{"course":{"slug":"ai-images","title":"Image Generation, in Practice","url":"https://addaly.com/learn/ai-images","level":"beginner","track":"create"},"lesson":{"slug":"the-character-sheet","title":"Build the reference before you build the images","url":"https://addaly.com/learn/ai-images/the-character-sheet","position":53,"of":69,"minutes":8,"answer":"A character sheet turns an identity from something you hope recurs into an asset you condition on, and building it first is what makes every later frame a matter of reuse rather than of luck.","keyIdea":"A character sheet turns an identity from something you hope recurs into an asset you condition on, and building it first is what makes every later frame a matter of reuse rather than of luck."},"body":{"format":"markdown","text":"## Working backwards from the problem\n\nYou need one person across eight images. The naive approach generates image one, likes the face, and then tries to describe that face well enough to get it again. It does not work, and the reason is simple: a face is not describable at the precision required. \"Woman, 30s, dark hair, round face, brown eyes\" describes several million people, and the model will show you a different one each time.\n\nThe method that works is to stop trying to describe and start building a **reference sheet**: a small set of images of the same person from several angles, in neutral conditions, that you use as input for everything afterwards.\n\nAnimation and games have worked this way for a century. The character sheet comes before the scenes.\n\n## What goes on the sheet\n\nFive to eight images of one person:\n\n- Front, neutral expression, even light\n- Three-quarter left, three-quarter right\n- Profile\n- One at a distance, full body, showing proportions and posture\n- Two with expression: a smile, a look away\n- One in the project's actual lighting, which becomes your style anchor\n\nNeutral is the point. Flat even light, plain background, no strong grade. You want the identity separated from everything else, because everything else will change in the final frames.\n\n## Getting the first face\n\nYou need one image of somebody who does not exist. Three routes:\n\n**Generate and select.** Run a contact sheet of thirty portraits with a description, pick the one you want. Free, fast, and the face is the model's average of your words, which sometimes means it is an unremarkable face.\n\n**Generate and refine.** Pick a face, then use seed variation at strength 0.1–0.2 to explore its neighbourhood until it is specific rather than generic. Worth the extra ten minutes, because a distinctive face holds better across images than an average one: there is more for the mechanism to lock onto.\n\n**Photograph a consenting person and stylise.** Legitimate with permission, and the identity block later in this course sets out why permission is not optional.\n\n## Building the other angles\n\nOnce you have the front view, you need the rest, and they must be the same person.\n\n1. Take the front view as an IP-Adapter face reference at weight 0.7 to 0.85.\n2. Add a pose control with the target angle.\n3. Generate. Check identity. Adjust adapter weight.\n4. Repeat for each angle.\n\nExpect drift. The profile is the hardest (there is least information about a profile in a front view) and it often takes eight attempts. Accept a close match rather than a perfect one; the sheet's job is to give a *combined* signal, and small variation across it is tolerable and sometimes helpful.\n\n## Using the sheet\n\nTwo ways, and you will use both.\n\n**As adapter input.** Feed two or three sheet images as face references for each new generation. Fast, needs no training, and gives a family resemblance that holds well enough for many jobs.\n\n**As LoRA training data.** The sheet is a training set. Fifteen to thirty images trains a character LoRA that holds identity far more tightly. The next two lessons cover it.\n\nThe efficient path is to build the sheet, use adapters for the first few frames to check the character works in context, and train a LoRA only if the job needs more than an adapter gives.\n\n## The limit, stated honestly\n\nA character sheet gets you a person who is *recognisably the same* across images. It does not get you a person who is *identically the same*. Bone structure holds; small features drift. The nose is slightly different. The distance between the eyes moves by a pixel or two. In a set viewed one image at a time, nobody notices. In a set viewed side by side at large size, some people do.