Upload a reference sheet (front, side, 3/4 view, and face close-up) before writing any prompt, and reuse a fixed Character DNA String—a verbatim text block describing your character—every session. Visual anchors carry far more identity signal than text alone, preventing the drift that accumulates when you rephrase descriptions.
AI image models are stateless between sessions, so each generation reinterprets your text description slightly differently. Without a visual reference sheet or a locked character prompt block, small variations compound into significant drift. Providing consistent image inputs and copy-pasting the same character description verbatim eliminates most of this variation.
A Character DNA String is a fixed, verbatim text block that describes your character's core physical traits—hair color, eye color, build, skin tone, and distinguishing features. You write it once and copy-paste it unchanged into every prompt. Paraphrasing introduces variation; exact repetition signals identity consistency to the model.
Aim for 3–5 reference images: a front-facing neutral shot, a side profile, a 3/4 angle, and a tight face close-up. This set gives the AI model enough visual data to anchor identity across different scenes, lighting conditions, and styles without requiring manual correction between generations.
Combine a visual reference sheet with a verbatim Character DNA String, and upload both at the start of every new session. Store reference images in a dedicated folder. Avoid describing your character from scratch each time—even small wording changes cause measurable drift in face shape, hair color, and build across generations.
Yes, if you anchor the character with a reference sheet and a fixed prompt block before changing scene or style variables. The character's identity traits—face, hair, build—should be locked in the prompt while scene descriptors like lighting, environment, and mood are the only variables you change per generation.
CRAISEE supports image input and style reference uploads, allowing you to anchor a character visually before prompting. By uploading your reference sheet directly into CRAISEE and pairing it with a verbatim DNA String, you can achieve stable character output across dozens of consecutive generations with zero manual correction.
The most common mistakes are: describing the character from scratch each session, paraphrasing instead of copy-pasting the character prompt, skipping a visual reference sheet, and mixing identity descriptors with scene descriptors in the same prompt block. Each of these introduces variation that compounds across multiple generations.
