{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"slug":"sketch-to-image","title":"A biro sketch is the most reliable composition tool you own","url":"https://addaly.com/learn/ai-images/sketch-to-image","position":35,"of":69,"minutes":8,"answer":"Five minutes of rough drawing photographed on a phone gives the model a composition it will follow, and it requires no drawing skill because the conditioning reads structure rather than quality of line.","keyIdea":"Five minutes of rough drawing photographed on a phone gives the model a composition it will follow, and it requires no drawing skill because the conditioning reads structure rather than quality of line."},"body":{"format":"markdown","text":"## The objection first\n\n\"I cannot draw.\" Almost nobody reading this can, and it does not matter. What the model needs from a sketch is where things are, how big they are, and roughly what shape. A shaky rectangle in the lower left is a table. Two circles and a line are a person. The conditioning model extracts structure, and structure survives bad drawing entirely.\n\nThis is worth taking seriously because it is the highest-leverage technique in this block, and the barrier to it is psychological rather than technical.\n\n## The method\n\n1. **Draw it.** Paper, biro, five minutes. Or in Krita or GIMP with a mouse, which is free and just as good. Block shapes, not detail. Mark the horizon and where the light comes from.\n2. **Photograph it** with a phone, flat and evenly lit, or scan it.\n3. **Clean it up**: raise contrast so the lines are black on white. Thirty seconds with Levels in GIMP.\n4. **Feed it as a control**, using scribble or lineart conditioning, at strength 0.6 to 0.8 and ending around 0.6 to 0.7 of the steps.\n5. **Prompt normally**, describing the scene in the slot structure from the previous block.\n\nThe scribble control type exists precisely for this and is forgiving of rough input. Lineart is cleaner and stricter. Canny edge detection is the wrong choice for a sketch, because it will faithfully preserve every wobble in your line.\n\n## Why it beats prompting for composition\n\nA prompt describes a composition in words the model interprets probabilistically. \"A figure on the left, a doorway behind, the table in the foreground\" is three spatial relationships, and spatial relationships are exactly what language conditioning binds worst. The previous block listed this as a structural failure: three or more objects in stated positions do not reliably land.\n\nA sketch is not a description. It is the answer. The model does not have to infer the arrangement, because the arrangement is the input.\n\nThe practical consequence is that a five-minute sketch replaces about forty minutes of prompt attempts, and it lands the first time.\n\n## Where to let go\n\nThe single most common mistake is over-drawing. A detailed sketch at high control strength produces a result that looks exactly like a coloured-in drawing, because it is one. The model has been given no room to contribute anything.\n\nSo:\n\n- Draw the **big shapes only**. Five to ten elements, no more.\n- Mark **the horizon and the light direction**: these two lines carry an enormous amount.\n- Leave faces, hands and detail completely blank. Draw an oval for a head.\n- Use **strength 0.6 to 0.75, not 1.0**, and **end the control around 0.6 to 0.7** so the last third of the sampling is free to make it look real.\n\nThat end-step setting is the one people miss. Running a control to the final step means the structure is enforced right through the phase where fine texture is formed, and the result acquires a flat, traced quality. Releasing it early gives you the composition you drew and a surface the model built.\n\n## The iteration this unlocks\n\nSketching changes the shape of the work, not just its accuracy. Instead of generating and judging, you draw three compositions in ten minutes, generate each, and compare. You are making compositional decisions on paper, where they take seconds and cost nothing, rather than in a generator where each attempt is a render and a wait.\n\nThis is how art direction has always worked (thumbnails first, then execution) and it transfers here without modification.\n\n## What a sketch cannot give you\n\nIt fixes arrangement. It does not fix perspective, and this catches people. If your drawn box does not converge correctly, the model will often follow your wrong perspective and produce a subtly impossible room that nobody can quite explain. The conditioning is faithful, including to your mistakes.