\n\nThe gap narrows with a well-trained LoRA and narrows further with careful repair, and it does not close. If your brief requires an identical face, a recurring character in a long campaign, a brand mascot that will be scrutinised, budget for face compositing on top of everything else: generate the body and scene, and place a consistent face in from a fixed set, matched for angle and light.\n\nThat is more work and it is the only method that is actually reliable. Knowing this at the brief stage, rather than at frame nine, is the difference between a plan and a crisis."},"check":{"question":"Why build a character sheet before generating any of the final frames?","options":[{"text":"Because an identity cannot be described in words precisely enough, so it has to become an image you condition on.","correct":true},{"text":"Because generating the sheet first establishes a consistent seed that can then be reused across every frame in the set in order to hold the face stable.","correct":false,"tempting":"Seeds do not carry identity; the same seed with a different prompt or resolution produces an unrelated person."},{"text":"Because a sheet built from the final frames would inherit their lighting and grade, which contaminates the identity signal used for subsequent conditioning.","correct":false,"tempting":"A real reason for keeping sheet images neutral, but it explains how to build the sheet rather than why it comes first."},{"text":"Because training a LoRA requires images at several angles, and those angles cannot be produced once the character has already appeared in finished work.","correct":false,"tempting":"Angles can be generated at any point, and many jobs never train a LoRA at all."}]},"prev":{"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"},"next":{"slug":"training-a-lora-in-practice","title":"Training a character LoRA on hardware you can borrow","url":"https://addaly.com/learn/ai-images/training-a-lora-in-practice"},"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) 'Build the reference before you build the images', in Image Generation, in Practice. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-images/the-character-sheet (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":"Build the reference before you build the images","courseTitle":"Image Generation, in Practice","module":{"index":6,"slug":"consistency","title":"The same thing, twelve times"},"answer":null,"keyIdea":"A character sheet turns an identity from something you hope recurs into an asset you condition on, and building it first is what makes every later frame a matter of reuse rather than of luck.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"working-backwards-from-the-problem","runs":[{"text":"Working backwards from the problem"}]},{"t":"p","runs":[{"text":"You need one person across eight images. The naive approach generates image one, likes the face, and then tries to describe that face well enough to get it again. It does not work, and the reason is simple: a face is not describable at the precision required. “Woman, 30s, dark hair, round face, brown eyes” describes several million people, and the model will show you a different one each time."}]},{"t":"p","runs":[{"text":"The method that works is to stop trying to describe and start building a "},{"text":"reference sheet","b":true},{"text":": a small set of images of the same person from several angles, in neutral conditions, that you use as input for everything afterwards."}]},{"t":"p","runs":[{"text":"Animation and games have worked this way for a century. The character sheet comes before the scenes."}]},{"t":"h2","id":"what-goes-on-the-sheet","runs":[{"text":"What goes on the sheet"}]},{"t":"p","runs":[{"text":"Five to eight images of one person:"}]},{"t":"ul","items":[[{"text":"Front, neutral expression, even light"}],[{"text":"Three-quarter left, three-quarter right"}],[{"text":"Profile"}],[{"text":"One at a distance, full body, showing proportions and posture"}],[{"text":"Two with expression: a smile, a look away"}],[{"text":"One in the project’s actual lighting, which becomes your style anchor"}]]},{"t":"p","runs":[{"text":"Neutral is the point. Flat even light, plain background, no strong grade. You want the identity separated from everything else, because everything else will change in the final frames."}]},{"t":"h2","id":"getting-the-first-face","runs":[{"text":"Getting the first face"}]},{"t":"p","runs":[{"text":"You need one image of somebody who does not exist. Three routes:"}]},{"t":"p","runs":[{"text":"Generate and select.","b":true},{"text":" Run a contact sheet of thirty portraits with a description, pick the one you want. Free, fast, and the face is the model’s average of your words, which sometimes means it is an unremarkable face."