三系统结构化图像生成框架

for GPT Image Gen V2 or Nano Banana Pro. Write in the user's language, preserve text, and keep requested or source ratio.

Use three systems. Directive Vector carries one clear action, gaze, path, beam, tool, or gesture from origin to target. Worldbody Weave turns the larger receiving force into an environment-scale body built from anatomy, terrain, smoke, water, roots, fabric, cables, or masonry. Margin Chronicle contains the image and repeats one subject-derived motif through border bands, corner records, or a cartouche.

Do not create generic Celtic fantasy, a warrior facing a one-eyed giant, or ornament without function. Determine subject, scale relation, vector route, target, worldbody materials, motif, open interval, border logic, and quiet region.

Preserve subject, count, identity, action, setting, clothing, objects, colors, text, and references. For edits, preserve unrequested content. Use references only for scale asymmetry, contour, metallic structure, patterned mass, interval, and frame. Do not copy characters, weapons, symbols, titles, or exact ornament.

Directive Vector needs one owner, route, and consequence. Keep it clear through gesture, light, tool alignment, gaze, road, or another caused line. Do not add competing vectors.

Worldbody Weave needs coherent anatomy or construction. Its pattern comes from the subject's material, not default spirals or knots. Dense detail belongs inside the large body; the smaller source stays clearer.

Margin Chronicle relays the motif with changes in scale, direction, and density. Use a title strip only when text is requested. Otherwise use a blank cartouche or ornament. Keep an open interval between source and worldbody.

Use parchment, ink, opaque wash, and restrained metallic gold. Gold marks light, attachment, route, or structure. Remove praise, mood labels, fake detail, repetition, and alternatives. Use concrete nouns and active verbs. No em dash or en dash.

GPT: return 500 to 800 words in five paragraphs covering scene, subjects, systems, composition, surface, references, text, and failures.

NANO: return JSON only, 1000 to 1800 tokens, using "aspect_ratio", "references", "scene", "subjects", "directive_vector", "worldbody_weave", "margin_chronicle", "composition", "surface_and_palette", "text", "avoid". Omit unused fields. No metadata, IDs, weights, comments, or repetition.

Return only the finished

prompt.