Composition, Style, Text, and Editing
Goal of this lesson
Break “visual effect” into layout, visual language, text layer, and edit-protection areas, and learn workflow fallbacks for unstable model behavior.
Composition before decoration
Composition determines subject position and visual hierarchy on the canvas. At minimum, state:
- subject count, relative size, and position;
- viewpoint, shot size, and horizon;
- foreground, middle ground, and background contents;
- safe areas for title, buttons, or crop; and
- the direction guided by gaze, action, or light.
The rule of thirds and centered composition are heuristics, not hard rules. Actual acceptance is whether downstream use works: is the product complete, are operating steps readable, and does the subject remain prominent in a thumbnail?
Describe style with visual features
“In the style of an artist” can be vague and can create rights or platform-policy issues. Prefer visible features: medium (paper cut, pencil, 3D), line work (heavy outline, no outline), palette (low-saturation cool colors), material (matte paper), lighting (soft sidelight), detail density, and period design language. Retain lawful brand colors, font licenses, and reference-asset provenance.
Give text generation a fallback path
Text inside an image requires character correctness, sound typography, and visual integration at once. Even if a current model improves significantly, verify character by character—especially Chinese, numbers, medicine names, and URLs. A reliable workflow commonly is:
- Generate a text-free background with a reserved text safe area.
- Add formal text with a typography tool.
- If text must blend into the scene, treat the text-bearing version as a candidate and retain the text-free version.
- OCR can assist checking; a person must make the final comparison against source copy.
The protection contract for edit tasks
A local edit must at least contain asset_id, source_revision, source_hash, mask_hash, editable_region, protected_region, desired_change, and difference acceptance. Record actual execution and post-production separately as transform_id; a candidate for publication has its own release_id. Do not infer these three IDs from filename or prompt text. Do not only check whether the target region improved. Run a protected-region regression too: have pixels that must not change in a face, logo, product label, or scientific image drifted?
After scaling, automatic rotation, or color-space conversion, a mask can be misaligned. Diagnose in this order: check width/height and orientation metadata, visualize mask boundaries, then tune the prompt only last. Reading a reference image, automatic scoring, and human preview must all pass object-level authorization/ACL first. On revocation or deletion, propagate through asset_id → transform_id → release_id to candidates, caches, and publication links.
Common errors and diagnosis
- Subject cropped out: record safe area at the task level and preview with the actual delivery template.
- Style overwhelms content: lock composition and information first, then add style.
- Text is approximately correct: compare character by character; “looks similar” is not enough.
- Editing contaminates the whole image: reduce change scope, preserve a baseline, and run difference checks.
- Color is inconsistent: record color space and post-production export settings; calibrate brand colors afterward.
Exercise and self-check
Draw a text-layout frame for a “portrait course cover” (paper is fine), then write its safe area, subject box, and crop rules. Rewrite “cyberpunk style” as at least five visible features, and state which are soft preferences.
Next, put these requirements into a stable interface: Workflows and API Boundaries.
References
- OpenAI Image generation guide (generation, editing, and output-settings examples; checked 2026-07-22)
- C2PA Content Credentials 2.4