{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"slug":"the-same-face-twice","title":"The same face twice is the hard part","url":"https://addaly.com/learn/ai-images/the-same-face-twice","position":52,"of":69,"minutes":9,"answer":"One-photo identity adapters get you a family resemblance; a specific real person needs a LoRA trained on varied photographs, or their actual photographed face composited in an editor.","keyIdea":"One-photo identity adapters get you a family resemblance; a specific real person needs a LoRA trained on varied photographs, or their actual photographed face composited in an editor."},"body":{"format":"markdown","text":"## The requirement that separates demos from jobs\n\nOne striking image is easy now. Ten images of *the same person* in different scenes, or forty shots of *the* product with *the* label, is the requirement almost every real commission has, and it is hard.\n\nThere are four approaches. They are not alternatives to each other so much as rungs, and it is worth knowing exactly how far up each one gets you.\n\n## Rung one: seeds. Almost useless for this\n\nA seed is the starting noise. Same seed, same model, same prompt, same settings gives you the same image, bit for bit. Change one word and the face changes.\n\nSo seeds reproduce an image, not an identity. They matter enormously for controlled experiments (change one clause, keep the seed, see what that clause did) and they matter for going back to a frame you liked. They do not give you the same person in a new pose. People waste days on this.\n\n## Rung two: identity adapters. Family resemblance\n\nFeed one photograph of a face; get back a generated face that resembles it. This is IP-Adapter FaceID, InstantID and PuLID in the open world, and \"character reference\" or a trained identity object like Higgsfield's Soul ID on hosted platforms.\n\nHonest ceiling: these carry the broad geometry of a face (the proportions, the general structure) and they are **family-resemblance accurate, not identity accurate**. The person's sister comes out. Skin texture drifts, apparent age drifts, and unusual angles break it. Expression is often flattened toward neutral.\n\nThat is fine for a stylised character, a game avatar, an illustrated persona. It is not fine when the client is looking at a picture of themselves, because humans have dedicated neural hardware for faces and will spot the wrongness without being able to name it.\n\nThe platform versions that ask for a set of photographs rather than one are doing something closer to rung three, and get closer to the person.\n\n## Rung three: a LoRA. The real answer\n\nTrain a small adapter on images of your subject. Twenty to thirty images, an hour or so of training, and the concept becomes something you can name in a prompt and place anywhere.\n\nWhat actually determines whether it works:\n\n- **Variety beats quantity.** Twenty photographs across different lighting, angles, distances and clothing beat two hundred taken in one session. If every training image was shot against the same blue wall, the LoRA has learned the blue wall as part of the person, and it will follow them into every scene.\n- **Crop consistently and caption honestly.** Caption what varies, not what stays constant.\n- **Stop early.** Overtraining bakes in the training set. The face becomes rigid and one of the photographs starts reappearing wholesale.\n- **Free to do.** Kohya_ss, OneTrainer and ai-toolkit are free, and Kaggle's or Colab's free GPU hours are enough for a character LoRA. [Fine-tuning](/learn/fine-tuning) covers the general method.\n\n## Rung four: do it in an editor\n\nGenerate the scene. Composite the real photographed head, or the real product, onto it. Match colour, match grain, match the light direction.\n\nThis is unglamorous and it is the most reliable thing in this lesson. Half of what looks like extraordinary AI work in commercial studios is a generated background with a photographed subject on top. [Image editing](/learn/image-editing) is the skill that makes it invisible.\n\n## Products are harder than faces\n\nA face can be five per cent wrong and read as the person. A logo cannot be five per cent wrong: a wrong letterform is a counterfeit. Bottle proportions, the exact green of a brand, the position of a seam: all of these must be exact and none of these are things a model holds precisely.\n\nThe professional answer: photograph the product, generate the environment, composite. Do not ask a model to draw a product that exists. Ask it to build a room to put the photograph in.\n\n## The consent line\n\nA LoRA of a person is a tool for producing unlimited images of that person doing things they did not do. Training one on someone who has not agreed is not a technical decision, and in several countries it is now a legal one. Get permission in writing, and say what the images will be used for. [Making things with AI](/learn/making-things-with-ai) sets out the wider consent and likeness position.