Character Reference
Use a face or body reference to lock identity across generations and keep consistent subjects in multi-panel work.
When to use character reference
- Storytelling projects where the same person must appear across panels.
- Avatar or portrait series with consistent age, ethnicity, wardrobe, and mood.
- Product testimonials where a single model appears in multiple scenes.
What the model derives from the reference
Primarily: face proportions, skin tone range, hair texture, and posture hints. Secondary: wardrobe color family.
Practical tips
- Crop tightly to the head and shoulders to reduce background noise.
- Use consistent lighting between refs (mixed key light directions confuse the model).
- Add hair texture language in text if you want it to survive across generations (“thick, wavy dark hair with natural part”).
- Keep wardrobe neutral in the reference so it does not bleed into scenes where it is wrong.
Multi-image consistency
For a character that must appear in very different contexts, use two separate refs:
- Ref A: face close-up
- Ref B: full-body pose with environment
This lets the model lock identity while adapting to scene.
360° and pose iterations (official pattern)
Google’s documentation shows character consistency by reusing the last render and asking for a new angle—for example, “A studio portrait of this man against white, in profile looking right.” Treat each step as image + text: attach the previous output when you need the same person from another viewpoint. Details and examples: Gemini image generation doc.
Common failure modes
- Very low resolution refs (< 512 px) lose detail and produce inconsistent faces.
- Mixing refs with very different ages or dramatically different makeup splits the face.
Related
- Cross-image character consistency guide
- Negative prompts to suppress style leakage from refs