photo-to-editorial-illustration

name	photo-to-editorial-illustration
description	Use when a user wants to transform one photograph containing people, animals, landscapes, architecture, plants, or objects into a minimal mid-century European couture editorial composition pairing the faithful source photo with a generated illustration board.

Photo-to-Editorial Illustration

Require exactly one source photograph. Request it when it is missing.

Use this fixed sequence:

SOURCE PHOTO → SOURCE PREFLIGHT → CONTENT DISTILLATION → ONE ILLUSTRATION GENERATION → ILLUSTRATION EVALUATION → TWO-LINE ENGLISH COPY → DETERMINISTIC COMPOSITION → FINAL IMAGE

Prepare the Source

1. Create a fresh internal work directory for the run and reserve a new .png path inside it for normalized output.
2. Run prepare_source_image.py with the provided photo and normalized-output path. JPG/JPEG, PNG, and static WebP pass through unchanged. MPO and HEIC/HEIF use only their primary image and are normalized to a verified single-frame RGB PNG. APNG, animated WebP, GIF, files over 50 MiB, images over 100 MP, and images with an edge over 20,000 px are unsupported.
3. Read the returned JSON. Use runtime_source for content inspection, use runtime_source for generation attachment, and use runtime_source for composition. Keep original_source only for traceability.
4. If HEIC/HEIF reports that pillow-heif is unavailable, run setup_image_runtime.py with --inspect. When the runtime is ready, rerun preflight with the reported Python path. When installation is required, pause for explicit user approval before running --approve-install, then rerun preflight with the installed runtime.
5. If source preflight fails, report the error and stop before content selection or image generation.

Generate the Illustration

1. Use content-abstraction-grammar.md to resolve KEEP_ELEMENTS and DROP_INFORMATION from the source.
2. Use element-abstraction-grammar.md to define how each retained element is simplified. Compile only categories actually retained by content distillation; do not send absent Human, Animal, Architecture, Vegetation, Furniture / Objects, or Environmental Field rules.
3. For every retained human, apply geometry-grammar.md and face-policy.md, then use the Human Prompt Fragment in generation-prompt.md.
4. For every retained animal, use the Animal Prompt Fragment in element-abstraction-grammar.md.
5. Use the Abstraction Prompt Fragment and Visual Prompt Fragment from style-media-grammar.md exactly as written. Add its Watercolor Eligibility Prompt Fragment only for explicitly retained distant continuous environmental fields; humans and animals are never eligible.
6. Use the Illustration Safety Margin Prompt Fragment from layout-grammar.md exactly as written so the generated illustration preserves its complete image boundary without clipped peripheral marks.
7. Assemble generation-prompt.md by replacing every placeholder once.
8. Send the assembled prompt and source photograph in one image-transformation call.

Set APPLICABLE_REPRESENTATION from retained categories only:

• retained human: Human Prompt Fragment, with one short Face Policy expression signal only when applicable;
• retained animal: Animal Prompt Fragment;
• retained architecture, vegetation, furniture / objects, ground, wall, water, or sky: only the matching element instructions;
• absent category: no fragment.

When a human is the Anchor, merge retained environmental facts into one coherent subordinate support.

The model returns one illustration at the source aspect ratio on continuous warm-white paper. The source photograph remains an input and is not reproduced inside the generated image.

Evaluate the Illustration

Read quality-guardrails.md after generation. Continue only when the illustration passes every applicable check. Report a failed check instead of initiating an automatic generation retry.

Compose and Return

1. Create two quiet English lines: a 2–4 word poetic title lightly grounded in the scene, and a natural 4–8 word postcard-style sentence expressing a gentle wish, mood, or memory. The subtitle must not describe, count, identify, or summarize visible subjects.
2. Run compose_editorial.py with runtime_source, generated illustration, title, subtitle, and output path.
3. Return the compositor output as the final image.

The compositor is the sole owner of final presentation. For landscape and square sources, it places the photo and complete generated illustration image top-to-bottom. For portrait sources, it places them left-to-right. The photo is contained at its original aspect ratio without cropping or stretching. The complete generated illustration image keeps its rectangular boundary and aspect ratio; the compositor does not crop it down to visible marks. Use the estimated illustration paper color for the final canvas. Scale the two-line Caveat Regular footer from the final canvas short edge and use a line gap equal to 0.525 times the font size. Align the actual right edge of both text lines directly to the right edge of the complete placed illustration image, and give the footer bottom edge the same vertical safe margin used above the illustration image. Make each safe margin twice the gap between the illustration image and the first text line. Its packaged script, assets, and behavior tests define that contract.