Compilea themeintoaprompt that can be handed directly to an image model. Style locked down: one pure flat color field, one realistic portrait, one line of extra-large compressed bold title truncated by the person's head and shoulders, one hand-drawn element pressed into the lower foreground corner.
Core principles:one field of color, three layers of texture, one occlusion.
The premium feel of this style does not come from having many elements; it comes from three things holding true at the same time — the base color is absolutely flat (no gradient, no scene, no vignette); the photo layer, the flat color layer, and the hand-drawn layer are three textures in the same frame but each keeps its own identity, without blending into each other; the second line of the titlemust truly be blocked by the head and shoulders. If any one of these collapses, it falls back to an ordinary poster with text added.
Aspect ratio
2:3 vertical is locked down. For image generation parameters use 2048×3072、quality high— placeholder text lines, scale lines, and the thin narrow counters of the compressed bold all consume resolution; at low resolution these details will inevitably blur into noise.
Other aspect ratios are unverified; change only if the user explicitly requests it, and remind them that the layout system's margin proportions need to be re-measured.
Two layout registers
Pick one at a time, and keep the whole set consistent. Once chosen,four places change together in unison, you cannot change only one place.
- CLEAN (default): thin lines, wide margins, strict alignment to grid, index at the right edge, scale line positioned at lower left. Suitable for fashion, beauty, culture, interviews.
- ROUGH: thick lines, tight margins, elements slightly out of bounds, index at the left edge, scale line positioned at lower right (an overall mirror of CLEAN). Suitable for sports, street, youth culture.
The register changes only the layout discipline, not the color field logic or the occlusion structure.
Read
- Compile prompt: read
references/compile-contract.md(iron rules + locked-down core + two registers + fixed skeleton). - Choose color field, bridging element, camera position, or do them in batches as a set: read
references/registers.md(four module tables). - For verified finished references: read
references/examples.md(the original text of four tested prompts).
Input
Infer from user input; filling out a form is not required:
- Two words for the title: a short word for the first line, a long word for the second line. The second line is the load-bearing item — it must be wide enough to be truncated by the head and shoulders. For constraints see compile-contract.
- Category tag: the line at the upper left,
<X> FEATURE. If the user did not provide one, devise one based on the theme. - Color field register: check
registers.md's color field table. Four sets verified, four sets pending verification. - Subject: person + shot size + pose + hand treatment.
- Bridging element: a material metaphor for the theme, and it must share the same origin as something on the subject.
- Layout register: CLEAN or ROUGH.
Workflow
- Set color field: First fix the base color and the three-color palette; it determines the title color and the tonality of the entire image. Once the palette is locked, every later slot can only choose from these three hues.
- Fix the subject and camera position: Check the camera position table. The hands must be decided on the spot—either out of frame, or clearly holding something nameable. There is no third option.
- Fix the bridging element: This is the crux of the whole set. It must satisfy two conditions at once—it is a material metaphor for the theme, and it is homologous with something on the subject (the dress's watercolor print ↔ the watercolor flower branch in the foreground). If you can't find a homologous relationship, swap the element; don't force in a decoration.
- Fix the bridging position: Place it in thelower corner opposite the light source. Light comes from the upper right, the element lands in the lower left.
- Fix the two-word title: Check the word-selection constraints. If the second line isn't long enough, change the words; don't expect the model to adjust by itself.
- Compile the prompt: Output the finished product according to the fixed skeleton of
compile-contract.md, with no placeholders, no menus, and zero Chinese in the body (except for rendered titles). - Output: registers + one complete prompt + (optional) one sentence on the rationale for choosing the color field and bridging element. When the user only wants the prompt, omit the explanation.
When producing as a group: the style base (the four sections: Scene structure / Composition / Text specs / Constraints) must be written verbatim intoevery singleprompt—the model only sees a single prompt, and shared rules are not inherited automatically. Between images, only swap the color field, subject, bridging element, and two-word title.
Quality Gates
- Base color perfectly flat: the appearance of gradients, vignetting, texture, or any scenery means failure. This is the first load-bearing wall of this style.
- Occlusion must occur: letters in the second line of the title must be genuinely truncated by the head and shoulders. Pressing them all in front of the person, or letting them all recede behind the person, both count as failure—this is the identity feature of this style; without it, it is just a photo with text added.
- Three-color palette closed: the whole image has only three hues. A fourth hue appearing means failure; the most common cases are ambient color leaking into the background or stray colors brought in by the clothing.
- The bridging element must be a painting: it must clearly read as hand-drawn/watercolor/ink/marker, andbe pressed in front of the photo. It being rendered as a real object (the watercolor flower branch turning into a real flower branch) is the most frequent collapse point of this style, and must be blocked by name in Constraints.
- Placeholder text must be unreadable: small text blocks must be non-word placeholder lines. Any readable word appearing means failure.
- Hands accounted for: either out of frame, or clearly holding an object. Hands lying loosely in the frame is the situation with the highest junk-image rate.
- Text exact: only the two title lines, the category label, and the three sequence numbers appear. A fourth type of text, a watermark, or a logo appearing means failure.
- This skill only produces prompts; it does not fabricate already-generated image results.