How to Write Better Prompts for GPT Image 2 (With Real Examples)

A strong GPT Image 2 prompt describes the subject clearly, specifies a visual style, names the lighting conditions, and includes any composition or camera details. Vague prompts produce generic results; specific prompts produce exactly what you intend. This guide breaks down each element with real examples you can copy and adapt.

GPT Image 2 is built on a language model that interprets your prompt as a set of instructions — not as keyword soup. It understands spatial relationships ('in the foreground', 'behind the subject'), relative importance ('focus on the face'), stylistic modifiers ('Baroque oil painting', 'Unreal Engine 5 render'), and negations ('no text', 'no watermarks').

Unlike older models where more keywords meant better results, GPT Image 2 responds better to natural language descriptions. Think of writing a prompt as briefing a professional illustrator: describe what you want, not just what elements should appear.

1. Subject: Who or what is the main focus? Be specific. Instead of 'a woman', write 'a 30-year-old Japanese woman in a tailored navy suit, standing in a glass-walled office'.

2. Style: What visual aesthetic should the image have? Options include photorealistic, cinematic, flat illustration, isometric, watercolor, 3D render, anime, editorial photography, and many more.

3. Lighting: How is the scene lit? Cinematic lighting, golden hour, studio lighting, soft diffuse light, neon backlight, and high-contrast rim lighting all produce dramatically different results.

4. Composition: Where should elements be placed? Describe camera angle (eye-level, bird's eye view, low angle), framing (close-up portrait, wide shot, over-the-shoulder), and depth of field (bokeh, f/2.8 aperture, everything in sharp focus).

5. Technical details: For maximum realism, add camera or render specifications: 'shot on Sony A7 IV, 85mm f/1.4', 'Octane render', '8K resolution', 'RAW photo quality'.

Weak prompt: 'a coffee shop' → Generic, low-detail result.

Strong prompt: 'A cozy Parisian coffee shop interior at dawn, warm amber light filtering through lace curtains onto marble tabletops, two espresso cups with steam rising, shallow depth of field, shot on Leica M10, cinematic color grading, no people' → Specific, art-directed result.

Weak prompt: 'product label for tea' → The model guesses the style and content.

Strong prompt: 'A premium loose-leaf tea canister label in the style of Japanese woodblock prints (ukiyo-e), featuring a stylized cherry blossom with Mount Fuji in the background. Text reads "Sakura Sencha" in both English and Japanese kanji. Deep indigo and gold color palette. High-resolution, ready for offset printing.' → Precise commercial asset.

GPT Image 2's text rendering is one of its defining strengths. To get accurate text, quote the exact words you want in your prompt: 'a billboard reading "Summer Sale — 50% Off" in bold red letters on white'. Keep text short and specific — fewer than 20 words per element renders most reliably.

For multi-language text, specify the script: 'the sign reads "咖啡" in traditional Chinese characters, written in a brushwork calligraphy style'.

Avoid asking the model to render body copy or long paragraphs. Use it for headlines, labels, signs, and short UI copy where legibility is critical.

For product photography: always specify the background color or texture ('on a matte white surface with soft studio lighting'), the product angle ('45-degree three-quarter view'), and whether you want reflections ('with a subtle floor reflection').

For character consistency across multiple images: establish a detailed character description in a reference prompt first, then reference those exact physical details in subsequent prompts. GPT Image 2 does not have a native 'style reference' memory across sessions.

For the best results with inpainting: upload a high-quality source image and use a precise mask. Describe what should replace the masked area as if starting from scratch — the model reads your text description, not the surrounding context.