An NSFW AI character reference sheet is a single artifact combining a turnaround (front, side, back), an expression set, and an outfit/detail callout, generated in Stable Diffusion with sheet-style prompts and OpenPose control. You then reuse it as an IPAdapter reference, a LoRA dataset seed, and a prompt-token checklist to keep your original adult character on-model.
Every serious AI character starts with a reference sheet. It is the master document that turns a vague idea into a repeatable design. Without one, you regenerate the character from memory every session and the look drifts. With one, you have a fixed visual contract: this is her face, these are her three angles, this is the outfit, these are the tokens. This guide covers how to build each section of the sheet and, just as importantly, how to feed it back into your generation pipeline downstream.
Why a reference sheet is worth the effort
A reference sheet does three jobs at once. First, it forces you to lock design decisions instead of improvising them, so your character has a stable identity. Second, it becomes literal input for consistency tools: the turnaround images feed IPAdapter or a LoRA dataset, and the callout sheet becomes a text checklist you paste into every future prompt. Third, it is a communication artifact if you ever collaborate, since one image tells another creator exactly who the character is.
The underlying methods for holding a face steady across images are covered in depth elsewhere. This post stays at the artifact level: what the sheet contains and how to produce it. When you need the deep mechanics of identity locking, use the consistency techniques pillar. When you are ready to convert the sheet into a trained model, follow how to train a nsfw character LoRA.
| Sheet section | Purpose | Key prompt tokens |
|---|---|---|
| Turnaround | Lock the character in 3D from every angle | character sheet, multiple views, front view, side view, back view, reference sheet, white background |
| Expression set | Define the emotional range for reuse | same face, neutral, happy smile, blush, angry, expression sheet |
| Outfit and detail callout | Record clothing, accessories, colors | full outfit reference, detail closeups, color palette, accessories, labeled callouts |
| Token checklist | A pasteable text recipe for future prompts | the exact hair, eyes, body, outfit descriptors as words |
Section 1: the character turnaround
The turnaround is the backbone. It shows the character from the front, the side, and the back in a consistent pose and neutral lighting, usually against a plain white background so the design reads clearly. This is what an artist or a training pipeline uses to understand the character in three dimensions.
A strong turnaround prompt:
(original female character, adult), character sheet, character turnaround,
multiple views of the same character, front view, side view, back view,
reference sheet, full body, T-pose, neutral expression,
white background, even flat lighting, consistent design, high detail,
[her hair, eyes, body type, and signature outfit recipe]
Negative: different characters, inconsistent design, busy background, extra limbsText-to-image alone often gives you three slightly different people rather than three views of one. This is where pose control matters. Feed an OpenPose skeleton showing three figures in front, side, and back stances so the model places one character in three defined poses. The OpenPose pose control guide shows how to build that multi-figure pose input, and stacking pose control with other conditioning gives even tighter results.
Section 2: the expression set
The expression sheet fixes your character’s emotional vocabulary. Generate the same face at the same seed across a row of moods: neutral, happy, blush, angry, and any signature look your character needs (a smirk, teary eyes). The goal is that every face is unmistakably the same person, only the emotion changing.
Hold the seed and the full character recipe constant, then vary only the mood token. Precise phrasing matters here, so lean on a tested library of facial expression prompts rather than guessing words. If the face shifts identity between expressions, generate them as a grid from one seed or inpaint only the mouth and eyes so the underlying face stays fixed. A clean expression set is what lets you drop the character into any scene later without her feeling like a different woman each time.
Section 3: the outfit and detail callout
The callout sheet documents the specifics that a full-body shot glosses over: the exact cut of the outfit, accessories, tattoos or markings, hair clips, footwear, and the color palette. Generate closeups of each element and a clear full-outfit shot. The point is to pin down details you will otherwise render inconsistently, like which shoulder the strap sits on or the precise color of her eyes.
Record the palette explicitly. Note the hex-like descriptions in words (warm auburn hair, teal eyes, charcoal jacket) because those exact phrases go into your token checklist. If your character’s design leans on a specific art style, the art style prompts library helps you phrase the aesthetic so it holds across every render.
How to use the sheet as an IPAdapter reference
Once the turnaround exists, its front-view portrait becomes a reference image for IPAdapter. You pass that clean face into IPAdapter and it steers new generations toward that identity, letting you place the character in new poses and scenes while keeping her recognizable. This is the fastest way to reuse a design without training anything.
