How to Build a Character Sheet for Consistent AI Video

Generate the same character twice, same prompt, same style, and you'll usually get two different people. Similar vibe, different face, different proportions, maybe different hair. That's not a prompting mistake. Diffusion models don't remember what they just made, every generation is a fresh draw. A character sheet is the fix: build the reference once, then point every future generation back at it instead of re-describing the character from scratch each time.

What a character sheet actually is

At minimum, it's a turnaround: front, side, and back views of the same character, same proportions, same outfit, same details, in one reference. More complete production sheets add a three-quarter angle, an expression sheet showing the face in a few different emotional states, and detail panels for outfit pieces, props, or materials that need to stay consistent across a whole project.

Think of it as the difference between describing a person to someone once versus handing them a photo. Every prompt after the sheet exists is pointing back at the photo.

How to decide what kind of sheet you're building

Not every character sheet serves the same job, and the pose you choose should follow from what you're actually building, not habit.

Illustration and reference sheets

For storytelling, concept art, general creative reference, or feeding a character into a video model, choose a pose with some personality in it. A stiff neutral stance reads as generic. A relaxed stance, weight shifted slightly, a hint of attitude, communicates who the character is, not just what they look like.

This is the type covered by the simple and advanced examples further down: front, side, and back at minimum, with a three-quarter angle, an expression panel, and detail panels layered in for anything longer or more complex.

Production and rigging sheets

If the sheet is heading toward 3D modeling, a game engine, or an actual character rig, personality is the wrong priority. Predictability is. This calls for a T-pose or A-pose specifically, arms straight out to the sides for a T-pose, or angled down at roughly 45 degrees for an A-pose, not a natural standing pose.

The two aren't interchangeable, and the choice genuinely matters. T-pose is the older industry default, easier to align mathematically since the arms sit on clean axes, and most auto-rigging tools expect it. A-pose has become more common in modern games because it deforms more naturally at the shoulder, a T-pose's fully extended arm tends to compress and distort once it's rigged and animated. If you're not sure which your pipeline expects, A-pose is the safer modern default.

Either way, keep the lighting flat and the background neutral. Dramatic shadows might look better in isolation, but they obscure the actual silhouette a modeler needs to see, and silhouette clarity is the entire point of this type of sheet.

Getting the layout right

A multi-view turnaround belongs on a landscape canvas, laid out horizontally so every angle sits at the same scale, side by side, with even spacing between them. A single hero portrait, one character, one pose, no rotation, works better on a portrait canvas instead, closer to 2:3.

If you're planning to actually use the sheet as a reference later, plan for it to get split apart. Most workflows generate the full turnaround as one image, then break it into individual panel images before referencing a specific view, tagging just the front view as @image1 for a face-focused generation, say, rather than the whole grid. Building the sheet with clean, even separation between panels from the start makes that split painless instead of a cropping headache later.

Building a simple sheet

Describe the character in one detailed prompt, or upload a single reference image if you already have one, a sketch, a portrait, an existing character design, and ask for a three-view expansion from it. Either path works. The more specific the description or the cleaner the reference image, the more the multi-view output actually holds together.

A simple sheet covers front, side, and back. That's enough for most short-form work, social content, a single video, a one-off illustration.

If you're generating the views separately rather than all at once, build front first, it's the foundation every other angle checks against. Back can mirror front's proportions directly. Three-quarter is consistently the hardest of the four to get right, it involves real foreshortening rather than a flat profile, so leave it for last and expect an extra pass or two to land it.

Building an advanced sheet

For anything longer, a multi-episode series, a recurring brand character, a game asset, add layers on top of the basic turnaround: a three-quarter angle for more natural framing options, an expression sheet with three or four emotional states, and dedicated detail panels for anything that needs to read correctly up close, a logo on a jacket, a specific prop, a distinctive accessory.

Build these in stages rather than one giant prompt. Generate the turnaround first, confirm it holds up, then generate the expression sheet and detail panels as separate passes referencing the approved turnaround. Trying to get everything in one shot is where advanced sheets tend to fall apart.

How many views do you actually need

Three to five is the working range for most illustration, concept, and reference work: front, side, back, and optionally a three-quarter angle for depth. That's enough to keep a character on-model for most creative projects, ours included.

Production pipelines sometimes need more. A full eight-angle compass sheet, front, two front three-quarters, two profiles, two back three-quarters, and back, is standard when a character needs to read correctly from every direction a camera or game engine might approach it from, common in detailed 3D reference or 8-directional sprite work.

