IMAGE-TO-IMAGE GUIDE · REVERSE PROMPTING · UPDATED 2026-09-04

How to Reverse-Engineer a Prompt from a Reference Image

When you like a reference image but do not know how to describe it, use AI to analyze the image, turn the analysis into a prompt, and generate a new result. This article documents the complete workflow.

How does this workflow work?

This workflow does not directly copy the source. It first translates the reference into visual information: subject, scene, composition, lighting, style, color, typography, and materials.

The resulting prompt remains editable. You can preserve composition and texture while changing the subject, environment, colors, or typography.

Upload reference→Analyze→Build prompt→Generate→Compare

Step 1: analyze the reference image

Upload the reference to a vision-capable AI such as DeepSeek, then use the template below. It asks for spatial relationships, camera angle, lighting direction, and material qualities—not just object recognition.

请扮演专业的图像分析师和提示词设计师，分析我上传的参考图，并生成一段适合 AI 绘图工具使用的中文图片提示词。请按照以下维度分析： 1. 主体：说明人物、产品或主要对象的外观、动作、表情、服装和数量。 2. 场景：说明室内或室外环境、背景元素、空间关系和前后景层次。 3. 构图：说明主体位置、画面比例、取景范围、镜头视角和构图方式。 4. 光线：说明主光方向、明暗关系、阴影、反光和整体氛围。 5. 风格：说明是商业摄影、电影剧照、插画、3D 渲染、胶片或其他视觉风格。 6. 色彩：说明主色、辅助色、冷暖倾向、饱和度和对比度。 7. 文字与图形：准确记录画面中的标题、副标题、Logo、按钮、标签和二维码位置。 8. 质感：说明材质、纹理、景深、颗粒、雾气、反射和其他关键细节。最后将分析结果整合成一段不超过 600 字的中文提示词，使用逗号分隔关键词，并在末尾给出推荐图片比例。不要虚构参考图中不存在的文字。

Step 2: extract an actionable prompt

The reference below is a portrait assembled from torn newspapers and vintage posters. “Collage portrait” alone is not enough to reproduce its scale, layers, palette, or lighting.

Reference image of a face assembled from torn newspapers and vintage poster fragments

Reference: vintage torn-paper collage portrait

The reverse-analysis template turns the image into executable information: what to draw, where it is, how it is framed and lit, and which materials define the look.

主体：由多层撕裂报纸和海报碎片拼接而成的人脸特写，表情平静，双眼直视前方，服装不可见，主体数量为一个面部。场景：类似贴满旧海报的破旧墙面，没有明确室内外环境。前景是撕裂、翘起的纸张碎片，后景是平整的背景海报，纸张与面部交织。构图与视角：脸部位于画面中央，占据约90%的面积，极近景特写，正面平视，中心对称构图。光影：均匀的漫反射平光，没有强烈定向主光，阴影主要来自纸张边缘的微小立体投影。色彩：复古黄、蓝、红为主，米白和黑色为辅，冷暖交错，整体偏暖，高饱和度，高对比。风格或媒介：混合媒介拼贴艺术、街头艺术、复古海报和插画风格。细节：清晰的撕裂纤维、多层纸张堆叠、卷边、剥落、粗糙表面和印刷颗粒，呈现平面媒介的立体化效果。限制条件：不要出现真实可读的现代标题、Logo、按钮或二维码，画面中的字母只作为旧海报残片出现，推荐4:3横向构图。

Enter the reverse-engineered prompt into Seedream 5.0 to generate a new image that carries forward the reference image’s visual language. The goal is not a pixel-perfect copy, but to preserve its subject scale, centered composition, vintage palette, and torn-paper texture.

Torn-paper face collage generated from the reverse-engineered image prompt

Result: a new collage portrait generated from the extracted prompt

Step 3: compare models with the same prompt

Reverse analysis solves how to describe an image; model comparison helps identify which model handles the task best. Keep the prompt and aspect ratio consistent for a useful comparison.

Purpose → Subject → Action or state → Environment → Composition and camera → Lighting → Color → Style or medium → Details → Constraints

This test generates an AI image-tool campaign poster. The prompt defines the desk scene, laptop position, left-side negative space, brand copy, and 16:9 ratio.

图片用途：面向创作者的 AI 图片工具宣传海报，适合网站首页和社交媒体分享。主体：画面中心是一台打开的笔记本电脑，屏幕显示由文字生成的彩色视觉作品，旁边摆放几张不同风格的打印图片和一支黑色记号笔。动作或状态：笔记本处于展示状态，打印图片自然散落在桌面上，画面传达正在进行创作的感觉。环境：简洁的创意工作室桌面，浅灰色背景，桌面上有纸张、色卡和少量设计工具，背景干净并留出排版空间。构图与视角：横版海报，三分法构图，笔记本位于画面右侧偏下，打印图片形成斜线引导，左侧保留大面积留白用于标题，轻微俯视视角，中近景。光影：柔和的侧上方自然光，物体边缘有细腻高光，阴影方向统一，屏幕带有轻微玻璃反光。色彩：米白、浅灰和深黑为基础色，加入蓝紫和珊瑚橙作为视觉强调色，中等饱和度，冷暖平衡。风格或媒介：高端科技品牌广告，照片级写实，干净的商业产品摄影，细节清晰，带有轻微胶片颗粒。细节：笔记本金属机身、纸张纤维、桌面细微纹理、屏幕玻璃反光和柔和接触阴影，主体清晰，背景轻微虚化。限制条件：左上角放置小型品牌文字“LNKLING IMAGE”，左侧主标题为“Create More With AI”，副标题为“Turn ideas into images”，右下角放置“Explore Image”，所有文字清晰可读，层级分明，不添加其他无关文字，16:9 横版。

AI image tool campaign poster generated by Seedream 5.0 Pro from the shared prompt — **Seedream 5.0 Pro**Rich desk detail, strong commercial photography, and clear type hierarchy.

AI image tool campaign poster generated by ChatGPT Image 2 from the shared prompt — **ChatGPT Image 2**Stable composition, generous left-side space, and natural prop placement.

AI image tool campaign poster generated by Imagen 3 from the shared prompt — **Imagen 3**A cleaner frame with a prominent product subject and different typography handling.

What do the three results reveal?

Even with identical wording, models interpret premium styling, natural placement, and negative space differently. Compare whether each result fulfills the intended purpose, composition, and typography—not only which looks prettiest.

How to improve reverse prompting

Check subject count, position, scale, and camera distance first.
Translate abstract style into observable lighting, material, grain, and color properties.
Record reference typography separately and manually verify important brand copy.
Fix the model, ratio, and resolution, then change only one or two variables per iteration.
Do not aim for pixel-perfect copying; extract the visual method and create something new.

The same method works for product images, posters, portraits, illustrations, and social covers. Once a reference becomes structured visual information, it can be edited, reused, and tested across models.