rendering was impossible, then unreliable, and is now workable on several models for short phrases. Counting and spatial binding have both improved with newer architectures. Anything written today about what models cannot do will be partly wrong within a year, and you should treat the list as a starting point rather than a law."}]},{"t":"p","runs":[{"text":"What does not change is the method: three attempts, watch for the "},{"text":"same","i":true},{"text":" failure repeating, then switch. That rule stays correct regardless of which items are on the list this year, and it is the actual skill."}]},{"t":"h2","id":"telling-the-difference-between-close-and-stuck","runs":[{"text":"Telling the difference between close and stuck"}]},{"t":"p","runs":[{"text":"The distinction that matters is between a "},{"text":"random","i":true},{"text":" failure and a "},{"text":"systematic","i":true},{"text":" one, and you can usually read it from three attempts."}]},{"t":"p","runs":[{"text":"Random failure","b":true},{"text":": each attempt fails differently. The hands are wrong in one, the light is flat in another, the composition is off in the third. This means the model can produce what you want and you are sampling around it. Keep going, or run a contact sheet."}]},{"t":"p","runs":[{"text":"Systematic failure","b":true},{"text":": each attempt fails the same way. Always the wrong number. Always the label illegible. Always the same anatomical impossibility at the same joint. This means the thing you want is outside what the model represents, and the twentieth attempt will look like the third."}]},{"t":"p","runs":[{"text":"Write down the failure after each attempt, in three words. Doing so converts a vague sense of frustration into a readable pattern, and the pattern answers the question."}]},{"t":"h2","id":"say-it-out-loud-on-a-shared-job","runs":[{"text":"Say it out loud on a shared job"}]},{"t":"p","runs":[{"text":"If you are working with a client or a team, name it early: “This element is not going to come out of the generator: I’ll shoot it and composite.” Said on day one, that is expertise. Said on day four after a silent struggle, it reads as an excuse, even though it is the same sentence and the same fact."}]}],"check":{"question":[{"text":"Three attempts at “seven birds on a wire” give five, nine and four birds. What does the consistency of the failure tell you?"}],"options":[{"text":[{"text":"That the prompt is ambiguous about arrangement, and specifying the spacing along the wire would allow the model to resolve the count correctly."}],"correct":false,"tempting":[{"text":"Adds detail to a description that was already clear; the count is unreliable because nothing in the model enumerates objects, not because the arrangement was underspecified."}]},{"text":[{"text":"That the seed is unlucky, and sampling considerably more seeds will eventually produce a frame containing exactly seven of them."}],"correct":false,"tempting":[{"text":"Occasionally true by chance, and it is the reasoning that produces the sunk-cost hour; hunting a distribution for an exact value is not a method."}]},{"text":[{"text":"That quantity is represented as a distribution rather than counted, so no rephrasing fixes it and the number has to come from compositing."}],"correct":true,"tempting":null},{"text":[{"text":"That the guidance scale is too low for the numeric part of the prompt to be enforced against the rest of the description around it."}],"correct":false,"tempting":[{"text":"Guidance amplifies the whole conditioning signal, so raising it produces a more confident wrong number rather than a correct one."}]}],"order":[3,0,1,2],"why":null},"prev":{"slug":"one-change-at-a-time","title":"Change one thing, or you have learned nothing","minutes":7},"next":{"slug":"prompt-from-an-image","title":"Interrogating a picture for vocabulary","minutes":7}}}