{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"slug":"adapter-or-lora","title":"The decision, with the numbers","url":"https://addaly.com/learn/ai-images/adapter-or-lora","position":56,"of":69,"minutes":7,"answer":"An adapter gives roughly seventy per cent of an identity in under a minute and a trained LoRA gives most of the rest in about an hour.","keyIdea":"An adapter gives roughly seventy per cent of an identity in under a minute and a trained LoRA gives most of the rest in about an hour, so the decision turns on how many images the identity has to survive and how closely it will be examined."},"body":{"format":"markdown","text":"## Stop treating it as a preference\n\nPeople adopt one method and defend it. It is a decision with two inputs: **how many images**, and **how closely will it be looked at**.\n\n## The comparison\n\n**Image-prompt adapter (IP-Adapter and relatives)**\n\n- Setup: seconds. Load a reference, set a weight.\n- Cost: nothing beyond generation.\n- Result: a family resemblance. Bone structure and general appearance carry; small features drift between images.\n- Flexibility: works with any prompt, though high weights start to compete with it.\n- Requires: one image.\n\n**Trained LoRA**\n\n> Figure: The decision, with the numbers rather than the preference. A two by two grid, rows Adapter and Trained LoRA, columns What it costs and What you get: Adapter and What it costs: Seconds; Adapter and What you get: About 70%; Trained LoRA and What it costs: About an hour; Trained LoRA and What you get: About 95%. One to three images seen separately: adapter. Four to twelve side by side: LoRA, because a set is always compared and adapter drift becomes visible the moment frames sit next to each other.\n\n- Setup: about an hour end to end, once you have done it before.\n- Cost: free on Kaggle or Colab; a few pennies of electricity locally.\n- Result: identity holds tightly, including small features, across dozens of images.\n- Flexibility: composes cleanly with prompts and controls at sensible strengths.\n- Requires: 15 to 30 varied images, which for a fictional character means building the sheet first.\n\n## The decision rule\n\n**One to three images, seen separately**: adapter. The hour of training buys nothing a viewer will register.\n\n**Four to twelve images, a campaign, side by side**: LoRA. Adapter drift becomes visible when frames are compared, and a set is always compared.\n\n**A recurring character across months**: LoRA, without question. It is an asset you keep, and the alternative is re-deriving the character every time.\n\n**Close scrutiny at large size**: LoRA, plus face compositing for the frames that matter most.\n\n**You do not yet know if the character works**: adapter first. Test the concept cheaply, then train once it is approved. Training a LoRA for a character the client then rejects is an hour spent on nothing.\n\n## Using both\n\nThe strongest arrangement for a demanding set is both together: the LoRA at 0.7 to 0.8 carrying the identity, and a face adapter at 0.3 to 0.4 from the single best reference nudging it toward one specific rendering.\n\nThe caution is the stacking budget from the previous block. A LoRA and an adapter are both constraints, and adding a pose control and a depth control on top leaves nothing free. If four seeds look identical, drop something.\n\n## What neither solves\n\nBoth fix *who the person is*. Neither fixes *how they are lit, framed or graded*, and those are what make a set look like a set. A perfectly consistent face across twelve differently-lit, differently-framed images still reads as twelve unrelated pictures.\n\nThat is the next lesson, and it is the part people skip. Identity consistency is the technically interesting problem, so it absorbs all the attention, while the visual system (one light, one lens, one palette, one grain) is what a viewer actually perceives as coherence and costs almost nothing to enforce.\n\nPut differently: a set with a slightly drifting face and a rigorously consistent look reads as consistent. A set with a perfect face and twelve different lighting setups does not. If you have limited time, spend it on the system.\n\n## Testing which one you need, cheaply\n\nYou do not have to guess. The test takes fifteen minutes and settles the question with evidence rather than opinion.\n\nGenerate four frames with an adapter at 0.75: the four most different situations the brief contains, so different angles, different distances, different lighting. Lay them out side by side at the size they will be delivered.\n\nThen ask one question: **would a stranger say this is the same person?** Not \"can I see the differences\", because you can, having stared at them. A stranger, at delivery size, in the order the audience will see them.\n\nIf yes, ship the adapter and keep the hour. If no, train. If you are unsure, show it to somebody who has not been in the project: the answer is usually immediate and it is usually right.\n\n## Strength values worth starting from\n\nFor an adapter: 0.6 to 0.8 for a face. Below 0.5 the identity barely registers; above 0.9 it starts overriding the prompt and the pose.\n\nFor a LoRA: 0.6 to 0.85, depending on how heavily trained it is. A LoRA that needs 1.0 to show up is undertrained; one that dominates at 0.5 is overtrained, and the middle epoch you did not pick is probably the right one.