If you're not sure which you need, build the five and add only the angles a later step actually demands. More views mean more consistency work to maintain across the sheet, not automatically a better reference.

The technique that actually holds consistency

Generate the reference sheet first, before anything else in the project, and save it. Every prompt after that references the sheet directly plus a short one-line identity string, something like "woman, late 20s, short dark hair, oversized cream sweater, navy tumbler in left hand." That short string matters more than it seems like it should, it's a compact anchor the model can hold onto even when the rest of the prompt changes scene, angle, or lighting.

Worth knowing going in: there's no universal method that solves character consistency completely. Turnaround sheets are the most reliable starting point, but different techniques have different failure points, and it's worth testing your specific character on a few throwaway generations before committing a full project to any one approach.

Common mistakes worth avoiding

Locking the pose before locking the design. Settle the silhouette, proportions, and defining details on a single front view first. Every other angle should build from that one approved view, not the other way around, changing something after the turnaround exists means redoing every angle, not just one.

Mixing pose types on one sheet. A production sheet in A-pose next to an illustration-style angle in a relaxed stance doesn't read as one consistent reference, it reads as two different characters that happen to share a face. Pick one pose type and hold it across every view on that sheet.

Dramatic lighting on a production sheet. It looks better in isolation, and it actively works against the sheet's job. A modeler needs to see the silhouette clearly, and hard shadows hide exactly the edges and forms they're trying to read.

What you can actually build from it

This is the part that makes the sheet worth building in the first place. Once you have a locked reference:

Feed it into a video model as the character reference. Seedance, Kling, and Nano Banana 2 all accept a tagged reference image, @image1 in Seedance and Nano Banana's syntax, that locks the character's face and identity across a generation. A character sheet is the strongest possible input for that tag, stronger than a single portrait, because it's already proven to hold up across angles.

Build a storyboard with the same character across every panel. Our storyboard guide covers this directly, generate the sheet first, then reference it in every panel prompt that follows.

Keep one mascot or spokesperson consistent across a whole campaign. A brand character that needs to appear in a dozen pieces of content over months holds up far better against a saved sheet than against a fresh description each time someone on the team generates something new.

Feed it into a rigging or Live2D pipeline. A production-style turnaround, correctly posed and flatly lit, is standard input for both 3D rigging tools and Live2D setups used for VTuber-style avatars, where the rigger needs to see exactly how hair, face, and clothing behave as the character rotates.

Five worked examples

Contemporary, simple turnaround

Front, side, and back view character turnaround, same character in all three views. Subject: a man in his early 30s, short beard, olive green field jacket, dark jeans, worn leather boots. Consistent proportions and outfit details across all three angles. Neutral studio background, even lighting, no shadows, flat reference-sheet style.

Contemporary, advanced production sheet

Building on that same character:

Character reference sheet for the man in the referenced turnaround, defined by @image1. Include: three-quarter angle view, an expression panel with neutral, laughing, and concerned expressions, and a detail panel showing the jacket's collar and buttons up close. Maintain identical proportions, outfit, and facial features from the reference. Clean grid layout, labeled panels, neutral background throughout.

Video game character, production sheet

This one's a production sheet, not an illustration reference, so the pose and lighting rules from earlier apply: A-pose, flat lighting, neutral background.

Front, side, and back view character turnaround, same character in all three views, in A-pose with arms angled down at roughly 45 degrees. Subject: a fantasy game character, an elven ranger, weathered leather armor, a longbow slung across the back, short auburn hair, a scar through one eyebrow. Consistent proportions and gear details across all three angles. Neutral grey studio background, flat even lighting with no dramatic shadow, game-art rendering, clean silhouette.

For the advanced version, add a three-quarter angle, an expression panel (neutral, focused, alarmed), and a detail panel on the bow and quiver, keeping the same A-pose and flat lighting throughout.

Brand mascot

Front, side, and back view character turnaround, same character in all three views. Subject: a friendly rounded mascot character, a small orange fox wearing a tiny backpack, big expressive eyes, simple rounded shapes. Consistent proportions, color, and details across all three angles. Neutral white background, flat vector-style illustration, even lighting, clean bold outlines.

Expression sheet

A standalone expression grid, useful once a character's base design is locked, referencing that design directly rather than describing the face again from scratch:

An expression reference sheet for the character in @image1, arranged as a clean 3-by-3 grid of head-and-shoulder portraits. Expressions, left to right, top to bottom: neutral, joy, concern, anger, surprise, suspicion, sadness, determination, relief. Fixed eye-level camera angle and identical framing across all nine panels, flat even lighting, light grey background throughout.