Image Creator

image-creator

Turn a Canon Spec + a scene into a maximally-detailed, reproducible generation prompt that renders a character to spec — then hand the result to image-analyser to verify. The loop partner. Schema + rubric + loop: canon-spec.md.

When to use

• "Generate / render / create this character", "render Veikko in scene X to spec", "make the image match the canon".
• Inside the fidelity loop, fed by image-analyser's correction directives.

NOT for: scene blocking / motion (→ video-director, motion-choreographer, which take a verified still from here), non-character art (→ canvas-design).

Input

• Character id / Canon Spec (agents/reference/ai-video/<project>/characters/<id>.json).
• Scene brief (setting + pose). Optional: prior image-analyser diff (loop mode).

Procedure

1. Governance gate FIRST (per media-governance-routing): if the character is a real-person likeness, consult agents/settings/policies/media/likeness.md
   • public-figures.md + disclosure.md before emitting anything. Fictional characters (odins-beard trio) are exempt; the routing decision is in-session.
2. Provider gate (per provider-lifecycle-discipline): read the resolved provider's tier; if non-stable (experimental/deprecated/community), surface the tier and ask before running. Never default to a non-stable provider silently. Name the provider + tier in the run summary.
3. Assemble the prompt from the spec — anchors first. Order matters: the hard-to-render identity_anchors go at the TOP (the canon's lesson — heterochromia + hair-split get forgotten if buried). Then physique, face (+ marks), per-location tattoos (incl. exact text), outfit, jewelry.
4. Asymmetry block — for split / heterochromatic characters, an explicit left/right section ("LEFT half black / RIGHT half blond", "blue LEFT eye / green RIGHT eye") with concrete comparison refs for engines that drop it.
5. Negative block + engine settings — reuse the canon's proven structure (negatives that kill "single hair colour", "both eyes same colour", etc.; per-engine --ar/--style/CFG/steps). Do not reinvent; the character book's prompt format is the template.
6. Generate through the existing provider/adapter layer (scripts/ai-video/adapters/, the /video|image surface). Do not add a new provider path where one exists.
7. Verify — hand the output to image-analyser; in loop mode, fold its correction directives into the next prompt (see the loop in canon-spec.md).

Anchors-first — why

The fidelity-loss evidence (img_2.png): Veikko's split hair + heterochromia were missed because they are hard for the engine and were not front-loaded. The analyser-derived fixes go to the TOP of the next prompt, with negatives, and the loop re-verifies — that is how the smallest mole comes back.

Output format

1. Generation prompt — anchors block · positive · asymmetry (if any) · negative · engine settings.
2. Provider + tier line (the audit entry).
3. Verify call — the image-analyser invocation on the result + the loop stop-state.

Example (anchors-first vs buried)

• Safe (Veikko): prompt opens with HAIR: exact vertical centre split, LEFT pitch-black, RIGHT platinum-blond + EYES: heterochromia — LEFT ice-blue, RIGHT forest-green, then the rest.
• Unsafe: physique/outfit first, the split + heterochromia in a trailing sentence → the engine drops them (the observed img_2.png failure).

Gotchas

• Never claim "canon-perfect" without an image-analyser pass (per verify-before-complete).
• DALL-E drops heterochromia + split hair — use concrete comparisons; expect multiple generations.
• Keep the spec the single source: regenerate from the Canon Spec, never paraphrase it into a fresh prompt.

Do NOT

• Do NOT claim "canon-perfect" / "matches the canon" without an image-analyser pass on the output (per verify-before-complete — no completion without evidence).
• Do NOT bury the identity_anchors — the hard-to-render features (heterochromia, hair-split) go at the TOP of the prompt, never in a trailing sentence.
• Do NOT default to a non-stable provider silently — surface the tier and ask first (per provider-lifecycle-discipline).
• Do NOT emit a prompt for a real-person likeness before the governance gate (step 1) clears it.
• Do NOT add a new provider/generation path where the existing adapter layer (scripts/ai-video/adapters/) already covers it.
• Do NOT paraphrase the Canon Spec into a fresh prompt — regenerate from the spec so it stays the single source of truth.

Policies

media-governance-routing + agents/settings/policies/media/ (likeness, style, public-figures, disclosure) — consulted in step 1 before any prompt is emitted for a real-person likeness.

Related skills

• image-analyser — the verify/loop partner.
• character-consistency — supplies the locked identity tokens.
• video-director / motion-choreographer — take the verified still into scene + motion.