# AI Image Prompt Writing Guide

Different models interpret prompts in different ways. The gpt-image series, Flux series, and Seedream series each have their own emphases.

---

## gpt-image Series

> Applies to `gpt-image-1`, `gpt-image-2`, `gpt-image-1-mini`

### Basic Structure

Organize content in this order: **Background/Scene → Subject → Key Details → Constraints**

Both Chinese and English are supported, but English tends to produce more consistent results.

### Writing Principles

**Be specific, not abstract**

| Vague | Specific |
| --- | --- |
| `a woman` | `a 30-year-old woman in a red coat, looking at the camera` |
| `a city` | `Tokyo street at night, neon signs, rain-soaked pavement` |
| `nice lighting` | `soft golden hour sunlight from the left, warm tones` |

**Specify a style**

State the desired style directly:

```
photorealistic, professional photography, DSLR, shallow depth of field
```

```
children's book illustration, soft watercolor, pastel colors
```

**When text appears in the image**

Use quotation marks to indicate the exact text:

```
a product label with the text "Morning Blend" in elegant serif font, white background
```

**Start simple, iterate gradually**

Generate with a core description first, then refine with follow-up instructions:

> First pass: `a white ceramic coffee mug on a wooden table, studio lighting`
>
> Follow-up: `add steam rising from the mug, keep everything else the same`

**When editing an image, explicitly state what to preserve**

```
change the background to a snowy mountain, keep the person's face, clothing, and pose exactly the same
```

### Full Examples

**Realistic product photo:**
```
a minimalist white ceramic coffee mug centered on a light oak wooden table,
soft natural light from the left window, shallow depth of field,
photorealistic, studio photography, clean white background,
ultra-detailed, no shadows on background
```

![gpt-image coffee mug result](https://cdnv2-cache.udelivrs.com/2026/06/161d866d482f7666b9c27acf5f77bd2a_1781261163388.png)

**Portrait:**
```
a young Asian woman in a black trench coat standing on a rainy Tokyo street at night,
looking back over her shoulder, half body shot,
neon signs reflecting on wet pavement, rim lighting,
cinematic, film grain, muted colors, photorealistic
```

![gpt-image Tokyo portrait result](https://cdnv2-cache.udelivrs.com/2026/06/b9e99129c9281f68a2d692390a7d3eeb_1781261163406.png)

---

## Flux Series

> Applies to `flux-2-pro`, `flux-kontext-pro`, `flux-kontext-max`

`flux-2-pro` is the latest generation released by Black Forest Labs in November 2025, supporting up to 10 reference images, up to 4MP output, with noticeably improved text rendering and complex composition capabilities.

### Basic Structure

**Subject + Action + Style + Scene/Environment**

Word order matters — place the most important elements first, as the model prioritizes content at the beginning of the prompt.

### Prompt Length

| Length | Word Count | Best For |
| --- | --- | --- |
| Short | 10–30 words | Quick concept exploration |
| Medium | 30–80 words | Most use cases |
| Long | 80+ words | Complex multi-element compositions |

### Writing Principles

**Negative prompts are not supported**

Flux does not recognize "no xxx" — describe what you want instead:

| Incorrect | Correct |
| --- | --- |
| `no blur` | `sharp focus throughout` |
| `no noise` | `clean, smooth image` |
| `not dark` | `bright, well-lit` |

**Realistic photography style**

Specifying camera and film parameters works better than writing "realistic" generically:

```
shot on Sony A7IV, 85mm lens, sharp, high dynamic range, natural colors
```

```
shot on Kodak Portra 400, natural grain, warm tones, film photography
```

**Color control**

You can use hex color codes directly, but they must be tied to a specific object:

```
the car color is #FF0000, glossy paint
```

**Text and typography**

Wrap text content in quotation marks and specify its position and font style:

```
a poster with the text "Summer Sale" in bold sans-serif font, centered at the top, white on dark blue background
```

**Chinese prompts are supported**

Flux officially supports multiple languages; Chinese descriptions tend to be more accurate when depicting culturally specific scenes.

### Full Examples

**Realistic portrait:**
```
a young woman with short dark hair sitting in a cafe,
reading a book, soft afternoon sunlight through the window,
shot on Fujifilm X-T4, 35mm lens, warm tones,
shallow depth of field, sharp focus on face
```

![Flux cafe portrait result](https://cdnv2-cache.udelivrs.com/2026/06/6ace53efad79a290d555537183b8af11_1781261163340.png)

**Stylized illustration:**
```
a fox wearing a vintage explorer outfit standing on a mountain peak at sunset,
detailed digital illustration, rich colors, dramatic sky,
painterly style, highly detailed
```

![Flux fox illustration result](https://cdnv2-cache.udelivrs.com/2026/06/dd78d7ffd8715169c80989956538b3b0_1781261163361.png)

---

## Seedream Series

> Applies to `doubao-seedream-5-0-260128`, `doubao-seedream-4.5`

Made by Doubao, Chinese prompts work very well. Version 5.0 has deep reasoning capabilities and can understand descriptions of complex logical relationships.

### 5.0 vs 4.5

| Dimension | 4.5 | 5.0 |
| --- | --- | --- |
| Focus | Text rendering, image editing, production scenarios | Complex composition, reasoning, data visualization |
| Prompts | Concise, 2–4 sentences | Can be more detailed |
| Best for | Ad images, product photos, text layouts | Infographics, scientific diagrams, complex scenes |

### Writing Principles

**4.5: Keep it concise**

Describe the visual result directly, in 2–4 sentences:

```
An orange cat sitting on a windowsill, rainy street outside the window, soft natural light, realistic photography style
```

**5.0: Can be more detailed; describe logical relationships**

The reasoning capability benefits from complex instructions — suitable for describing relationships between elements:

```
An infographic showing the orbital relationships of the eight planets in the solar system,
planets arranged from nearest to farthest from the sun, each labeled with name and relative size,
deep blue space background, clean science-education style, with English annotations
```

![Seedream solar system infographic result](https://cdnv2-cache.udelivrs.com/2026/06/831d21a578538a105a60b4ee5d1de4f8_1781261163420.png)

**Text rendering is a strong suit**

Seedream excels at text rendering; Chinese text graphics, posters, and title images all look great:

```
一张简洁的活动海报，主标题"夏日音乐节"用大号黑体居中，
副标题"2026年8月15日·上海"在下方，
渐变橙色背景，现代设计风格
```

![Seedream summer poster result](https://cdnv2-cache.udelivrs.com/2026/06/188cce907e4655e1d9f7324f2a84aa1c_1781261163435.png)

---

## Which Model to Choose

| Use Case | Recommended Model |
| --- | --- |
| General use, Chinese prompts | `gpt-image-2` |
| Realistic portraits / photography style | `flux-2-pro` |
| Reference-image-based editing / style transfer | `flux-kontext-pro` |
| Highest image quality / complex layouts | `flux-kontext-max` |
| Chinese posters / text typography | `doubao-seedream-4.5` |
| Infographics / data visualization | `doubao-seedream-5-0-260128` |