GPT Image 2 — Specific Rules

The skill Copy

Production default OpenAI: `gpt-image-2`. Migration-only: `gpt-image-1.5`, `gpt-image-1`. Бюджетный: `gpt-image-1-mini`.

## Структура промпта — 5 slots

GPT Image 2 сильнее всего реагирует на разделение по секциям. Пиши лейблами, не сплошным текстом.

```
Scene: location, time of day, background, environment
Subject: primary focus — who or what is central
Important Details: materials, textures, lighting, camera angle, mood, composition
Use Case: editorial, product mockup, UI, poster, infographic
Constraints: what must NOT change/appear (no watermarks, preserve face, no extra text)
```

> «The fifth slot is where most mediocre prompts fail silently.» Без явных constraints модель дрейфует.

### Минимальный пример

```
Scene: small Lisbon florist storefront at blue hour, wet cobblestones
Subject: woman in navy apron locking the front door, half-turned to camera
Important Details: warm interior glow spilling onto pavement, 50mm feel,
  soft contact shadows, brushed brass door handle, "Florista" hand-painted sign
Use Case: editorial photography
Constraints: no extra signage, no people in background, no text other than the sign
```

## Anti-Slop Rules

GPT Image 2 особенно чувствителен к качеству формулировок. Vague praise = деградация результата.

| ❌ Не пиши | ✅ Пиши |
|-----------|---------|
| stunning, incredible, epic, gorgeous, masterpiece | overcast daylight, brushed aluminum, chipped paint, 50mm feel |
| «minimalist brutalist luxury photoreal» (стиль-теги) | «cream background, heavy black sans-serif, asymmetrical type block, one hero object, generous negative space» |
| «как в Apple-рекламе» | конкретные визуальные факты |
| мудовый язык, в котором тонут функциональные требования | прямое заявление: «image must contain a transit kiosk» |

## Quality Settings — рычаг fidelity / latency

| Setting | Когда |
|---------|-------|
| `quality: low` | High-volume, превью, exploratory, latency-sensitive, draft |
| `quality: medium` | **Default starting point** |
| `quality: high` | Маленький/плотный текст, infographics, портреты, identity-sensitive edits, brand assets |

Стартуй с `low`, апгрейди по необходимости. Часто `low` уже достаточно.

## Размеры (gpt-image-2)

- Max edge: <3840px
- Обе стороны: кратны 16
- Aspect ratio: max 3:1 (long:short)
- Total pixels: 655 360 – 8 294 400
- Reliable upper bound: 2560×1440

**Ходовые:**
- Portrait 1024×1536
- Landscape 1536×1024
- Square 1024×1024
- 2K 2560×1440

> Экстрим вроде 1:8 / 8:1 GPT Image 2 НЕ умеет — иди в Nano Banana.

## Text in Image

- Литеральный текст в `"..."` или ALL CAPS.
- Шрифт, размер, цвет, позиция — явно.
- Сложные слова и бренды: спеллинг по буквам.
- Защита от мусора: «**no extra words, no duplicate text, no watermarks**».
- Для нечитаемого мелкого текста: «**100 percent readable and physically believable**».
- Маленький/плотный/multi-font → `quality: high` обязательно.

## Editing — двухколоночная логика

Endpoint: `openai/gpt-image-2/edit` (на fal.ai) или соответствующий через OpenAI API.

```
Change: [single concrete change]
Preserve: face, identity, pose, lighting, framing, background, geometry, text, layout
Constraints: no extra objects, no redesign, no drift
```

**Правила edit:**
- **Один edit за итерацию.** Не пытайся менять всё разом.
- **Preserve list повторять каждую итерацию.** Иначе дрейф.
- **Surgical edits:** явно перечисли что НЕ трогать (saturation, contrast, layout, arrows, labels, camera angle).
- Опционально: `mask_image_url` для точечных edits.