\n\nWrite the working value next to the LoRA's filename, because you will not remember it in a month and rediscovering it costs a dozen renders."},"check":{"question":"When is an image-prompt adapter the right choice over a trained LoRA?","options":[{"text":"When the base model already renders the desired character type well, so only a light adjustment toward the specific reference is required.","correct":false,"tempting":"Base-model competence does not reduce drift between frames; an approximate identity is approximate whether or not the model handles the category well."},{"text":"When the character appears in one or two images seen separately, where drift between frames will never be observed.","correct":true},{"text":"When the character must appear alongside controls such as pose and depth, since a LoRA cannot be combined with structural conditioning in the same generation.","correct":false,"tempting":"LoRAs compose with controls routinely; the constraint is the total strength budget, which applies to adapters equally."},{"text":"When there are fewer than fifteen reference images available, which is the minimum a LoRA needs in order to train at all.","correct":false,"tempting":"Fifteen is a practical floor for good results rather than a hard requirement, and for a fictional character the sheet can simply be generated."}]},"prev":{"slug":"dataset-hygiene","title":"Four dataset faults, and the LoRA each one ruins","url":"https://addaly.com/learn/ai-images/dataset-hygiene"},"next":{"slug":"product-consistency","title":"A product is harder than a face","url":"https://addaly.com/learn/ai-images/product-consistency"},"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 decision, with the numbers', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/adapter-or-lora (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 decision, with the numbers","courseTitle":"Image Generation, in Practice","module":{"index":6,"slug":"consistency","title":"The same thing, twelve times"},"answer":null,"keyIdea":"An adapter gives roughly seventy per cent of an identity in under a minute and a trained LoRA gives most of the rest in about an hour, so the decision turns on how many images the identity has to survive and how closely it will be examined.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure","figure":{"cols":["What it costs","What you get"],"rows":["Adapter","Trained LoRA"],"type":"matrix","cells":[["Seconds","About 70%"],["About an hour","About 95%"]],"title":"The decision, with the numbers rather than the preference","caption":"One to three images seen separately: adapter. Four to twelve side by side: LoRA, because a set is always compared and adapter drift becomes visible the moment frames sit next to each other.","cellLabels":[["one reference image","family resemblance"],["15–30 varied images","small features hold"]]},"alt":"Figure: The decision, with the numbers rather than the preference. A two by two grid, rows Adapter and Trained LoRA, columns What it costs and What you get: Adapter and What it costs: Seconds; Adapter and What you get: About 70%; Trained LoRA and What it costs: About an hour; Trained LoRA and What you get: About 95%. One to three images seen separately: adapter. Four to twelve side by side: LoRA, because a set is always compared and adapter drift becomes visible the moment frames sit next to each other."}},"blocks":[{"t":"h2","id":"stop-treating-it-as-a-preference","runs":[{"text":"Stop treating it as a preference"}]},{"t":"p","runs":[{"text":"People adopt one method and defend it. It is a decision with two inputs: "},{"text":"how many images","b":true},{"text":", and "},{"text":"how closely will it be looked at","b":true},{"text":"."}]},{"t":"h2","id":"the-comparison","runs":[{"text":"The comparison"}]},{"t":"p","runs":[{"text":"Image-prompt adapter (IP-Adapter and relatives)","b":true}]},{"t":"ul","items":[[{"text":"Setup: seconds. Load a reference, set a weight."}],[{"text":"Cost: nothing beyond generation."}],[{"text":"Result: a family resemblance. Bone structure and general appearance carry; small features drift between images."}],[{"text":"Flexibility: works with any prompt, though high weights start to compete with it."}],[{"text":"Requires: one image."}]]},{"t":"p","runs":[{"text":"Trained LoRA","b":true}]},{"t":"ul","items":[[{"text":"Setup: about an hour end to end, once you have done it before."}],[{"text":"Cost: free on Kaggle or Colab; a few pennies of electricity locally."}],[{"text":"Result: identity holds tightly, including small features, across dozens of images."}],[{"text":"Flexibility: composes cleanly with prompts and controls at sensible strengths."}],[{"text":"Requires: 15 to 30 varied images, which for a fictional character means building the sheet first."}]]},{"t":"h2","id":"the-decision-rule","runs":[{"text":"The decision rule"}]},{"t":"p","runs":[{"text":"One to three images, seen separately","b":true},{"text":": adapter. The hour of training buys nothing a viewer will register."}]},{"t":"p","runs":[{"text":"Four to twelve images, a campaign, side by side","b":true},{"text":": LoRA. Adapter drift becomes visible when frames are compared, and a set is always compared."