A character bible is your cast’s source of truth: identity, visuals, personality, and do/don’t rules. Ingestion makes that canon usable by AI systems—templates, datasets, and parameters that guide generation at scale.

Benefits:

• Identity stability across shots and episodes
• Faster iteration with fewer retakes
• Cross-team alignment (writers, designers, animators)
• Safer outputs via explicit constraints and negatives

Model only what you intend to enforce. Recommended fields:

• Identifiers: character name, aliases, unique token(s)
• Logline: one-sentence identity hook
• Visual anchors: hair/eye color, silhouette, key shapes, proportions, signature items
• Palette: primary/secondary HEX/RGB swatches and material notes
• Wardrobe: base outfit + approved variants (seasonal, arc-based)
• Expressions/poses: approved range with examples and limits
• Behavior/voice: tone, speech patterns, catchphrases
• Relationships: key dynamics that affect staging or framing
• Negative constraints: out-of-scope styles, colors, props, time periods
• References: curated images with alt text and captions
• Licensing/rights: usage bounds for any third-party elements
• Version: bible version, date, owner

Organize for machines and humans:

• Folder layout: /characters/{name}/ with subfolders for text, images, palettes, poses
• Manifests: a sidecar JSON (or YAML) defining traits, palettes, and file references
• File names: include role, view, variant, and tags (e.g., hero_fullbody_v1_redcoat.png)
• Captions/alt text: concise and trait-aligned (support training or retrieval)
• Palettes: swatch image + HEX list; include tolerances (e.g., ±ΔE 3)
• Image curation: diverse angles, clean backgrounds; avoid near-duplicates
• Dataset balance: equalize outfits/expressions you want reproduced
• Repro seeds: store known-good seeds and prompts with outputs for baselines

Common routes, from light to heavy:

• Prompt templates only: encode canon in reusable prompts, negatives, sampler/CFG defaults, and seed recipes. Best for simple, stylized casts.
• Reference conditioning: use curated reference crops (face/fullbody) and conditioning adapters to anchor identity without training. Good for quick deployment.
• Textual inversion/embeddings: train a token (or small set) linked to the character. Pair with strict negatives and palette constraints to limit drift.
• LoRA/adapter fine-tunes: small, composable fine-tunes from a balanced dataset. Best for high fidelity, multiple outfits, and side-cast control. Watch for overfit and style bleed.
• Retrieval-assisted prompt assembly: index bible text and use retrieval to auto-build scene prompts and negatives consistently.

Choose based on fidelity requirements, runtime constraints, and team skills.

1. Draft the bible: capture only enforceable traits and rules.
2. Normalize: write atomic trait statements; remove ambiguous terms.
3. Tag and caption: consistent tags across images; align with prompts.
4. Pack a manifest: traits, palettes, references, and defaults in JSON.
5. Choose ingestion path: template → reference → embedding → LoRA.
6. Build templates: base prompt, scene prompt, style prompt, negatives; set default sampler/CFG/steps.
7. Calibrate: generate a grid across seeds, angles, and lighting; log metrics.
8. Lock and publish: freeze versions, attach repro seeds, and share kit.
9. Integrate: wire into your scene/panel pipeline and project management.

Track quality with objective and subjective checks:

• Identity similarity: face/feature embeddings vs. reference set
• Color fidelity: palette ΔE to approved swatches
• Proportions: key ratios (head:height, eye spacing) within tolerance
• Caption alignment: trait recall/precision from auto-captions
• Prompt reproducibility: variance across seeds and lighting
• Human review: pass/fail on do/don’t rules and brand notes

Procedure: run a fixed test grid per version; log metrics and examples; only publish when thresholds are met.

Treat the bible as a versioned product:

• Semantic versions: MAJOR (identity or style shifts), MINOR (new outfit), PATCH (prompt tweaks)
• Change log: what changed, why, and impact on back-catalog
• Variant strategy: separate adapters/tokens per outfit or era
• Environment notes: model/backbone, sampler, CFG, steps, control weights
• Deprecation policy: retire old variants with guidance for replacements

• Style drift: increase negative strength, tighten palette constraints, rebalance dataset, reduce CFG or steps
• Overfit/plastic look: lower learning rate/epochs (for embeddings/LoRA), add diverse lighting/backgrounds, remove near-duplicates
• Token collisions (name clashes): use unique, non-dictionary tokens
• Outfit bleed: split per-outfit adapters; reduce mixing; add clear captions
• Inconsistent faces between angles: add side/back angles; use multi-view references
• Prompt brittleness: simplify templates; move scene details to separate clauses; lock core identity first