### Edit-паттерны

**Virtual try-on:** «Change garments only. Preserve exact face, body shape, pose, hair, expression, background, camera angle. Match lighting/shadows so outfit looks naturally worn.»

**Object removal:** «Remove [X]. Do not change anything else. Use `input_fidelity: high` to maintain surrounding context» (только gpt-image-1.5/1, в gpt-image-2 high-fidelity по умолчанию).

**Lighting/weather swap:** «Change ONLY environmental conditions: lighting direction/quality, shadows, atmosphere, precipitation. Preserve identity, geometry, camera angle, object placement.»

**Interior swap:** «Swap [furniture]. Preserve camera angle, lighting, shadows, surrounding context. Photorealistic contact shadows.»

## Multi-Image — до 16 рефов

Индексируй с **ролью**, не только номером:
```
Image 1: base scene
Image 2: jacket reference (apply only the jacket fabric/cut to subject in Image 1)
Image 3: lighting reference (apply golden-hour quality from Image 3)
```

## Style Transfer

Не пиши абстрактно («minimalist», «editorial»). Назови конкретные визуальные свойства референса: палитра, edge treatment, силуэт, обработка теней, plane логика.

## World Knowledge

GPT Image 2 умеет домысливать контекст: «Bethel, NY, August 1969» → выведет Woodstock-эстетику. Используй: дай исторический/культурный анкер, не расписывай каждую деталь.

## Iteration Strategy

- Стартуй с **чистого** базового промпта.
- Один change за раунд. «Make lighting warmer», «remove extra tree», «restore original background».
- При drift — перечисли invariants заново.
- Для длинных промптов — labeled sections, не одна простыня.

## Use-Case Templates

### Photoreal Editorial
```
Scene: [location, time, weather]
Subject: [who, action, framing]
Important Details: [lens feel, light source, surface wear, imperfections, real texture]
Use Case: editorial photograph, looks like a real photo
Constraints: no glamorization, no heavy retouching, no studio gloss
```

### Product Mockup (Clean Background)
```
Scene: plain white opaque background
Subject: [product] centered
Important Details: crisp silhouette, no halos/fringing, light contact shadow,
  preserve label legibility exactly, preserve geometry
Use Case: product mockup
Constraints: no restyling, only background removal + light polish
```

### UI Mockup
```
Scene: [device frame, e.g. iPhone 15 Pro]
Subject: [screen/app name] — describe AS IF IT EXISTS, not concept art
Important Details: layout, hierarchy, real interface elements, exact copy in quotes,
  typography behavior, spacing, state
Use Case: shipped product screenshot
Constraints: no sketch language, no placeholder text, no Lorem Ipsum
Quality: high (for small UI text)
```

### Marketing Creative with Text
```
Scene: [environment]
Subject: [hero element]
Important Details: [composition, palette, mood]
Use Case: ad creative for [audience]
Text: "EXACT HEADLINE" in [font style], [color], [position]
      "exact subhead" in [font style], [color], [position]
Constraints: no extra text, no duplicate text, no watermarks, no unrelated logos
Quality: high
```

### Infographic / Diagram
```
Title: "[TITLE]"
Content flow: [step 1] → [step 2] → [step 3]
Visual format: [layout type — flowchart, pyramid, isometric, etc.]
Use Case: educational infographic for [audience]
Constraints: readable labels at all sizes, clear hierarchy, no clutter,
  no decorative noise, ample whitespace
Quality: high
Size: 1536×1024
```

## Migration from Older GPT-Image

- Промпты в основном переносятся как есть.
- После переноса — посмотри качество, latency, retry-rate; ретюнь.
- `gpt-image-1-mini` — только если главное снизить цену batch'а на низкорисковых задачах.

---

*Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*

How to install it Copy

Gemini

1. Open Gems → New Gem.
2. Paste the text below into the instructions field.
3. Save. The Gem appears in the sidebar and keeps the instructions
   across conversations.

In Google AI Studio the same text goes in System instructions.

Did it work for you?

Where it came from