
原文
GPT Image 2 · en
Using the provided reference image, redraw and upscale it into a cleaner, sharper, more polished version while preserving the same overall infographic layout, all original text content, and the Longjing tea theme. Keep the product title {argument name="tea name" default="龙井茶"} and the English subtitle {argument name="English tea name" default="Longjing Tea"} unchanged.
Goal: Make the image look less like a rough GPT-generated graphic by removing broken, noisy, meaningless textures and replacing them with crisp, intentional detail.
Enhancements to apply: Increase resolution and clarity; sharpen typography, icons, borders, tea leaves, glass, and card edges; make the lighting softer and more natural; improve material realism for the tea, cup, dried leaves, and ceramic dish; clean up any fuzzy or fragmented artifacts.
Meaningful detail additions: In the main hero panel, replace the vague pale background with a refined scenic tea plantation landscape of layered green terraces, misty mountains, and a wooden tabletop beneath the tea cup and plate. In the origin card, add a small matching tea-terrace landscape illustration at the bottom. Make the bottom green banner more vivid with a natural leaf-vein texture.
Layout constraints: Preserve exactly 7 visible content panels/sections: the large hero panel, origin card, features card, flavor card, brewing suggestion card, grade-and-picking card, and bottom slogan banner. Preserve exactly 4 feature rows in the features card, exactly 3 small icon benefits in the hero panel, and exactly 3 brewing parameter icons along the bottom of the brewing card. Do not add extra sections, extra labels, watermarks, logos, or decorative clutter.
Style: Elegant Chinese tea brochure design, cream-and-sage palette, refined minimal UI cards, soft shadows, crisp vector-like icons, realistic tea photography elements, calm premium product-advertising finish.使い方
- パラメータを含む元のプロンプト全体をコピーします。
- 被写体、文字、構図を自分の要件に置き換えます。
- 対応する外部ツールで使用し、利用条件と結果を確認します。
開始前の確認
- 参考画像の要否を確認してください。互換モデルの表示は、そのモデルでの生成実績を意味しません。
- 名前、数値、画像内の文字を確認します。モデルの版で結果は変わります。