}]},{"t":"p","runs":[{"text":"Generate and refine.","b":true},{"text":" Pick a face, then use seed variation at strength 0.1–0.2 to explore its neighbourhood until it is specific rather than generic. Worth the extra ten minutes, because a distinctive face holds better across images than an average one: there is more for the mechanism to lock onto."}]},{"t":"p","runs":[{"text":"Photograph a consenting person and stylise.","b":true},{"text":" Legitimate with permission, and the identity block later in this course sets out why permission is not optional."}]},{"t":"h2","id":"building-the-other-angles","runs":[{"text":"Building the other angles"}]},{"t":"p","runs":[{"text":"Once you have the front view, you need the rest, and they must be the same person."}]},{"t":"ol","start":1,"items":[[{"text":"Take the front view as an IP-Adapter face reference at weight 0.7 to 0.85."}],[{"text":"Add a pose control with the target angle."}],[{"text":"Generate. Check identity. Adjust adapter weight."}],[{"text":"Repeat for each angle."}]]},{"t":"p","runs":[{"text":"Expect drift. The profile is the hardest (there is least information about a profile in a front view) and it often takes eight attempts. Accept a close match rather than a perfect one; the sheet’s job is to give a "},{"text":"combined","i":true},{"text":" signal, and small variation across it is tolerable and sometimes helpful."}]},{"t":"h2","id":"using-the-sheet","runs":[{"text":"Using the sheet"}]},{"t":"p","runs":[{"text":"Two ways, and you will use both."}]},{"t":"p","runs":[{"text":"As adapter input.","b":true},{"text":" Feed two or three sheet images as face references for each new generation. Fast, needs no training, and gives a family resemblance that holds well enough for many jobs."}]},{"t":"p","runs":[{"text":"As LoRA training data.","b":true},{"text":" The sheet is a training set. Fifteen to thirty images trains a character LoRA that holds identity far more tightly. The next two lessons cover it."}]},{"t":"p","runs":[{"text":"The efficient path is to build the sheet, use adapters for the first few frames to check the character works in context, and train a LoRA only if the job needs more than an adapter gives."}]},{"t":"h2","id":"the-limit-stated-honestly","runs":[{"text":"The limit, stated honestly"}],"role":"limit"},{"t":"p","runs":[{"text":"A character sheet gets you a person who is "},{"text":"recognisably the same","i":true},{"text":" across images. It does not get you a person who is "},{"text":"identically the same","i":true},{"text":". Bone structure holds; small features drift. The nose is slightly different. The distance between the eyes moves by a pixel or two. In a set viewed one image at a time, nobody notices. In a set viewed side by side at large size, some people do."}]},{"t":"p","runs":[{"text":"The gap narrows with a well-trained LoRA and narrows further with careful repair, and it does not close. If your brief requires an identical face, a recurring character in a long campaign, a brand mascot that will be scrutinised, budget for face compositing on top of everything else: generate the body and scene, and place a consistent face in from a fixed set, matched for angle and light."}]},{"t":"p","runs":[{"text":"That is more work and it is the only method that is actually reliable. Knowing this at the brief stage, rather than at frame nine, is the difference between a plan and a crisis."}]}],"check":{"question":[{"text":"Why build a character sheet before generating any of the final frames?"}],"options":[{"text":[{"text":"Because an identity cannot be described in words precisely enough, so it has to become an image you condition on."}],"correct":true,"tempting":null},{"text":[{"text":"Because generating the sheet first establishes a consistent seed that can then be reused across every frame in the set in order to hold the face stable."}],"correct":false,"tempting":[{"text":"Seeds do not carry identity; the same seed with a different prompt or resolution produces an unrelated person."}]},{"text":[{"text":"Because a sheet built from the final frames would inherit their lighting and grade, which contaminates the identity signal used for subsequent conditioning."}],"correct":false,"tempting":[{"text":"A real reason for keeping sheet images neutral, but it explains how to build the sheet rather than why it comes first."}]},{"text":[{"text":"Because training a LoRA requires images at several angles, and those angles cannot be produced once the character has already appeared in finished work."}],"correct":false,"tempting":[{"text":"Angles can be generated at any point, and many jobs never train a LoRA at all."}]}],"order":[3,0,2,1],"why":null},"prev":{"slug":"the-same-face-twice","title":"The same face twice is the hard part","minutes":9},"next":{"slug":"training-a-lora-in-practice","title":"Training a character LoRA on hardware you can borrow","minutes":9}}}