GPT 이미지 2.5 프롬프팅 가이드: 8가지 원칙

Official Image prompting guide: Model selection: - Flare (gpt-image-2.5-flare): small model, speed-oriented; image quality roughly equivalent to GPT Image 2 - Sunburst (gpt-image-2.5-sunburst): base model, quality-oriented; image quality higher than GPT Image 2 8 principles for writing prompts: 1. Define the final output first: purpose (product photo / advertisement / diagram), composition, aspect ratio, placement. For complex requests, use sections like "scene / subject / details / constraints". 2. Format is flexible, readability first: short sentences, paragraphs, JSON-like structures, or tags are all fine - don't rely on special syntax. 3. Write what is visible: materials, light, color, medium. For real photos, write photorealistic / real photograph. Lens parameters are only visual cues, not physical simulation. Mood words are not enough - include scale, weather, and color temperature. 4. Be specific about people and actions: full body / whether feet are in frame, eye direction, how hands grip objects. 5. Put precise text in quotes: number of occurrences, position, font; spell out uncommon words letter by letter; request no extra text; use medium/high for small text. 6. Separate changes and locks: for edits, write change only X, then list identity, geometry, layout, lighting, and labels that must be preserved; also state what not to change. 7. Divide reference images by role: specify by number which is the subject, style, clothing, background, and how to composite them. 8. Change only one thing at a time: use the previous output as the next input; rewrite key constraints each time; if pixel-level consistency is needed, composite in post - don't rely on prompts alone. Migration advice: first save baselines (hard-to-edit, precise text, faces, product geometry, transparent images), compare instruction following, identity/product fidelity, text, unwanted changes, and transparency, then gradually shift traffic.