character-generator

Generate a character reference sheet first, then lock the identity as a structured description of face, hair, clothing, and body. Each new pose is generated from the reference image plus the locked identity, never from text alone.

Provide a CharacterDefinition with appearance, clothing, art style, and a pose list. The workflow creates a reference sheet, locks identity attributes, then generates and validates one full-body image per pose such as standing, walking, running, or sitting.

Supported poses include standing, walking, running, sitting, lying, jumping, waving, pointing, looking back, arms crossed, hands in pockets, thinking, talking, laughing, angry, surprised, and scared. Custom pose descriptions are also accepted.

Yes, expressions are treated as independent assets stored in an expressions folder, such as neutral, happy, sad, angry, surprised, and scared. Expressions must not alter facial identity, hairstyle, apparent age, or body proportions.

A Consistency Reviewer compares each pose against the reference and scores identity, clothing, and pose accuracy. If the score falls below the configured threshold, the pose is marked FAIL and regenerated until it passes.

Consistency depends on the underlying image model's ability to follow reference images, so identity drift can still occur and require regeneration. Complex environments are excluded initially because characters, environments, and props are generated as separate assets.