{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"slug":"the-img2img-ladder","title":"Three passes down the denoise ladder","url":"https://addaly.com/learn/ai-images/the-img2img-ladder","position":34,"of":69,"minutes":8,"answer":"Running several img2img passes at decreasing denoise strength converges on a finished image more reliably than one pass at a middling value, because each pass only has to make a small correction.","keyIdea":"Running several img2img passes at decreasing denoise strength converges on a finished image more reliably than one pass at a middling value, because each pass only has to make a small correction."},"body":{"format":"markdown","text":"## One big jump, or three small ones\n\nThe usual way people use img2img is a single pass: load the reference, pick a denoise value, generate, look, adjust the number, repeat. It works, and it is worse than the alternative for a reason that is easy to see once stated.\n\nAt a single middling value, 0.55, say, the model has to do two jobs at once: reorganise the image substantially, and finish it convincingly. It is being asked to make a large change and a refined one in the same pass, and one of them suffers. Either the composition survives and the surface stays rough, or the surface is beautiful and the composition has drifted.\n\nThe ladder splits the jobs.\n\n```text\npass 1   denoise 0.75   establish the look, accept large change\npass 2   denoise 0.50   fix what pass 1 got wrong, hold composition\npass 3   denoise 0.30   finish surface and light, change nothing structural\n```\n\nEach pass takes its predecessor as input. Same prompt, same seed or not, only the strength moves.\n\n> Figure: Three rungs, each starting from something better. Steps: 1. Pass one, denoise 0.75 (Establish the look. Accept large change.); 2. Pass two, denoise 0.50 (Fix what pass one got wrong. Hold the composition.); 3. Pass three, denoise 0.30 (Finish surface and light. Change nothing structural.). One pass at 0.55 asks for a large change and a refined one at the same time, and one of them suffers. Correct the compounded colour and contrast drift once, at the end, rather than on every rung.\n\n## What the numbers actually do\n\nDenoise strength decides how far back toward pure noise the image is pushed before denoising begins. At 1.0 nothing of the input survives; at 0.0 nothing changes. Between those, the number is roughly the proportion of the sampling schedule that runs.\n\nUseful landmarks, which hold across most models:\n\n- **0.15 to 0.25**: texture and grain only. The picture is identical.\n- **0.30 to 0.40**: surfaces, light quality, small detail. Composition untouched.\n- **0.45 to 0.60**: the working range. Real change, recognisable source. Objects can move slightly, expressions change, materials change completely.\n- **0.65 to 0.75**: the source is now a suggestion. Composition survives loosely, content may not.\n- **0.80+**: the input is contributing little beyond a colour distribution.\n\nThe ladder works because 0.30 applied to a nearly-right image is a completely different operation from 0.30 applied to a rough one. Each rung starts from something better than the last did.\n\n## Where this is the right tool\n\n**Bringing a rough render up.** A sketch, a 3D blockout, a crude collage. Start at 0.7 to make it photographic, then descend.\n\n**Changing a material.** Concrete to timber, plastic to ceramic. Two passes at 0.5 and 0.3 with the material named in the prompt.\n\n**Matching a look across a set.** Take frame nine, run it at 0.3 with the anchor's prompt, and its light and grade move toward the rest of the set without its content moving.\n\n**Rescuing a frame that is nearly there.** One pass at 0.35 often repairs the general softness that makes an image feel unfinished.\n\n## The way it goes wrong\n\n**Colour drift.** Each pass nudges the palette, usually warmer and more saturated, because each pass re-imposes the model's own bias. Three passes compound it. Check the grade after the ladder and correct it once at the end rather than fighting it on every rung.\n\n**Contrast crush.** The same compounding applies to contrast; blacks get blacker each time. If your third pass looks heavy, this is why.\n\n**Loss of the thing you liked.** The reason you chose that frame was some specific quality, and three passes of averaging tends to sand it off. Keep the original open in another window and compare, not just the previous rung.\n\n**Detail invention.** At 0.5 and above the model will add things that were not there: a pattern on a wall, a reflection, an extra object. On a product shot this is a correctness problem, not a style one.\n\n## The limitation to understand\n\nImg2img cannot move something across the frame. It has no notion of objects; it is denoising a whole latent toward a distribution consistent with the prompt. A chair on the left stays on the left, or becomes something else on the left.\n\nSo the ladder is the wrong tool for \"move the subject right and give me more sky\". That is outpainting, or a control image, or a new generation. When somebody spends forty minutes raising denoise to try to reposition a subject, this is what they have missed: by the time the strength is high enough to move it, the strength is also high enough that nothing else survives either.\n\nThe right division is that img2img changes *what things are made of and how they are lit*. Controls, which the rest of this block covers, change *where things are*. Reaching for the wrong one is the single commonest waste of time in this part of the work."