--- name: comic-galaxy-image-production description: "Produce and review coherent text-to-image sets for the project, including twelve fixed theme groups, three-IP character consistency, nine-image prompt planning, batch generation, seed and model metadata, visual QA, and AI compliance. Use for this project's image production or review; not for general UI design or unrelated image generation." metadata: short-description: Image set production for the project codename --- > Original skill · commissioned by the user · English edition, 2026-11-12 # Image Production Use this skill to plan, generate, review, and archive image sets for the project. Preserve the locked twelve-group structure and keep the three main IP characters consistent across all 108 images. ## Non-negotiable structure - Keep exactly 12 groups with 9 images each: 108 images total. - Do not add a 13th group or rename a group unless the user explicitly changes the content architecture. - The locked group names are: 1. `` 2. `` 3. `` 4. `` 5. `` 6. `` 7. `` 8. `` 9. `` 10. `` 11. `` 12. `` - `` contains 2-3 images of `` and ``, replacing ordinary mountain images rather than increasing the group size. - `` contains `` night scenes. Do not make `` a separate song or group unless the user changes the locked structure. - Each group has one distinct visual style. All nine images inside a group share the same style, palette, line quality, light logic, texture, and rendering rules. - The three-IP presence target across the full library is 30%-35%, with all three together in 15%-20%. The Q-version group has all three in 9/9 images. Normal groups have 2-3 images with characters; key story groups may have 3-4. - Every image needs a unique image ID, a unique visual idea, a prompt record, a model and version record, and a review status. ## Style-content contract - The project uses 12 distinct primary style families, one per group. - All nine images inside one group must use the same style lock: same line quality, palette logic, light logic, texture, material language, and rendering rules. - A different style must be a different visual language, not a palette swap, filter, crop, or colour grade. - The style must be derived from the group's content: material, subject, culture, weather, time, emotion, and song function. Do not choose a style only because it looks attractive in isolation. - No two groups may share the same primary style family. If two groups look confusable in a thumbnail grid, revise one of them. - Series-level consistency remains required: the three IP characters, aspect ratio, resolution floor, quality bar, composition discipline, and AI compliance rules stay consistent across all groups. - Before generating a group, read the style isolation matrix in `references/style-bible.md` and confirm the group's style, content driver, and forbidden overlap. ## Workflow 1. Confirm the target group, song, and nine-image story before generating. Read `references/style-bible.md` and `references/character-consistency.md`. 2. Design the nine images first. Each image must differ in subject, scene, action, time, weather, emotion, or camera. A filter-only, crop-only, or pose-only variant is not a new image. 3. Build prompts with `references/prompt-templates.md`. Lock the style tokens for the group and the character reference tokens for any image that includes ``, ``, or ``. 4. Generate 3-5 candidates per image with the same model version and style lock. Do not accept the first output as final. 5. Select the best candidate, then correct character details, hands, faces, clothing, and composition through inpainting or outpainting. Upscale and color-grade only after the selection is stable. 6. Run QA with `references/qa-and-compliance.md`. Reject images with character drift, wrong clothing, extra limbs, text artifacts, protected IP, brand logos, celebrity likeness, duplicated composition, or style mismatch. 7. Record the prompt, negative prompt, model, version, seed, LoRA, IP-Adapter, ControlNet, resolution, sampler, review status, AI label, and commercial-use evidence. 8. Update `images.json`, `prompts.json`, `styles.json`, `characters.json`, and `music-videos.json` before marking a group complete. ## Tool boundaries - Use `$imagegen` for concept exploration, reference sheets, single-image edits, and small-batch generation. - Use a production text-to-image model or API with commercial-use terms and reference-consistency support for the final 108 images. - Use `$frontend-design` or `$ui-ux-pro-max` only when broader visual direction is requested. The group style rules in this skill take priority for image production. - Use `$wechat-mini-program-skill` or `$wechat-mini-program-ui` only for app integration, data structures, display, and performance. They do not generate images. ## Hard guardrails - Do not put Chinese text, calligraphy, inscriptions, shop signs, slogans, or readable text inside generated images. Add text in post-production. - Do not imitate a living artist, celebrity, photographer, or protected style as a selling point. - Do not generate modern brands, film or television characters, celebrity faces, protected logos, or copyrighted characters. - Do not present an AI image as a real photograph, historical record, or factual reconstruction. - Keep every approved original, candidate, prompt, model version, seed, and review note. Do not overwrite approved images. - If commercial-use rights, character consistency, or content safety cannot be established, stop and surface the blocker rather than shipping the image.