}]},{"t":"p","runs":[{"text":"A recurring character across months","b":true},{"text":": LoRA, without question. It is an asset you keep, and the alternative is re-deriving the character every time."}]},{"t":"p","runs":[{"text":"Close scrutiny at large size","b":true},{"text":": LoRA, plus face compositing for the frames that matter most."}]},{"t":"p","runs":[{"text":"You do not yet know if the character works","b":true},{"text":": adapter first. Test the concept cheaply, then train once it is approved. Training a LoRA for a character the client then rejects is an hour spent on nothing."}]},{"t":"h2","id":"using-both","runs":[{"text":"Using both"}]},{"t":"p","runs":[{"text":"The strongest arrangement for a demanding set is both together: the LoRA at 0.7 to 0.8 carrying the identity, and a face adapter at 0.3 to 0.4 from the single best reference nudging it toward one specific rendering."}]},{"t":"p","runs":[{"text":"The caution is the stacking budget from the previous block. A LoRA and an adapter are both constraints, and adding a pose control and a depth control on top leaves nothing free. If four seeds look identical, drop something."}]},{"t":"h2","id":"what-neither-solves","runs":[{"text":"What neither solves"}]},{"t":"p","runs":[{"text":"Both fix "},{"text":"who the person is","i":true},{"text":". Neither fixes "},{"text":"how they are lit, framed or graded","i":true},{"text":", and those are what make a set look like a set. A perfectly consistent face across twelve differently-lit, differently-framed images still reads as twelve unrelated pictures."}]},{"t":"p","runs":[{"text":"That is the next lesson, and it is the part people skip. Identity consistency is the technically interesting problem, so it absorbs all the attention, while the visual system (one light, one lens, one palette, one grain) is what a viewer actually perceives as coherence and costs almost nothing to enforce."}]},{"t":"p","runs":[{"text":"Put differently: a set with a slightly drifting face and a rigorously consistent look reads as consistent. A set with a perfect face and twelve different lighting setups does not. If you have limited time, spend it on the system."}]},{"t":"h2","id":"testing-which-one-you-need-cheaply","runs":[{"text":"Testing which one you need, cheaply"}]},{"t":"p","runs":[{"text":"You do not have to guess. The test takes fifteen minutes and settles the question with evidence rather than opinion."}]},{"t":"p","runs":[{"text":"Generate four frames with an adapter at 0.75: the four most different situations the brief contains, so different angles, different distances, different lighting. Lay them out side by side at the size they will be delivered."}]},{"t":"p","runs":[{"text":"Then ask one question: "},{"text":"would a stranger say this is the same person?","b":true},{"text":" Not “can I see the differences”, because you can, having stared at them. A stranger, at delivery size, in the order the audience will see them."}]},{"t":"p","runs":[{"text":"If yes, ship the adapter and keep the hour. If no, train. If you are unsure, show it to somebody who has not been in the project: the answer is usually immediate and it is usually right."}]},{"t":"h2","id":"strength-values-worth-starting-from","runs":[{"text":"Strength values worth starting from"}]},{"t":"p","runs":[{"text":"For an adapter: 0.6 to 0.8 for a face. Below 0.5 the identity barely registers; above 0.9 it starts overriding the prompt and the pose."}]},{"t":"p","runs":[{"text":"For a LoRA: 0.6 to 0.85, depending on how heavily trained it is. A LoRA that needs 1.0 to show up is undertrained; one that dominates at 0.5 is overtrained, and the middle epoch you did not pick is probably the right one."}]},{"t":"p","runs":[{"text":"Write the working value next to the LoRA’s filename, because you will not remember it in a month and rediscovering it costs a dozen renders."}]}],"check":{"question":[{"text":"When is an image-prompt adapter the right choice over a trained LoRA?"}],"options":[{"text":[{"text":"When the base model already renders the desired character type well, so only a light adjustment toward the specific reference is required."}],"correct":false,"tempting":[{"text":"Base-model competence does not reduce drift between frames; an approximate identity is approximate whether or not the model handles the category well."}]},{"text":[{"text":"When the character appears in one or two images seen separately, where drift between frames will never be observed."}],"correct":true,"tempting":null},{"text":[{"text":"When the character must appear alongside controls such as pose and depth, since a LoRA cannot be combined with structural conditioning in the same generation."}],"correct":false,"tempting":[{"text":"LoRAs compose with controls routinely; the constraint is the total strength budget, which applies to adapters equally."}]},{"text":[{"text":"When there are fewer than fifteen reference images available, which is the minimum a LoRA needs in order to train at all."}],"correct":false,"tempting":[{"text":"Fifteen is a practical floor for good results rather than a hard requirement, and for a fictional character the sheet can simply be generated."}]}],"order":[3,2,0,1],"why":null},"prev":{"slug":"dataset-hygiene","title":"Four dataset faults, and the LoRA each one ruins","minutes":8},"next":{"slug":"product-consistency","title":"A product is harder than a face","minutes":8}}}