10 Common Prompt Mistakes for AI Image Generation and How to Fix Them

Version used: gpt-image-2 (released April 2026)

01 | Vague Description

Bad example: 一个好看的咖啡馆场景 Good example:

俯拍视角，一张白色大理石圆桌，桌上摆着一杯拿铁拉花咖啡和一本翻开的书，背景是落地窗，下午三点的暖光从右侧斜射进来，摄影风格，浅景深

Root cause: The term “好看” provides no actionable information, causing the model to generate a generic, mediocre image.

Fix template:

[拍摄视角] + [主体描述] + [场景/背景] + [光线方向和类型] + [画面风格] + [景深/构图]

Example:

近景平视，一位穿白色衬衫的年轻女性坐在咖啡馆靠窗位置，桌上有笔记本电脑，窗外是模糊的街道，柔和的侧逆光，纪实摄影风格，浅景深

02 | Changing Too Many Variables at Once

Bad example: Change background to forest, protagonist to a boy, color to warm, and add a cat in a single request.

Good example: Modify only one variable per round, e.g., first change the background, then the protagonist, then the color.

Root cause: Simultaneous changes make the model compromise on all variables, and you cannot pinpoint which step caused an issue.

Fix template:

轮次1：把[元素A]从[旧值]改为[新值]，其他不变
轮次2：在上图基础上，把[元素B]从[旧值]改为[新值]，保持[A不变]
轮次3：...（每轮只动一个）

03 | Ignoring Aspect Ratio

Bad example: Generate a horizontal poster with the default 1024x1024 (square).

Good example: Specify size="1536x1024" for a horizontal layout or describe the desired composition in the prompt.

Root cause: Mismatched image ratio forces post‑generation cropping, breaking composition.

Fix template: Use the appropriate size for the target scenario.

公众号封面 – recommended size 1536x1024 (3:2) – 1536x1024 朋友圈竖图 – recommended size 1024x1536 (2:3) – 1024x1536 头像/图标 – square – 1024x1024 手机壁纸 – recommended size 1024x1792 (9:16) – 1024x1792 Also add a prompt note, e.g., “横版构图，左三分之一放主体，右侧留白用于放文字，画面比例 3:2”。

04 | Stacking Conflicting Style Words

Bad example:

赛博朋克风格，水彩画风格，极简主义，复古胶片，未来感，手绘插画风

Good example: Choose one primary style and up to two complementary modifiers.

Root cause: The model tries to fuse contradictory directions, resulting in an image with no clear style.

Fix template: [主风格]，[辅助修饰1]，[辅助修饰2] Recommended style combinations (excerpt):

主风格: 赛博朋克 – 辅助: 霓虹光效，雨夜 – 不推荐: 水彩、极简、手绘

主风格: 水彩插画 – 辅助: 柔和色调，纸质感 – 不推荐: 3D渲染、赛博朋克

主风格: 极简主义 – 辅助: 留白，几何，单色 – 不推荐: 复古、繁复花纹

主风格: 摄影写实 – 辅助: 自然光，浅景深 – 不推荐: 卡通、手绘线稿

05 | Prompt Length Too Long or Too Short

Bad (too long): 600‑word description with 20 detail requests.

Bad (too short): 一个人在城市里 (10 characters).

Good range: 80‑200 Chinese characters (or 50‑150 English words).

Root cause:

Too short: insufficient information, model fills randomly.

Too long: exceeds the model’s effective context window; later instructions receive little weight.

Fix template: Split the prompt into layers.

[核心层，20-40字] 主体 + 场景 + 风格
[细节层，30-80字] 光线 + 颜色 + 材质 + 构图
[约束层，10-30字] 不要什么 + 必须保留什么

Keep total length 100‑150 words; excess details should be moved to API parameters.

06 | Ignoring Lighting Description

Bad example: 一位女性站在户外，背景是树林 (no lighting info).

Good example:

一位女性站在户外树林边缘，黄昏时分，背后是逆光的橙色夕阳，轮廓光打亮发丝，面部补光，柔和阴影

Root cause: Lighting defines image texture; without it the model defaults to flat, generic lighting.

Fix template: Add scene‑specific lighting terms.

产品图 – 柔和漫射光，无阴影，纯白背景 人像 – 蝴蝶光，侧光，伦勃朗光，逆光轮廓 风景 – 黄金时刻，蓝调时刻，正午硬光 室内 –

窗口自然光，暖色台灯，霓虹灯环境光

07 | Not Using Negative Constraints

Bad example: 一张简洁的产品海报，手机放在白色桌面上 (no “不要” clause).

Good example: Append “不要 X，不要 Y，不要 Z” to the prompt.

Root cause: The model adds common but unwanted elements when not explicitly prohibited.

Fix template:

[正向描述]。不要[元素1]，不要[元素2]，不要[元素3]，背景保持干净

Typical exclusion lists (excerpt):

产品图 – 多余道具、阴影过重、水印、反光

人像 – 多余人物、变形手指、奇怪背景

UI截图 – 多余弹窗、中文乱码、不一致字体

Logo – 多余装饰线、阴影、背景纹理

08 | Not Locking Invariant Elements During Iteration

Bad example: “把杯子颜色改成红色” without stating other elements stay the same, causing the whole scene to change.

Good example: “只把杯子颜色从白色改成深红色，桌面、背景、光线方向、其他所有元素严格保持不变”。

Root cause: GPT‑Image‑2’s local editing works only when unchanged parts are explicitly locked; otherwise the model performs global optimization.

Fix template:

在上图基础上，只修改[目标元素]的[具体属性]：
- 改变：[目标元素] 从 [原状态] 改为 [新状态]
- 严格保持不变：[锁定元素1]、[锁定元素2]、[锁定元素3]
- 光影方向、整体构图、色调完全不变

09 | Not Defining Character Traits in Multi‑Image Tasks

Bad example: Generating a comic series with only “a young girl”; each image shows a different face.

Good example: Include a fixed character description in every prompt.

Root cause: GPT‑Image‑2 lacks a persistent character ID; without repeated description it generates random appearances.

Fix template (character card):

# 角色固定描述（每张图都要包含）
角色：25岁亚洲女性，长直黑发，杏仁眼，圆脸，高挑身材，常穿白色T恤配牛仔裤

# 场景描述（每张图可以不同）
场景：[这张图的具体场景]

Store the character card as a variable and concatenate it into each prompt, e.g.:

CHARACTER = "25岁亚洲女性，长直黑发，杏仁眼，圆脸，高挑身材，常穿白色T恤配牛仔裤"
for scene in scenes:
    prompt = f"角色：{CHARACTER}。场景：{scene}"
    # call API

10 | Skipping Post‑Generation Quality Checks

Bad example: Look at the result once and either accept it or randomly tweak the prompt.

Good example: Use a standardized checklist to inspect each output and target fixes.

Root cause: Without a checklist you cannot locate deficiencies, leading to inefficient trial‑and‑error.

Quality‑check checklist:

□ 主体是否清晰，没有变形？
□ 光线方向是否一致，无矛盾光源？
□ 构图是否符合预期（画幅比例、主体位置）？
□ 风格是否统一，没有混搭冲突？
□ 细节（手指、文字、logo）是否正常？
□ 有没有出现不想要的元素？
□ 整体色调是否和使用场景匹配？

Each checklist item maps to a specific prompt‑repair direction.