A graphic style of flat large color blocks + fine dense grain. Both modes share the same style DNA.
Style DNA (six items, unchanged for any subject)
- Fine, even dotted grain, like spray paint through a stencil;
- Grain only on the subject, background absolutely clean(a single flat color, zero noise, zero texture) — this is a condition of validity, not an option;
- Edges are grain-eroded: shapes are hard, recognizable, sharply focused, with boundaries showing fine dotted burrs; neither clean vector lines nor softened blur;
- A small number of large planes, one color per plane, no tonal splits within a plane;
- Closed 5–6 colors, no gradients, no outlines;
- The silhouette is recognizable at first glance, graphic feel takes priority over realism.
Item 2 is the easiest one to lose. Most models will by default spread grain over the entire image, and once the background is dirty, the style immediately collapses into 'an illustration with added noise'.
Mode discrimination
- The user provideda photoand wants to change the style →
references/photo-restyle.md - The user wantsgeneration from scratch(a cover, an illustration, an image of some subject) →
references/from-scratch.md - The user wantsa version that can be sent to someone else as a whole block, where they only fill in the subject and it works →
ARTICLE-COPY.md - If it is not clear, ask one question first, do not default
Photo style change
Already tested and finalized, copy the whole block without changing a word. The user's action is three steps: upload photo → paste prompt → generate. No API needed, no reference image needed.
- Do not use the multi-image style reference path— that requires preparing a reference image first, and the deliverable changes from the prompt into that image.
- Do not 'optimize' this prompt: three improvement attempts all degraded in testing and have been reverted; before changing anything, read the failure records in the reference.
- To change the temperament, replace only the
COLOUR —section (nine alternatives are in the reference),do not touch any other section。
Generation from scratch
Just fill in the template:subject type × color logic × camera angle and compositionfreely combine. The DNA fixed block is reused verbatim, only the palette is swapped. Six reference examples are in references/examples.md。
Before combining, look at the combination taboos table at the end of the reference — three of the five (top-down view plus perspective, bright colors appearing in multiple places, distant subject too small) will be violated if not written into the prompt.
Not tested item by item: the DNA fixed block itself has been verified in photo style change mode; the template and the six examples are combinations compiled according to it. If the output deviates, first suspect the naming in the Subject slot, do not touch the fixed block.
Prompt writing rules
Seven sections, with section labels, in this order:Use case / Scene / Subject / Key details / Composition / Text in image / Constraints。
Merge exclusions directly into the Constraints section,never output them separately as a negative prompt block— Models like GPT Image 2 have no negative channel; constraints placed in standalone blocks will only be lost or drawn as positive descriptions.
Every prompt must be complete and directly copyable; do not split key constraints into vague notes for users to assemble themselves.
Known trade-offs (accepted items, do not repeatedly fix)
| Phenomenon | Status |
|---|---|
| Background hue shifts blue | Accepted item, no longer fixed. History see references/test-log.md |
| Face shapes converge (all face shapes get elongated/sharper) | Accepted item, no longer fixed — style strength takes priority over face shape fidelity; the fix would weaken the styling strength, not worth the loss. History see references/test-log.md |
| Exaggeration magnitude and recognizability of the person trade off against each other | Accepted item. When dialing back, remove a single operation, do not reduce the overall magnitude — reducing the overall magnitude will directly fall back to realism |
Guardrails
- Describe style with visual features,do not name specific artists。
- Other people's photos require their consent.
- Front-facing flat-lit selfies work worst; source photos with clear single-side lighting are most stable — you can remind users, but do not make it a required step.
- If unsatisfied with the output, just regenerate once directly; do not guide users through multiple rounds of correction (it would break the three-step flow).