\n\nFor anything where perspective must be correct (architecture, interiors, a product on a surface) the next tool in this block is the answer: build the blockout in three dimensions and let the software get the perspective right. A sketch is for arrangement. A depth render is for geometry. Knowing which problem you have saves you from fixing a perspective error by redrawing, which mostly produces a different perspective error."},"check":{"question":"Why should a sketch control be set to end around 60 to 70% of the steps rather than running to completion?","options":[{"text":"Because control models lose accuracy in the final steps, so continuing past that point introduces structural drift rather than preventing it.","correct":false,"tempting":"Backwards: the control remains accurate throughout, and the problem is that it stays accurate during the phase where texture should be forming freely."},{"text":"Because the final steps form fine texture, and enforcing the drawn structure through them makes the result look traced.","correct":true},{"text":"Because running a control to the end consumes additional memory in the later steps, where activation sizes are at their peak for the sampling process.","correct":false,"tempting":"Control cost is roughly constant across steps, and memory has nothing to do with the flat appearance the setting exists to prevent."},{"text":"Because ending the control early allows the prompt to take over and correct any compositional errors present in the original sketch.","correct":false,"tempting":"The prompt does gain influence, but it will not restructure the composition at that late stage; what it gains is freedom over surface, not over arrangement."}]},"prev":{"slug":"the-img2img-ladder","title":"Three passes down the denoise ladder","url":"https://addaly.com/learn/ai-images/the-img2img-ladder"},"next":{"slug":"colour-blocking","title":"Paint flat shapes, then let the model make them real","url":"https://addaly.com/learn/ai-images/colour-blocking"},"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) 'A biro sketch is the most reliable composition tool you own', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/sketch-to-image (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":"A biro sketch is the most reliable composition tool you own","courseTitle":"Image Generation, in Practice","module":{"index":4,"slug":"steering-with-images","title":"Steering with pictures instead of adjectives"},"answer":null,"keyIdea":"Five minutes of rough drawing photographed on a phone gives the model a composition it will follow, and it requires no drawing skill because the conditioning reads structure rather than quality of line.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"the-objection-first","runs":[{"text":"The objection first"}]},{"t":"p","runs":[{"text":"“I cannot draw.” Almost nobody reading this can, and it does not matter. What the model needs from a sketch is where things are, how big they are, and roughly what shape. A shaky rectangle in the lower left is a table. Two circles and a line are a person. The conditioning model extracts structure, and structure survives bad drawing entirely."}]},{"t":"p","runs":[{"text":"This is worth taking seriously because it is the highest-leverage technique in this block, and the barrier to it is psychological rather than technical."}]},{"t":"h2","id":"the-method","runs":[{"text":"The method"}]},{"t":"ol","start":1,"items":[[{"text":"Draw it.","b":true},{"text":" Paper, biro, five minutes. Or in Krita or GIMP with a mouse, which is free and just as good. Block shapes, not detail. Mark the horizon and where the light comes from."}],[{"text":"Photograph it","b":true},{"text":" with a phone, flat and evenly lit, or scan it."}],[{"text":"Clean it up","b":true},{"text":": raise contrast so the lines are black on white. Thirty seconds with Levels in GIMP."}],[{"text":"Feed it as a control","b":true},{"text":", using scribble or lineart conditioning, at strength 0.6 to 0.8 and ending around 0.6 to 0.7 of the steps."}],[{"text":"Prompt normally","b":true},{"text":", describing the scene in the slot structure from the previous block."}]]},{"t":"p","runs":[{"text":"The scribble control type exists precisely for this and is forgiving of rough input. Lineart is cleaner and stricter. Canny edge detection is the wrong choice for a sketch, because it will faithfully preserve every wobble in your line."}]},{"t":"h2","id":"why-it-beats-prompting-for-composition","runs":[{"text":"Why it beats prompting for composition"}]},{"t":"p","runs":[{"text":"A prompt describes a composition in words the model interprets probabilistically. “A figure on the left, a doorway behind, the table in the foreground” is three spatial relationships, and spatial relationships are exactly what language conditioning binds worst. The previous block listed this as a structural failure: three or more objects in stated positions do not reliably land."