character-pipeline

Character Pipeline

This skill is stills-only. It builds reusable identity assets so the same person holds across plates and composites. If the user wants motion after that, hand off the approved scene still + sheets to video-prompting / cinematic-motion — do not own generate_video here.

Pipeline (hard order)

1. Hero still          generate_image (or import_media / user photo)
2. Multi-angle sheet   generate_image with referenceMediaRefs = [hero]
                       bust sheet default; optional full-body turnaround
3. Environment plate   generate_image — empty set, locked camera/light
4. Scene placement     generate_image: env plate + sheet (+ optional wardrobe)
5. Fidelity close-up   optional generate_image or capture_frame from the scene still

Do not skip to scene placement without an approved sheet. Reroll stills until lighting and identity pass.

Session gates

• get_timeline → canGenerate
• list_models({ type: "image" })
• Propose each paid generation; wait for confirmation
• Folders e.g. Character/Hero, Character/Sheets, Character/Scenes

Model picks

Confirm via list_models:

Step	Prefer
Hero / sheet / env / scene	grok-imagine, gpt-image-2, or seedream-v5-pro (photoreal trio)

Prefer gpt-image-2 / seedream-v5-pro for any step that needs referenceMediaRefs if the current Grok build rejects refs. Trust each model's aspect/resolution enums from list_models.

Step 1 — Hero still

Clear identity plate. Prefer candid editorial over beauty-retouch.

Include: age/skin/hair/eyes/expression; simple stable wardrobe; repeatable light; skin texture preserved (pores — no smoothing).

Avoid: unrepeatable glam grade; extreme pose that fights the sheet; busy BG that leaks into identity.

If the user supplies a photo, import_media and skip generation.

Step 2 — Character sheet

generate_image with referenceMediaRefs: [heroMediaRef].

Head / bust sheet (default — identity lock)

• 4 views in one continuous image, left → right: front, left profile, right profile, back of head
• No white gutters, borders, frames, or panel dividers — seamless shared grey BG
• Neutral light grey BG; even studio light; no color cast
• Same crop, height, distance every view (chest-up / bust)
• Identical simple wardrobe; neutral expression unless asked otherwise
• Exact facial structure from hero
• Detail emphasis: razor-sharp focus, pores/microtexture, hair strands, fabric weave — no soft blur / beauty mush
• Clinical identity sheet — not fashion/editorial
• No vignette, no grade, no text watermark

For GPT Image 2: quality: "high" and prefer 3840x2160 (or at least 2560x1440) when detail matters.

Full-body turnaround (optional — wardrobe / silhouette)

When scene placement needs legs, stance, or clothing behavior:

• 4 full-body views in one continuous image — same no-gutter rules
• Shared grey floor+BG; identical wardrobe; consistent height/stance
• Head still matches hero; proportions consistent across views
• Same detail emphasis; quality: "high" + high res when available

Keep both when useful: bust sheet for face lock; full-body for whole-figure staging.

Step 3 — Environment plate

Empty set only — no people. Lock camera, furniture, and lighting before the insert.

State practicals explicitly (window cool vs lamp warm, etc.) so scene placement has sources to match.

Step 4 — Scene placement still

Composite the character into the locked environment plate.

Preserve the entire scene exactly: <env elements>.
Do not alter camera angle, framing, perspective, or environment.

Insert the subject from the character reference sheet — <short identity anchors>.
Pose: <body contacts, gaze, expression>.
Wardrobe: <from clothing ref if any — fabric behavior in THIS pose>.
Anatomy fully correct — natural joints, no elongation, believable foreshortening.
Match the scene's existing light direction, color temperature, intensity, and softness
exactly — subject lit only by sources visible in the plate (no extra beauty key).
Warm/cool splits, falloff, and contact shadows must match furniture in the plate.
Skin texture preserved — no retouching.

referenceMediaRefs order: [environmentPlate, characterSheet, …] — state each ref's job in the prompt.

Environment plate wins on camera, set, and lighting. Sheet wins on face. If the insert looks beauty-lit or flatter than the plate, reroll with explicit dual-source / falloff language.

Identity lock block

Keep a short stable paragraph and paste into every later still prompt (update to the actual character):

Olive-warm skin, dark brown eyes, strong brow, full lips, high cheekbones,
defined jawline, long dark brunette hair with natural wave.
Natural skin texture — visible pores, no smoothing.

Consistency > poetry.

Wardrobe changes

1. Neutral garment on the identity sheet when possible
2. Separate clothing still via generate_image / import
3. Scene placement binds fabric behavior to pose

Anti-patterns

• Jumping to video from this skill (hand off instead)
• Sheet panels with gutters or different light/crop → four different people
• Soft / low-res sheets when face lock matters
• Scene placement that “improves” the plate's camera or beauty-relights the subject
• Rewording identity every call → slow face melt

Handoff (optional)

When stills are approved and the user wants motion: pass scene still + sheets to video-prompting / cinematic-motion. Suggested ref ranking for those skills: env/scene → sheet → face close-up — declared in their prompts, not here.