How to Source and Create Prompts and Reference Assets for AI Personification Characters

In generative image applications, many users try to create personified versions of AI assistants or intelligent systems—often called “gijinka” or “moe anthropomorphism”—with specific outfits, hair colors, palettes, and personality traits. When searching for prompts for this kind of character, a common mistake is to look only for a complete block of text that can be copied and pasted, with the expectation that the same input will produce an identical result. In practice, the final character image is usually shaped by reference assets, the base model, sampling parameters, composition requirements, pose control, and post-processing. The prompt is only one part of the process.

Instead of searching for a so-called “standard prompt,” it is more useful to build a reusable character generation method. With such a method, you can change models, art styles, or actions while still preserving the character’s core identifying features.

Why “one-click reproducible” prompts are hard to find

A prompt is fundamentally a text description of a visual target, not a complete image file. It can describe a character’s appearance, clothing, mood, action, camera, and overall style, but it cannot precisely record every detail in a reference image.

For example, the following types of information are difficult to reproduce accurately through text alone:

• The exact structure and layering of hair accessories;
• Local patterns and decoration placement on clothing;
• The character’s exact proportions within the frame;
• Hand gestures, body orientation, and center of gravity;
• Light direction, material reflections, and background perspective;
• Line art, coloring, and composition habits specific to a derivative work.

In addition, different image generation models do not interpret the same words in exactly the same way. The same prompt can produce very different hairstyles, outfits, and facial features under different models, random seeds, or parameter settings. Prompts shared online should therefore be treated as creative references, not as “source code” with a fixed output.

A more accurate way to understand the workflow is this: the prompt states the visual requirements, reference images provide structural constraints, generation parameters influence the distribution of results, and human review and editing provide final quality control.

What to look for when sourcing reference assets

If the goal is to create an existing personification character, it is better to prioritize the following types of assets rather than searching only for a block of text.

1. Character key visuals

A character key visual can provide relatively complete information about hairstyle, clothing, color palette, and accessories, making it suitable for character design analysis. When choosing assets, try to use images that have:

• A clear main character;
• Few occlusions;
• High resolution;
• A relatively simple background;
• Complete facial and clothing details.

If the asset contains a complex background, multiple characters in the same frame, or large areas of text, the image generation model may incorrectly extract those elements and produce unstable character structure.

2. Three-view or multi-angle reference sheets

Three-view sheets usually include front, side, and back views, which help the model understand the character’s spatial structure. For tasks that require generating the same character repeatedly, a three-view sheet is more valuable than a single front-facing image because it adds information about the side profile, back hair structure, clothing back, and accessory placement.

However, a three-view sheet does not mean you should simply stitch three images together and upload them directly. A safer approach is to:

1. Confirm that the character proportions are consistent across the three views;
2. Remove irrelevant text, borders, and annotation labels;
3. Arrange the images into a clear reference board with appropriate white space;
4. Explicitly ask in the prompt for a single character, not a reproduction of multiple views from the reference sheet.

3. Pose and composition references

If the main goal is to generate a specific action—such as waving, running, sitting, or looking back—a character key visual alone is often not enough. In this case, you can provide an additional pose reference so the system understands the body skeleton, limb direction, and frame composition.

Character references and pose references should ideally serve different roles:

• Character reference: decides “who the character is”;
• Pose reference: decides “what the character is doing”;
• Composition reference: decides “how the character appears in the frame.”

Mixing these types of information together can make both the character appearance and the action unstable.

How to structure prompts in layers

A practical prompt can be divided into multiple layers instead of piling up many adjectives. The following structure is recommended.

Layer 1: Subject identity

First state what object should be generated in the image, for example:

• A single character;
• A young female figure;
• Anime-style illustration;
• Full-body or half-body framing;
• Facing the viewer or standing in profile.

The clearer the subject description, the less likely you are to get multiple people, duplicated limbs, or a missing subject.

Layer 2: Character appearance

This part describes the character’s stable identifying features, including:

• Hair color, hairstyle, and bangs;
• Eye color and eye shape;
• Clothing category and main colors;
• Structure of jackets, skirts, boots, and similar items;
• Badges, earrings, hair clips, or other accessories;
• Overall character temperament.

Prioritize the three to five most important features. Longer prompts are not always better; too many details may conflict with one another and weaken the core features.

Layer 3: Action and emotion

Specify the character’s action, pose, and facial expression, for example:

• One hand waving forward;
• Both hands holding a document;
• Body leaning slightly forward;
• A confident smile;
• Looking outside the frame.

Action descriptions should be as specific as possible. Writing only “natural pose” or “cute action” usually provides too little constraint, and the model has to guess the exact behavior.

Layer 4: Camera and composition

You can add:

• Close-up, medium shot, or full body;
• Front view, side view, or three-quarter view;
• Low angle or eye level;
• Character centered or placed to the left;
• Simple background with space reserved for text;
• Vertical or horizontal framing.