}]},{"t":"p","runs":[{"text":"A sketch is not a description. It is the answer. The model does not have to infer the arrangement, because the arrangement is the input."}]},{"t":"p","runs":[{"text":"The practical consequence is that a five-minute sketch replaces about forty minutes of prompt attempts, and it lands the first time."}]},{"t":"h2","id":"where-to-let-go","runs":[{"text":"Where to let go"}]},{"t":"p","runs":[{"text":"The single most common mistake is over-drawing. A detailed sketch at high control strength produces a result that looks exactly like a coloured-in drawing, because it is one. The model has been given no room to contribute anything."}]},{"t":"p","runs":[{"text":"So:"}]},{"t":"ul","items":[[{"text":"Draw the "},{"text":"big shapes only","b":true},{"text":". Five to ten elements, no more."}],[{"text":"Mark "},{"text":"the horizon and the light direction","b":true},{"text":": these two lines carry an enormous amount."}],[{"text":"Leave faces, hands and detail completely blank. Draw an oval for a head."}],[{"text":"Use "},{"text":"strength 0.6 to 0.75, not 1.0","b":true},{"text":", and "},{"text":"end the control around 0.6 to 0.7","b":true},{"text":" so the last third of the sampling is free to make it look real."}]]},{"t":"p","runs":[{"text":"That end-step setting is the one people miss. Running a control to the final step means the structure is enforced right through the phase where fine texture is formed, and the result acquires a flat, traced quality. Releasing it early gives you the composition you drew and a surface the model built."}]},{"t":"h2","id":"the-iteration-this-unlocks","runs":[{"text":"The iteration this unlocks"}]},{"t":"p","runs":[{"text":"Sketching changes the shape of the work, not just its accuracy. Instead of generating and judging, you draw three compositions in ten minutes, generate each, and compare. You are making compositional decisions on paper, where they take seconds and cost nothing, rather than in a generator where each attempt is a render and a wait."}]},{"t":"p","runs":[{"text":"This is how art direction has always worked (thumbnails first, then execution) and it transfers here without modification."}]},{"t":"h2","id":"what-a-sketch-cannot-give-you","runs":[{"text":"What a sketch cannot give you"}]},{"t":"p","runs":[{"text":"It fixes arrangement. It does not fix perspective, and this catches people. If your drawn box does not converge correctly, the model will often follow your wrong perspective and produce a subtly impossible room that nobody can quite explain. The conditioning is faithful, including to your mistakes."}]},{"t":"p","runs":[{"text":"For anything where perspective must be correct (architecture, interiors, a product on a surface) the next tool in this block is the answer: build the blockout in three dimensions and let the software get the perspective right. A sketch is for arrangement. A depth render is for geometry. Knowing which problem you have saves you from fixing a perspective error by redrawing, which mostly produces a different perspective error."}]}],"check":{"question":[{"text":"Why should a sketch control be set to end around 60 to 70% of the steps rather than running to completion?"}],"options":[{"text":[{"text":"Because control models lose accuracy in the final steps, so continuing past that point introduces structural drift rather than preventing it."}],"correct":false,"tempting":[{"text":"Backwards: the control remains accurate throughout, and the problem is that it stays accurate during the phase where texture should be forming freely."}]},{"text":[{"text":"Because the final steps form fine texture, and enforcing the drawn structure through them makes the result look traced."}],"correct":true,"tempting":null},{"text":[{"text":"Because running a control to the end consumes additional memory in the later steps, where activation sizes are at their peak for the sampling process."}],"correct":false,"tempting":[{"text":"Control cost is roughly constant across steps, and memory has nothing to do with the flat appearance the setting exists to prevent."}]},{"text":[{"text":"Because ending the control early allows the prompt to take over and correct any compositional errors present in the original sketch."}],"correct":false,"tempting":[{"text":"The prompt does gain influence, but it will not restructure the composition at that late stage; what it gains is freedom over surface, not over arrangement."}]}],"order":[2,0,1,3],"why":null},"prev":{"slug":"the-img2img-ladder","title":"Three passes down the denoise ladder","minutes":8},"next":{"slug":"colour-blocking","title":"Paint flat shapes, then let the model make them real","minutes":8}}}