},"check":{"question":"A designer runs img2img at increasing denoise values trying to shift the subject from the centre to the left of the frame. By 0.8 the subject has moved but everything else has changed too. Why does this approach fail?","options":[{"text":"Img2img denoises the whole latent toward the prompt rather than manipulating objects, so position only changes when the source stops surviving.","correct":true},{"text":"The denoise value applies uniformly across the frame, so shifting a subject requires a masked pass at a higher strength on the target region than on the surrounding area.","correct":false,"tempting":"A masked pass would regenerate one region but cannot relocate the subject into it; you would end up with two subjects or an empty space."},{"text":"The seed was not held constant between passes, so the composition was free to drift independently of the denoise strength being tested.","correct":false,"tempting":"Fixing the seed changes the starting noise but not the fact that img2img has no mechanism for moving content within the frame."},{"text":"The model's attention layers bind the subject to the centre of the latent, and only a positional control signal such as a pose skeleton can override that binding.","correct":false,"tempting":"Invents a centre-bias mechanism; the subject stays put because the source pixels stay put, not because of any positional preference in the architecture."}]},"prev":{"slug":"telling-it-the-composition","title":"The model is guessing your composition","url":"https://addaly.com/learn/ai-images/telling-it-the-composition"},"next":{"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"},"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) 'Three passes down the denoise ladder', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/the-img2img-ladder (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":"Three passes down the denoise ladder","courseTitle":"Image Generation, in Practice","module":{"index":4,"slug":"steering-with-images","title":"Steering with pictures instead of adjectives"},"answer":null,"keyIdea":"Running several img2img passes at decreasing denoise strength converges on a finished image more reliably than one pass at a middling value, because each pass only has to make a small correction.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"type":"flow","steps":[{"label":"Pass one, denoise 0.75","detail":"Establish the look. Accept large change."},{"label":"Pass two, denoise 0.50","detail":"Fix what pass one got wrong. Hold the composition."},{"label":"Pass three, denoise 0.30","detail":"Finish surface and light. Change nothing structural."}],"title":"Three rungs, each starting from something better","caption":"One pass at 0.55 asks for a large change and a refined one at the same time, and one of them suffers. Correct the compounded colour and contrast drift once, at the end, rather than on every rung."},"alt":"Figure: Three rungs, each starting from something better. Steps: 1. Pass one, denoise 0.75 (Establish the look. Accept large change.); 2. Pass two, denoise 0.50 (Fix what pass one got wrong. Hold the composition.); 3. Pass three, denoise 0.30 (Finish surface and light. Change nothing structural.). One pass at 0.55 asks for a large change and a refined one at the same time, and one of them suffers. Correct the compounded colour and contrast drift once, at the end, rather than on every rung."}},"blocks":[{"t":"h2","id":"one-big-jump-or-three-small-ones","runs":[{"text":"One big jump, or three small ones"}]},{"t":"p","runs":[{"text":"The usual way people use img2img is a single pass: load the reference, pick a denoise value, generate, look, adjust the number, repeat. It works, and it is worse than the alternative for a reason that is easy to see once stated."}]},{"t":"p","runs":[{"text":"At a single middling value, 0.55, say, the model has to do two jobs at once: reorganise the image substantially, and finish it convincingly. It is being asked to make a large change and a refined one in the same pass, and one of them suffers. Either the composition survives and the surface stays rough, or the surface is beautiful and the composition has drifted."}]},{"t":"p","runs":[{"text":"The ladder splits the jobs."}]},{"t":"code","lang":"text","label":"Text","text":"pass 1   denoise 0.75   establish the look, accept large change\npass 2   denoise 0.50   fix what pass 1 got wrong, hold composition\npass 3   denoise 0.30   finish surface and light, change nothing structural"},{"t":"p","runs":[{"text":"Each pass takes its predecessor as input. Same prompt, same seed or not, only the strength moves."}]},{"t":"h2","id":"what-the-numbers-actually-do","runs":[{"text":"What the numbers actually do"}]},{"t":"p","runs":[{"text":"Denoise strength decides how far back toward pure noise the image is pushed before denoising begins. At 1.0 nothing of the input survives; at 0.0 nothing changes. Between those, the number is roughly the proportion of the sampling schedule that runs."}]},{"t":"p","runs":[{"text":"Useful landmarks, which hold across most models:"}]},{"t":"ul","items":[[{"text":"0.15 to 0.25","b":true},{"text":": texture and grain only. The picture is identical."}],[{"text":"0.30 to 0.40","b":true},{"text":": surfaces, light quality, small detail. Composition untouched."