How to Make a Character Reference Sheet With AI

Dreamina is a useful starting point for creating and refining those images. Work from an approved character reference rather than asking the model to invent a new character in every panel. Our existing Dreamina example also shows why a requested “front view” needs inspection: the resulting panel was still a three-quarter view.

The goal is a reference another image or video workflow can use. You are defining the character's design, including decisions about surfaces the original picture may not reveal.

Step 1: Decide what the sheet needs to explain

Start with the next job. A front-facing mascot post may need one full-body view and a few expressions. A character who turns around in a film needs approved side and back information. A close-up may need facial details more than an extensive pose library.

Choose a small, useful set:

You do not have to generate every possible angle. Include a view because it answers a design question or supports a planned shot.

Write a short design note before generating. Record the character's defining traits, relative proportions, materials and anything that must not change. Separate these fixed decisions from pose, lighting and expression, which may vary later.

For more general image-tool selection, see AI image tools for character consistency. This tutorial focuses on making the reference package itself.

Step 2: Approve one master image

Choose the clearest existing image of the design you want. If you are starting from text, create candidate images first and choose one before asking for the sheet.

A useful master shows the face and costume clearly, with enough of the body to establish proportions. Avoid treating a heavily obscured action pose as the only authority for a full-body turnaround. If the lower body is missing, its design still needs to be decided.

Our example uses this courier robot:

Its fixed traits include a round amber eye on the viewer's left, a square blue eye on the viewer's right, a red-bead antenna, dark-blue hands and a mustard backpack. Those details give you specific things to compare across images.

Save the approved master separately. Do not overwrite it with an attractive variation that quietly changes the eyes or backpack. Add a brief note explaining which file is the current design authority.

The Dreamina AI photo editor can help refine a source image. Correct a distracting background or unsuitable composition without casually accepting a redesigned face. A cleaner starting image is valuable only if it retains the design you intend to use.

Step 3: Generate a simple first sheet

In Dreamina, open the image-generation workspace, select a model that supports your reference workflow and attach the master image. Write the views, framing, background and fixed traits explicitly. Confirm the current aspect ratio, output setting and credit quote before generating.

For the reference-sheet example, Seedream 5.0 Pro used one character image with a 16:9 layout and 2K output. Keeping the character reference fixed lets you concentrate the prompt on the views and details you need for the sheet.

The instruction was:

Create a clean production character sheet of this exact courier robot: front view, left side view and back view, full body at the same scale and evenly spaced on a warm-gray background. Preserve asymmetric eyes, red-bead antenna, dark-blue hands and mustard backpack. The back view must show a plausible backpack attachment. Consistent handcrafted clay-and-enamel materials, no labels or text.

Here is the native image returned by that attempt:

This 2560 × 1440 image is a draft. Its first panel is not a straight-on front view, and the backpack attachment visible from behind introduces design information that needs approval.

A combined sheet is a convenient first attempt because you can compare several views together. It can also produce different proportions or ignore a requested angle. Do not let an orderly layout hide those problems.

Free access and paid upgrade: Free accounts in the US receive 120 credits per day; daily allowances vary in other regions. Start developing images and short videos with Seedance 2.0 for video and Seedream 5.0 Lite for images. For regular production, Dreamina Basic is USD 15 per month on monthly billing, with 1,575 monthly credits. Paid plans add watermark-free downloads and a faster generation queue. Credit use varies by model and settings.

Step 4: Correct the view instead of regenerating the whole design

Inspect each panel against the master. For a front view, the character should face the camera directly; a view that reveals substantial side depth may still be three-quarter. For a profile, check which side is shown and whether accessories occupy plausible positions.

When one view fails, ask for that view separately. Keep the master attached and describe the specific correction. If your sheet has the same problem, an illustrative follow-up prompt is:

Create a straight-on front view of the referenced courier robot, facing directly toward the camera with the head and torso aligned forward. Show the complete body at a neutral standing pose. Preserve the original asymmetric eyes, antenna, dark-blue hands, backpack design and materials. Plain warm-gray background. One character only, no labels.

Compare the result with the master again. A technically frontal image still fails as a reference if the model swaps the eyes, changes the head shape or adds a different costume.

Keep each approved view as its own file. If an iteration drifts, return to the master and the relevant approved view rather than using the most recent flawed generation as the new starting point.

Kapwing's reference-sheet guide also develops the task through stages. The useful principle is to resolve the design before making the reference package larger; adding more panels cannot repair an inconsistent foundation.

Step 5: Make deliberate decisions about unseen details

A back view generated from a front or three-quarter image is partly a design proposal. The model cannot recover a hidden backpack attachment that was never supplied.

In the courier draft, plausible rear straps were introduced. Decide whether that attachment suits the character, whether it connects sensibly to the visible backpack and whether it should become part of the approved design. If it should not, describe the intended alternative and revise the rear view.

Record the decision in a short note. For example, you might specify that the rear attachment shown in the approved back image is part of the design. That gives later work a clear authority instead of asking every new generation to infer a different solution.

This distinction matters for costumes, hair, footwear, props and mechanical joints as well. An AI-generated unseen angle is useful concept work, but it becomes a reference only after you choose it.

Step 6: Add expressions and details without changing identity

Once the main views are stable, add only the expressions, hands or accessory details the project needs. Generate these from the master and a relevant approved view.

Keep expression separate from redesign. A smile can change the mouth and cheeks; it should not unintentionally change the eye shapes, antenna or materials. For a character with deliberate asymmetry, inspect whether that asymmetry is preserved or accidentally mirrored.

Use a small detail panel when an important feature would otherwise be too tiny on the sheet. A hand close-up can explain finger construction; a backpack detail can explain an attachment. These images should agree with the full-body views.

If several references are available, assign each a clear purpose. Do not include rejected alternatives merely to fill the input allowance. Our guide to combining multiple reference images explains how to keep character, setting and style roles understandable.

Step 7: Arrange, label and export the package

Bring the approved images into a layout editor such as Canva. Use a blank landscape page or another format suited to the recipient, then place the full-body views at a comparable visual scale. Align their feet where possible and leave enough space to distinguish silhouettes.

Add labels as ordinary text layers. Label the actual angle shown, the character name and the version or date if useful. Do not ask the image model to reproduce small technical labels when editable text is easier to correct.

Include a compact fixed-traits note beside the images. You can also sample colors from the approved master in an editor, but avoid inventing exact color codes from memory. For materials that change with lighting, a short description can be more informative than a color swatch alone.

Export a clear PNG or JPEG for visual use, or a PDF if your layout tool and recipient support it. Check the exported file at a normal viewing size: faces, hands and important notes should remain readable. Keep the editable layout and original approved images in the same organized package.

A useful handoff contains the finished sheet, separate approved views, the master image and the short design note. The sheet gives a human a quick overview; the individual images let a later generator use the relevant view without interpreting an entire collage.

Use the right reference for the next scene

Upload the image that explains the shot you are making. A profile walking shot may benefit from the approved side view; a frontal close-up may need the master and facial detail. The correct choice depends on the selected model and input role.

For example, Runway's Gen-4 Image References workflow supports named references and up to three images for that image workflow. That allowance is not a universal limit for every Runway video model or every platform.

A reference sheet helps you make consistent decisions; it does not lock every future output automatically. Keep the master available, use only relevant approved assets and correct drift before expanding the scene. When you move into animation, continue with keeping the same character across AI video scenes to connect the design package to shot-by-shot continuity.