\n\nToday: if you have a recurring character, count your reference photographs. If they were all taken in one place on one afternoon, you do not yet have a training set: you have one photograph repeated."},"check":{"question":"A client wants their founder to appear in six generated scenes. The designer uses a single-photo face reference feature and the client says the results look like the founder's cousin. What is the correct diagnosis?","options":[{"text":"The reference photograph is too low-resolution and a studio-quality headshot would fix it.","correct":false,"tempting":"Blames the input's quality for a ceiling that belongs to the method itself."},{"text":"The seed should be locked across all six generations so the face stays constant.","correct":false,"tempting":"Confuses reproducing one image exactly with carrying an identity across six different images."},{"text":"Single-photo adapters transfer the broad geometry of a face and are family-resemblance accurate rather than identity accurate; a specific real person needs a LoRA trained on 20-30 varied photographs, or her real photographed head composited in.","correct":true},{"text":"The prompt needs to describe her features in detail alongside the reference image.","correct":false,"tempting":"Expects a text description to carry identity information that the pixels themselves did not."}]},"prev":{"slug":"tiling-and-large-canvases","title":"Getting to poster size without a second horizon","url":"https://addaly.com/learn/ai-images/tiling-and-large-canvases"},"next":{"slug":"the-character-sheet","title":"Build the reference before you build the images","url":"https://addaly.com/learn/ai-images/the-character-sheet"},"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 same face twice is the hard part', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/the-same-face-twice (Accessed: 31 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 same face twice is the hard part","courseTitle":"Image Generation, in Practice","module":{"index":6,"slug":"consistency","title":"The same thing, twelve times"},"answer":null,"keyIdea":"One-photo identity adapters get you a family resemblance; a specific real person needs a LoRA trained on varied photographs, or their actual photographed face composited in an editor.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"the-requirement-that-separates-demos-from-jobs","runs":[{"text":"The requirement that separates demos from jobs"}]},{"t":"p","runs":[{"text":"One striking image is easy now. Ten images of "},{"text":"the same person","i":true},{"text":" in different scenes, or forty shots of "},{"text":"the","i":true},{"text":" product with "},{"text":"the","i":true},{"text":" label, is the requirement almost every real commission has, and it is hard."}]},{"t":"p","runs":[{"text":"There are four approaches. They are not alternatives to each other so much as rungs, and it is worth knowing exactly how far up each one gets you."}]},{"t":"h2","id":"rung-one-seeds-almost-useless-for-this","runs":[{"text":"Rung one: seeds. Almost useless for this"}]},{"t":"p","runs":[{"text":"A seed is the starting noise. Same seed, same model, same prompt, same settings gives you the same image, bit for bit. Change one word and the face changes."}]},{"t":"p","runs":[{"text":"So seeds reproduce an image, not an identity. They matter enormously for controlled experiments (change one clause, keep the seed, see what that clause did) and they matter for going back to a frame you liked. They do not give you the same person in a new pose. People waste days on this."}]},{"t":"h2","id":"rung-two-identity-adapters-family-resemblance","runs":[{"text":"Rung two: identity adapters. Family resemblance"}]},{"t":"p","runs":[{"text":"Feed one photograph of a face; get back a generated face that resembles it. This is IP-Adapter FaceID, InstantID and PuLID in the open world, and “character reference” or a trained identity object like Higgsfield’s Soul ID on hosted platforms."}]},{"t":"p","runs":[{"text":"Honest ceiling: these carry the broad geometry of a face (the proportions, the general structure) and they are "},{"text":"family-resemblance accurate, not identity accurate","b":true},{"text":". The person’s sister comes out. Skin texture drifts, apparent age drifts, and unusual angles break it. Expression is often flattened toward neutral."}]},{"t":"p","runs":[{"text":"That is fine for a stylised character, a game avatar, an illustrated persona. It is not fine when the client is looking at a picture of themselves, because humans have dedicated neural hardware for faces and will spot the wrongness without being able to name it."}]},{"t":"p","runs":[{"text":"The platform versions that ask for a set of photographs rather than one are doing something closer to rung three, and get closer to the person."