comic-image-generation-adapter

name	comic-image-generation-adapter
version	1.1.0
category	comic-consistency
description	Unified interface for image generation calls (GPT Image, Higgsfield, ComfyUI, FLUX, etc.). Defines the canonical prompt assembly contract that injects World Bible context, character DNA, shot-plan directives, style memory, and negatives — in a fixed order.

Comic Image Generation Adapter

Core principle: Generation calls should be simple for the caller while automatically carrying all consistency context — assembled in one canonical order, so two panels never differ because their prompts were organized differently.

This skill is the execution layer connecting the consistency stack to actual image models.

When to Use

• Any time a skill needs to generate images while respecting the full consistency system
• When switching between generation backends
• When debugging "same inputs, different look" — the first suspect is assembly-order drift

The Prompt Assembly Contract (Canonical Order)

Every generation prompt is assembled in exactly this order. Sections come from their owning artifacts — the adapter composes, it never authors:

#	Block	Source	Notes
1	STYLE LOCK	Style skill Prompt Block	Verbatim; never paraphrased
2	FORMAT DIRECTIVE	Project contract	Canvas, aspect, panel count, reading direction (e.g., "one wide 21:9 image, three panels, left to right")
3	CHARACTER DNA	World bible dna_template per character	Stacked in screen-position order (see multi-character rules); per-scene overrides applied
4	PANEL DIRECTIVES	Director shot plan	Per panel: beat, shot size, camera angle, staging, eyeline, dialogue text + placement zone
5	SCENE CONTEXT	World bible world_register	Location rules, lighting grammar, time/weather
6	CONSISTENCY ANCHORS	Consistency config	Seed policy, reference-image handles, LoRA/IP-Adapter weights (backend-dialect section)
7	NEGATIVE BLOCK	World bible negative_library	Project + per-character + per-style negatives, merged and deduplicated; never hand-edited

Budget rule: when the backend's prompt budget forces cuts, cut from the bottom of block 5 upward — style lock, format, and DNA are never truncated.

Worked Mini-Example (Blocks 1–4, Abbreviated)

[STYLE] 1970s-80s retro hand-inked Japanese manga style, black and white,
G-pen brush outlines, soft screentone shading, hand-lettered bubbles…
[FORMAT] One wide 21:9 image, exactly three panels, read left to right.
[CHARACTER: Rabot] young man, short dark hair, sharp features, navy
jacket over white shirt, small scar on left cheek.
[PANEL 1 — SETUP] medium shot, eye level; Rabot at a control-room
console, fg hands on console, bg monitors; gaze toward panel right;
speech balloon upper-left dead zone: "Quiet shift tonight."
[PANEL 2 — REINFORCE] close-up, slight low angle; …

Backend Dialects

Backend	Dialect Notes
GPT-Image-class (single-call multi-panel)	Whole strip in one call; format directive is load-bearing; consistency anchors live in the text + attached references
ComfyUI / SD-class	Panel-by-panel + stitch; blocks 6–7 become node parameters (LoRA weights, IP-Adapter refs, negative field); seed policy per comic-character-consistency-system
FLUX Kontext-class	Reference-image conditioning carries DNA; text DNA block is shortened to signature marks only
Higgsfield	Per its adapter (see comic-production); assembly order unchanged

The assembly order is invariant across backends — only the delivery mechanism of each block changes.

Key Responsibilities

• Accept high-level requests (character + scene + style + shot plan)
• Resolve and inject all seven blocks automatically from their owning artifacts
• Support single-panel and batched long-sequence generation (Consistent Self-Attention where available)
• Emit a generation log: blocks used, artifact versions (bible version, shot-plan id), backend, seed — so any panel can be reproduced or audited
• Provide hooks for the Director's final cut and Producer review gates

Failure Modes to Catch

• Paraphrased style locks ("retro manga style" instead of the verbatim block) — the top cause of style drift between sessions
• Negatives pasted manually instead of merged from the bible
• Shot-plan fields silently dropped when budgets are tight (cut scene context first, never directives)
• Generating without a shot plan at all — the adapter must refuse

Integration

• Depends on all consistency-layer skills; consumes comic-director shot plans and the Producer's contract
• Used by every comic-pipeline skill and all style skills

---

Generation should be the easy part. The hard part is making sure every call carries the right context — in the right order, every time.