Use the front view or a strong three-quarter face as the reference because those carry the most identity information. Crop tightly to the face for maximum identity transfer, or use a fuller shot when you want the body and outfit to carry through too. The mechanics of dialing IPAdapter weight and choosing the right reference crop live in the consistency pillar; the sheet simply gives you the clean source image that makes IPAdapter work well.
How to use the sheet as a LoRA dataset seed
A reference sheet is the ideal starting point for a training dataset. The turnaround gives you front, side, and back; the expression set gives you facial variety; the callout gives you outfit detail. Together they seed a diverse, consistent dataset that teaches a LoRA the character from every angle.
A sheet alone is not a full dataset, though. You typically want fifteen to thirty varied images, so expand from the sheet by generating the character (using IPAdapter from your reference) in different poses, lighting, and framings, then curating the best. The LoRA training dataset guide covers how many images, what variety, and how to caption them. Deciding whether a LoRA, DreamBooth, or textual inversion fits your case is covered in LoRA vs DreamBooth vs textual inversion.
How to use the sheet as a prompt-token checklist
The least technical but most-used output of a reference sheet is a plain text checklist. Translate every locked design decision into exact words: warm auburn shoulder-length wavy hair, teal eyes, athletic build, adult, freckles, charcoal cropped jacket, silver hoop earrings. This block is your canonical character token. You paste it into every prompt so even without IPAdapter or a LoRA, the model gets a precise description.
Keep the checklist versioned in a text file next to your images. When you tweak the design, update the checklist so your prompt and your reference sheet never disagree. This text recipe is also the anchor that a full persona system builds on for a recurring character.
A practical build order
Work in this sequence. First, settle the design in your head or a rough prompt. Second, generate the turnaround with OpenPose so you have the three canonical angles. Third, produce the expression set from a locked seed. Fourth, shoot the outfit and detail callouts. Fifth, write the token checklist from what you see. Sixth, pick your reuse path: IPAdapter for quick reuse, LoRA for heavy reuse, or the checklist alone for lightweight work.
Built in that order, the sheet compounds in value. Each section feeds the next, and the finished artifact serves you for the entire life of the character, whether you make one image a month or a full series. Archive the sheet, the seed, and the checklist together so months later you can regenerate a perfectly matching new pose without rediscovering the design from scratch.
Generating the sheet as one image versus a set of images
You can build a reference sheet two ways, and each has tradeoffs. The single-image approach prompts the model to lay out multiple views in one canvas, using tokens like character sheet, multiple views, front view, side view, back view on a wide canvas. It looks tidy and reads like a real model sheet, but the model often makes the three views subtly different people, and you have less control over each view’s quality. The multi-image approach generates the front, side, and back as separate renders, each with its own OpenPose skeleton, then arranges them side by side in an image editor. This takes longer but gives you a clean, controllable, genuinely consistent turnaround.
For serious work, prefer the multi-image approach. Generate each view individually with the same seed, the same character recipe, and a pose-specific OpenPose input, then composite. You get full-resolution control of every angle, and if the side view fails you regenerate only that one panel rather than the whole sheet. The single-image method is fine for a quick concept pass, but the composite method is what produces a sheet you can actually train on.
Common mistakes that ruin a reference sheet
The most common failure is skipping the seed lock, which lets the face drift between panels so your turnaround shows three cousins rather than one woman. Always fix the seed and the full recipe across every view. The second mistake is a busy background: patterns and props behind the character get baked into IPAdapter references and LoRA training as if they were part of her, so keep the backdrop plain white or flat grey. The third is inconsistent lighting across panels, which reads as different times of day and confuses both the eye and the training pipeline; use even, flat, neutral lighting on the sheet so the design itself is what stands out.
A fourth, subtler mistake is designing the sheet around a single flattering angle. If every panel is a three-quarter front view, you never actually document the profile or the back, and the first time you need the character from behind she is improvised. A true turnaround earns its name by showing genuinely different angles, even the unflattering ones, because those are exactly the views you cannot easily improvise later.
Versioning and archiving the sheet
A reference sheet is a living document early in a character’s life and a fixed contract later. While you are still designing, expect to revise: you tweak the hair, adjust the outfit, settle the eye color. Each time you change the design, regenerate the affected panels and update the token checklist in the same edit, so the images and the words never disagree. Keep old versions in a dated archive folder rather than overwriting, in case you decide a earlier choice was better.
Once the design is locked, freeze the sheet and stop editing it. From that point it is the reference every future image answers to. Store the finished sheet, the seed, the exact prompt recipe, the negative prompt, the base model, and any LoRA together in one folder. That bundle is what lets you, or a collaborator, generate a perfectly matching new image of the character a year later. Without it, a character you loved becomes unreproducible the moment you close the session, which is the single most common way solo creators lose their best original designs. The sheet plus its recipe is cheap insurance against that loss, and it is the foundation every recurring-character system is built on.