}]},{"t":"h2","id":"rung-three-a-lora-the-real-answer","runs":[{"text":"Rung three: a LoRA. The real answer"}]},{"t":"p","runs":[{"text":"Train a small adapter on images of your subject. Twenty to thirty images, an hour or so of training, and the concept becomes something you can name in a prompt and place anywhere."}]},{"t":"p","runs":[{"text":"What actually determines whether it works:"}]},{"t":"ul","items":[[{"text":"Variety beats quantity.","b":true},{"text":" Twenty photographs across different lighting, angles, distances and clothing beat two hundred taken in one session. If every training image was shot against the same blue wall, the LoRA has learned the blue wall as part of the person, and it will follow them into every scene."}],[{"text":"Crop consistently and caption honestly.","b":true},{"text":" Caption what varies, not what stays constant."}],[{"text":"Stop early.","b":true},{"text":" Overtraining bakes in the training set. The face becomes rigid and one of the photographs starts reappearing wholesale."}],[{"text":"Free to do.","b":true},{"text":" Kohya_ss, OneTrainer and ai-toolkit are free, and Kaggle’s or Colab’s free GPU hours are enough for a character LoRA. "},{"text":"Fine-tuning","href":"https://addaly.com/learn/fine-tuning"},{"text":" covers the general method."}]]},{"t":"h2","id":"rung-four-do-it-in-an-editor","runs":[{"text":"Rung four: do it in an editor"}]},{"t":"p","runs":[{"text":"Generate the scene. Composite the real photographed head, or the real product, onto it. Match colour, match grain, match the light direction."}]},{"t":"p","runs":[{"text":"This is unglamorous and it is the most reliable thing in this lesson. Half of what looks like extraordinary AI work in commercial studios is a generated background with a photographed subject on top. "},{"text":"Image editing","href":"https://addaly.com/learn/image-editing"},{"text":" is the skill that makes it invisible."}]},{"t":"h2","id":"products-are-harder-than-faces","runs":[{"text":"Products are harder than faces"}]},{"t":"p","runs":[{"text":"A face can be five per cent wrong and read as the person. A logo cannot be five per cent wrong: a wrong letterform is a counterfeit. Bottle proportions, the exact green of a brand, the position of a seam: all of these must be exact and none of these are things a model holds precisely."}]},{"t":"p","runs":[{"text":"The professional answer: photograph the product, generate the environment, composite. Do not ask a model to draw a product that exists. Ask it to build a room to put the photograph in."}]},{"t":"h2","id":"the-consent-line","runs":[{"text":"The consent line"}]},{"t":"p","runs":[{"text":"A LoRA of a person is a tool for producing unlimited images of that person doing things they did not do. Training one on someone who has not agreed is not a technical decision, and in several countries it is now a legal one. Get permission in writing, and say what the images will be used for. "},{"text":"Making things with AI","href":"https://addaly.com/learn/making-things-with-ai"},{"text":" sets out the wider consent and likeness position."}]},{"t":"p","runs":[{"text":"Today: if you have a recurring character, count your reference photographs. If they were all taken in one place on one afternoon, you do not yet have a training set: you have one photograph repeated."}]}],"check":{"question":[{"text":"A client wants their founder to appear in six generated scenes. The designer uses a single-photo face reference feature and the client says the results look like the founder’s cousin. What is the correct diagnosis?"}],"options":[{"text":[{"text":"The reference photograph is too low-resolution and a studio-quality headshot would fix it."}],"correct":false,"tempting":[{"text":"Blames the input’s quality for a ceiling that belongs to the method itself."}]},{"text":[{"text":"The seed should be locked across all six generations so the face stays constant."}],"correct":false,"tempting":[{"text":"Confuses reproducing one image exactly with carrying an identity across six different images."}]},{"text":[{"text":"Single-photo adapters transfer the broad geometry of a face and are family-resemblance accurate rather than identity accurate; a specific real person needs a LoRA trained on 20-30 varied photographs, or her real photographed head composited in."}],"correct":true,"tempting":null},{"text":[{"text":"The prompt needs to describe her features in detail alongside the reference image."}],"correct":false,"tempting":[{"text":"Expects a text description to carry identity information that the pixels themselves did not."}]}],"order":[1,3,2,0],"why":null},"prev":{"slug":"tiling-and-large-canvases","title":"Getting to poster size without a second horizon","minutes":8},"next":{"slug":"the-character-sheet","title":"Build the reference before you build the images","minutes":8}}}