}],[{"text":"0.45 to 0.60","b":true},{"text":": the working range. Real change, recognisable source. Objects can move slightly, expressions change, materials change completely."}],[{"text":"0.65 to 0.75","b":true},{"text":": the source is now a suggestion. Composition survives loosely, content may not."}],[{"text":"0.80+","b":true},{"text":": the input is contributing little beyond a colour distribution."}]]},{"t":"p","runs":[{"text":"The ladder works because 0.30 applied to a nearly-right image is a completely different operation from 0.30 applied to a rough one. Each rung starts from something better than the last did."}]},{"t":"h2","id":"where-this-is-the-right-tool","runs":[{"text":"Where this is the right tool"}]},{"t":"p","runs":[{"text":"Bringing a rough render up.","b":true},{"text":" A sketch, a 3D blockout, a crude collage. Start at 0.7 to make it photographic, then descend."}]},{"t":"p","runs":[{"text":"Changing a material.","b":true},{"text":" Concrete to timber, plastic to ceramic. Two passes at 0.5 and 0.3 with the material named in the prompt."}]},{"t":"p","runs":[{"text":"Matching a look across a set.","b":true},{"text":" Take frame nine, run it at 0.3 with the anchor’s prompt, and its light and grade move toward the rest of the set without its content moving."}]},{"t":"p","runs":[{"text":"Rescuing a frame that is nearly there.","b":true},{"text":" One pass at 0.35 often repairs the general softness that makes an image feel unfinished."}]},{"t":"h2","id":"the-way-it-goes-wrong","runs":[{"text":"The way it goes wrong"}]},{"t":"p","runs":[{"text":"Colour drift.","b":true},{"text":" Each pass nudges the palette, usually warmer and more saturated, because each pass re-imposes the model’s own bias. Three passes compound it. Check the grade after the ladder and correct it once at the end rather than fighting it on every rung."}]},{"t":"p","runs":[{"text":"Contrast crush.","b":true},{"text":" The same compounding applies to contrast; blacks get blacker each time. If your third pass looks heavy, this is why."}]},{"t":"p","runs":[{"text":"Loss of the thing you liked.","b":true},{"text":" The reason you chose that frame was some specific quality, and three passes of averaging tends to sand it off. Keep the original open in another window and compare, not just the previous rung."}]},{"t":"p","runs":[{"text":"Detail invention.","b":true},{"text":" At 0.5 and above the model will add things that were not there: a pattern on a wall, a reflection, an extra object. On a product shot this is a correctness problem, not a style one."}]},{"t":"h2","id":"the-limitation-to-understand","runs":[{"text":"The limitation to understand"}],"role":"limit"},{"t":"p","runs":[{"text":"Img2img cannot move something across the frame. It has no notion of objects; it is denoising a whole latent toward a distribution consistent with the prompt. A chair on the left stays on the left, or becomes something else on the left."}]},{"t":"p","runs":[{"text":"So the ladder is the wrong tool for “move the subject right and give me more sky”. That is outpainting, or a control image, or a new generation. When somebody spends forty minutes raising denoise to try to reposition a subject, this is what they have missed: by the time the strength is high enough to move it, the strength is also high enough that nothing else survives either."}]},{"t":"p","runs":[{"text":"The right division is that img2img changes "},{"text":"what things are made of and how they are lit","i":true},{"text":". Controls, which the rest of this block covers, change "},{"text":"where things are","i":true},{"text":". Reaching for the wrong one is the single commonest waste of time in this part of the work."}]}],"check":{"question":[{"text":"A designer runs img2img at increasing denoise values trying to shift the subject from the centre to the left of the frame. By 0.8 the subject has moved but everything else has changed too. Why does this approach fail?"}],"options":[{"text":[{"text":"Img2img denoises the whole latent toward the prompt rather than manipulating objects, so position only changes when the source stops surviving."}],"correct":true,"tempting":null},{"text":[{"text":"The denoise value applies uniformly across the frame, so shifting a subject requires a masked pass at a higher strength on the target region than on the surrounding area."}],"correct":false,"tempting":[{"text":"A masked pass would regenerate one region but cannot relocate the subject into it; you would end up with two subjects or an empty space."}]},{"text":[{"text":"The seed was not held constant between passes, so the composition was free to drift independently of the denoise strength being tested."}],"correct":false,"tempting":[{"text":"Fixing the seed changes the starting noise but not the fact that img2img has no mechanism for moving content within the frame."}]},{"text":[{"text":"The model’s attention layers bind the subject to the centre of the latent, and only a positional control signal such as a pose skeleton can override that binding."}],"correct":false,"tempting":[{"text":"Invents a centre-bias mechanism; the subject stays put because the source pixels stay put, not because of any positional preference in the architecture."}]}],"order":[1,2,3,0],"why":null},"prev":{"slug":"telling-it-the-composition","title":"The model is guessing your composition","minutes":9},"next":{"slug":"sketch-to-image","title":"A biro sketch is the most reliable composition tool you own","minutes":8}}}