From sheet to first real scene
The payoff of all this documentation is speed on every future image. With the sheet finished, generating a new picture of the character becomes a short, repeatable routine. Paste the token checklist into your positive prompt, load the LoRA or attach the front-view IPAdapter reference, set your scene and pose, and render. Because the identity is pinned three ways at once, by the checklist words, by the trained or referenced model, and by the design you have internalized from staring at the sheet, the face lands on-model far more often than it would from a cold prompt.
Compare that to working without a sheet: you half-remember the hair, guess the eye color, forget the earrings, and end up with a character who is close but clearly not the same woman as last week. The sheet eliminates that drift by making every design decision explicit and reusable. It is the difference between a character who exists only in your head and one who exists as a stable, shareable, reproducible asset. That stability is what lets a single original character carry a whole series of images, a visual novel, or a recurring content stream without ever losing her identity along the way.
How much detail is enough
New creators either under-document or over-document. Under-documenting leaves the design vague, so the model fills gaps differently each time and the character drifts. Over-documenting crams so many tokens into the recipe that they start fighting each other and the model ignores half of them. Aim for a middle: capture the load-bearing traits that define the character (hair, eyes, body type, one or two signature features, the core outfit) in clear tokens, and let the model handle incidental variation like exact skin pore detail or micro-folds in fabric. A reference sheet with six to ten strong, distinctive descriptors is far more reproducible than one with forty competing ones, and it leaves room for natural variety across poses without losing identity.
Frequently asked questions
What exactly goes on a character reference sheet?
A complete sheet has four parts. A turnaround shows the character front, side, and back in a neutral pose on a white background. An expression set shows the same face across moods like neutral, happy, blush, and angry. A callout sheet documents the outfit, accessories, and color palette in closeups. Finally, a written token checklist translates every design detail into exact prompt words you paste into future generations.
How do I get a consistent turnaround instead of three different people?
Text-to-image alone tends to produce three slightly different characters. Use OpenPose ControlNet with a skeleton showing three figures in front, side, and back stances, so the model renders one character in three defined poses. Combine that with a strong character-sheet prompt, a locked seed, and a plain white background. This forces the same design across all three angles rather than letting the model improvise each view.
Can I use my reference sheet directly to train a LoRA?
The sheet is an excellent seed but usually not a full dataset by itself. Training wants roughly fifteen to thirty varied images. Use the sheet’s front view as an IPAdapter reference to generate the character in more poses, lighting, and framings, then curate the strongest results into your dataset. The turnaround and expression set give you the angle and mood variety that makes the trained LoRA generalize well.
Which image from the sheet works best as an IPAdapter reference?
Use the front-view portrait or a strong three-quarter face, since those carry the most identity information. Crop tightly to the face when you want maximum identity transfer to new poses, or use a fuller shot when you also want the body and outfit to carry through. A clean, well-lit face on a plain background gives IPAdapter the clearest signal and the most recognizable results.
Why bother with a text token checklist if I have a LoRA?
Even with a LoRA, the checklist is your canonical description and a fallback. It keeps prompts precise, works on any model without loading your LoRA, and documents design decisions so your prompt and reference images never drift apart. It is also the easiest artifact to share or version. Think of it as the written source of truth that every visual output should match.
How many expressions should the expression set include?
At minimum include neutral, happy, blush, and angry, since those four cover most storytelling needs. Add any signature looks your character requires, such as a smirk, teary eyes, or a sultry glance. Generate them all from one locked seed so the underlying face stays identical and only the emotion changes. Four to six well-matched expressions give you plenty of range without ballooning the sheet.
Should the reference sheet use a white or transparent background?
Use a plain white or flat neutral background for the sheet itself, because it makes the design read clearly and helps training pipelines focus on the character. Transparent backgrounds matter later when you cut sprites for a scene, but for the reference document, a clean solid backdrop is standard. It also keeps IPAdapter and LoRA training from picking up unwanted background features as part of the character.
How do I keep my character original and not copy someone real?
Design from attributes, not references to real people. Choose an age-adult body type, face features, hair, and outfit as abstract descriptors rather than naming a celebrity or basing it on a photo. Never use real-person likeness, no matter how loosely. An original fictional adult character built from a written recipe is both safer and more reusable, since you fully control and document every trait on the sheet.
Related guides:
