How to Make a Character Reference Sheet for AI Image Generation

Character consistency is easier when the generator is given a designed identity, not just a name and a paragraph. A reference sheet—often called a model sheet— turns the important visual decisions into one controlled source: proportions, face shape, hair, clothing, colors, and the details that should survive a new pose or setting.

The goal is not to freeze the character into one image. It is to separate identity from performance. Identity should remain recognizable; pose, expression, camera, and environment should change. This guide explains how to build a compact reference sheet and use it as the highest-priority source in an AI-assisted illustration workflow.

Define identity anchors before generating views

Choose a small set of features that make the character identifiable. Useful anchors are specific and observable:

• overall proportions and apparent age;
• head and jaw shape;
• eye spacing, brow shape, and nose profile;
• hair silhouette, part, texture, and length;
• skin tone under neutral light;
• one or two distinctive marks, such as a freckle cluster or eyebrow notch;
• signature garment shapes and their color relationships;
• recurring accessories and which side they appear on.

Avoid making every detail an anchor. A character should be able to change shoes, remove a jacket, or appear in different light without becoming a new person. Prioritize features that affect recognition at thumbnail size.

Write the anchors in neutral terms. “Attractive young hero” leaves the model to invent a familiar stereotype. “Long triangular face, low straight brows, close-cropped curls with a high left temple, and a small gap between the front teeth” describes an identity that can be checked.

Choose a neutral base pose and light

The sheet should reveal form rather than sell a dramatic scene. Use a relaxed standing pose, arms slightly separated from the torso, hands visible, and feet unobstructed. A plain background and broad neutral light make colors and proportions easier to compare.

Strong colored light may create an appealing portrait but a poor reference. It can hide eye color, shift skin and fabric values, or merge the hair silhouette into the background. Save expressive lighting for later scene tests.

Use a camera around eye level with restrained perspective. Extreme wide-angle views distort head and limb proportions, while a long-lens portrait may conceal body relationships. The base reference should feel descriptive.

Build the core view set

A practical sheet does not need dozens of drawings. Start with:

1. Front view: neutral expression, full body.
2. Three-quarter view: the view most scenes will use.
3. Side view: profile of face, hair, chest, and footwear.
4. Back view: hair shape, garment construction, and rear accessories.
5. Face close-up: neutral light, high enough resolution to inspect anchors.

Keep scale aligned across full-body views. If one view is much larger or posed differently, comparison becomes harder. Label views outside the generated art in a layout tool; do not depend on the model to typeset the sheet.

If the final work is mostly waist-up, the full-body views still prevent costume details from being invented later. If the character is a creature or robot, add the angle that reveals its key construction—top view for a shell, underside for a flying object, or an exploded accessory detail.

Add an expression row

Identity can drift when expression changes. Create a small row using the same camera and light:

• neutral;
• genuine smile;
• concern or concentration;
• anger without exaggerated transformation;
• surprise.

Compare bone structure, eye spacing, hairline, and distinctive marks across the row. The mouth and brows should move; the underlying face should not. If every emotion produces a different nose or jaw, strengthen the identity reference before attempting complex scenes.

Include only expressions the project needs. A calm educational mascot may not need horror or slapstick extremes. The sheet is a production tool, not an acting catalogue.

Document wardrobe as rules

Clothing is one of the fastest sources of inconsistency. Record:

• garment layers and order;
• silhouette and fit;
• exact color roles;
• closures, pockets, seams, and trim;
• sleeve and trouser length;
• footwear;
• accessories and left/right placement.

Use color swatches sampled from the approved view. Store values in a normal design file, because a screenshot is affected by display and export. If an exact brand color matters, apply it during compositing rather than assuming a generator will reproduce a numeric value.

Separate permanent identity from outfit variants. The primary sheet may show the canonical outfit; secondary cards can define winter, formal, or work versions. Each variant should still preserve face, body, hair, and recurring identity anchors.

Generate the sheet in controlled passes

Trying to create the entire model sheet in one prompt can produce internally inconsistent panels. Build it in passes:

1. approve one neutral three-quarter design;
2. derive front and side views using that image as the identity reference;
3. create back and detail views;
4. create the expression row;
5. assemble approved images in a layout tool.

Repeat the same identity block in every pass. The reference-image brand workflow explains how to give each reference one job and prevent a mood image from overriding identity.

When a new view fails, change one variable. Do not simultaneously rewrite the character, pose, clothing, and rendering style. The prompt structure guide offers a useful layer order for targeted revisions.

Use a consistency rubric

Score candidate views against the approved three-quarter image:

Area	Check
silhouette	head, hair, body, and costume read as the same design
face geometry	jaw, eye spacing, brow, nose, and ears remain stable
distinctive marks	position, side, shape, and count are correct
palette	color roles stay consistent under neutral light
wardrobe construction	layers, closures, pockets, and accessory side match
proportions	relative head, torso, arm, and leg lengths remain stable

Mark a candidate “concept only” if it looks interesting but fails identity. Do not quietly make it the next reference. Reference drift compounds: a changed view generates a more changed view, and after several scenes the original character is gone.

For important projects, overlay face landmarks or silhouettes at partial opacity. Visual comparison catches changes that a verbal review misses.

Move from sheet to scene

When generating a scene, keep the reference relationship explicit:

Preserve the character identity, proportions, facial geometry, hairstyle, left-ear silver cuff, and canonical jacket from the supplied model sheet. Change only the pose, expression, camera, and environment described below.

Then write the scene as a separate block. If the character interacts with a prop, describe contact points and which hand is used. If several characters appear, give each a distinct identity reference and spatial position.

Start with simple scenes before crowds, action, or extreme perspective. A seated conversation is a better consistency test than a motion-blurred fight scene. Once the identity survives a few controlled settings, expand the range.

Keep exact assets deterministic

A generated model sheet does not make every downstream use safe. Licensed characters, client identities, and real people require appropriate rights and consent. Do not create deceptive depictions of real people. Keep official logos, uniform badges, and text as separate approved assets.

Archive the canonical sheet, source prompts, reference IDs, generation date, and manual edits. Number revisions. If a character intentionally changes, document the change instead of overwriting the old reference.

You can develop neutral view candidates with BananaLite, then assemble and label the approved sheet in your normal design tool. The assembled sheet—not a random favorite output—becomes the source for later scenes.

Sources and further reading

• Library of Congress still-image format guidance — useful for choosing durable archival masters.
• W3C Design Tokens format — a model for storing colors and design decisions as named reusable values.
• U.S. Copyright Office AI initiative — current primary materials on copyright questions involving AI-generated works.

A good reference sheet is a contract with the future. It says which properties make the character the same person, which properties may change, and how an editor can tell the difference.