These descriptions are especially important for avatars, posters, emote sets, and promotional illustrations.

Layer 5: Visual style

Finally, describe line art, color, and lighting, for example:

• Clean line art;
• Flat color combined with soft shading;
• High saturation without overexposure;
• Soft rim light;
• Clean background;
• Consistent anime illustration texture.

Style terms should not be stacked endlessly. Conflicting styles—such as requesting heavy realism, minimalist flat color, and strong watercolor texture at the same time—often cause the output to lose consistency.

Recommended image-to-image workflow

For tasks where a character design already exists, image-to-image generation is often more stable than pure text-to-image generation. A general workflow is as follows.

Step 1: Organize reference assets

First prepare a character key visual or three-view sheet, then prepare a pose reference based on the target action. Try to ensure the reference images are clear, the subject is complete, and irrelevant backgrounds are cropped out. If multiple references are used, clearly define the purpose of each one.

Step 2: Generate a simple composition first

Do not start with a complex scene, multiple interacting characters, and a large number of decorations. Generate a single-character version with a simple background and a clear action first, and focus on checking:

• Whether the hairstyle remains consistent;
• Whether the clothing structure is correct;
• Whether the color palette is close to the design;
• Whether the number and proportions of hands and feet are normal;
• Whether the face looks natural.

Only after the character foundation is stable should you gradually add backgrounds and props.

Step 3: Control the amount of variation

In image-to-image generation, a variation strength that is too low may only make minor edits to the original image; a strength that is too high may cause the character identity to be lost. In practice, adjust gradually based on the quality of the reference image and the desired degree of change:

• Changing only the expression: use a smaller variation;
• Changing the action: use a medium variation;
• Changing clothing or scene: may require a higher variation while strengthening text constraints;
• Completely reconstructing the image: generate in stages rather than trying to finish everything at once.

Different tools may use different parameter names, so think in terms of “reference retention” and “amount of image change” rather than mechanically applying a fixed numeric value.

Step 4: Local repair

Generated results often need repair in local areas such as hands, eyes, accessories, and clothing edges. Local inpainting is more efficient than regenerating the entire image repeatedly, and it makes it easier to preserve parts that are already satisfactory.

When repairing, avoid making the selected area too large; otherwise, the model may change the hairstyle, face shape, or even clothing again. For complex accessories, first ensure the overall outline is correct, then add details in the local repair stage.

Negative prompts and quality checks

Negative prompts are used to reduce common errors, but they cannot replace positive descriptions. You can constrain the output around issues such as:

• Extra fingers or limbs;
• Deformed hands and feet;
• Duplicated characters;
• Asymmetrical facial features;
• Blurriness or low clarity;
• Text, watermarks, or logos;
• Overly complex backgrounds;
• Cropped characters;
• Abnormal numbers of accessories.

Quality checks are best performed in three rounds:

The first round checks structure: confirm that the number of characters, proportions, pose, and composition are correct. The second round checks identity: confirm that hairstyle, clothing, colors, and accessories remain consistent. The third round checks details: confirm that hands, eyes, lines, lighting, and background look natural.

How to build a reusable character prompt template

A prompt can be divided into fixed template sections and variable fields:

Subject:
Number of characters, apparent age, framing, overall style
Character design:
Hairstyle, hair color, eye color, outfit, key accessories, dominant color palette
Action and expression:
Body pose, hand gesture, gaze direction, emotion
Composition and environment:
Camera angle, position in frame, background, lighting, negative space
Quality constraints:
Clean line art, natural proportions, intact structure, stable details
Exclusions:
Extra characters, limb errors, blur, text, watermarks, severe deformation

The value of the template is that it separates “character identity” from “current scene.” Later, you only need to replace the action, background, and camera to generate different works featuring the same character.

Copyright and publishing considerations for reference assets

Character images, derivative designs, and compiled reference assets found online are not necessarily free to use. Even if an asset can be found publicly, that does not mean it can be used directly for commercial publishing, account operations, or redistribution. Before use, consider:

• Whether the original author allows derivative works;
• Whether attribution is required;
• Whether commercial use is restricted;
• Whether model training or re-uploading the asset is prohibited;
• Whether authorization is required for the work to which the character belongs.

For personal study, it is usually reasonable to keep a record of the creative process. If you plan to publish the work publicly, it is advisable to cite the reference sources in the description and avoid claiming that you created the original character design. For works involving the personification of AI systems, also avoid wording that could make people mistake the work for an officially authorized, official character, or formal product image.

Overall, the right way to find prompts for AI personification characters is not to fixate on one fixed block of text. Instead, start with high-quality character references, three-view sheets, and pose assets, then use layered prompts, image-to-image generation, and local repair to complete the work step by step. Text expresses intent, reference images preserve identity, parameters adjust variation, and human review ensures final quality. Once you understand this method, you can use common image generation tools more consistently to create personification characters with a unified style, natural actions, and a complete design.