osgaa444/seedance-characters

seedance-characters

Character fidelity, identity anchoring, and first-frame art direction for Seedance 2.0.

Scope

• Reusable character card format
• Identity anchoring via @Tag
• Multi-character separate reference pattern
• Prop/weapon separation from character body
• Hand and face safety
• 360° consistency testing
• First-frame composition rules for I2V

Out of scope

• Style and visual aesthetic — see [skill:seedance-style]
• Camera positioning — see [skill:seedance-camera]
• Fight choreography — see [skill:seedance-motion]

---

Character Card Format

Write once. Reuse across all prompts for this character. Never change nouns mid-project.

[Name]: [age range], [build], [skin tone], [hair style/color], 
[defining features], [wardrobe description], [emotional energy].

Example:
Maya: woman mid-30s, lean build, warm brown skin, short natural hair,
sharp eyes, leather jacket over white tank, calm and precise energy.

---

Identity Anchoring

For I2V and R2V, always assign the character reference explicitly:

@Image1's character as the subject
@Image1 for facial features and clothing
Use @Image1 and @Image2 for the character's appearance from multiple angles

A bare @Image1 with no role instruction is weak.

---

Multi-Character Patterns

For two characters, use separate identity anchors:

Character A references @Image1.
Character B references @Image2.

Character A throws a right punch at Character B.
Character B blocks with crossed arms.

Attribute every action by name. Never use ambiguous pronouns in multi-character prompts.

---

Prop and Weapon Separation

Upload character body and prop/weapon as separate references:

Character appearance references @Image1.
Weapon design references @Image2.

This prevents the model from blending weapon details into the character's body geometry.

---

Hand Safety

If hands are not essential to the action: frame waist-up or specify "hands not in frame."

If hands are essential: specify one simple action only.

✅  picks up the glass with right hand
✅  places hand flat on the table
✅  open palm facing camera
❌  intricate finger gestures
❌  typing on keyboard (close-up)

---

Face Stability

• Prefer medium close-up with steady, locked camera for dialogue
• Avoid rapid head turns combined with extreme close-up
• Re-upload the original face reference when extending clips
• Never rely solely on the last frame of a previous clip to maintain face identity

---

360° Consistency Test

Before committing to a character reference, generate the same character from multiple angles (front, side, three-quarter, back). Place results side by side.

If identity holds across all angles → the reference is production-ready. If identity drifts at any angle → improve the reference or generate from a better image.

---

First-Frame Art Direction (I2V)

The first frame is the primary creative contract for I2V. Everything follows from it.

Composition rules for I2V first frames

1. Bake the camera angle. If you want low angle, compose the first frame from low angle. Do not contradict it in the prompt.
2. Bake the lighting direction. The model maintains established lighting. If you want side-lighting, the first frame must show it.
3. Pose at the START of motion. If the character swings a sword, pose them at wind-up, not mid-swing.
4. Clean, depth-separated background. Cluttered backgrounds warp during camera moves.
5. Match aspect ratio. Generate the first-frame image in the same AR as the target video.

What goes in the image vs. the prompt

| In the first-frame image | In the prompt | |--------------------------|---------------| | Character identity + costume | Motion (what changes) | | Pose at start of action | Timing (when things happen) | | Camera angle + lighting | Camera movement (how frame evolves) | | Environment composition | Sound | | Color palette | Constraints |

Common first-frame mistakes

❌  Wrong lighting direction → forces re-light → causes flicker
❌  Character mid-action → no room for motion in prompt
❌  Complex cluttered background → warp during camera movement
❌  Low resolution → model loses detail for consistency

---

Real-Face Restriction

Real human face references require identity verification on the Dreamina platform. Use AI-generated portraits or illustrated/3D character references instead. See [skill:seedance-prompt] for content policy.

---

Agent gotchas

1. If identity drifts mid-clip: add "@Image1's character as the subject" and reduce motion complexity.
2. Character card nouns are contractual. Renaming "wool coat" to "heavy jacket" mid-project breaks consistency.
3. One hero subject per shot. Two max if interaction is essential.
4. When extending a clip, always re-upload the face reference image. The last frame is not enough.
5. The 360° consistency test is cheap insurance. Run it before committing to a production pipeline.