GPT Image 2 Prompt Formulas and Commercial Use Boundaries

GPT Image 2 (released 2026-04-21) replaced DALL·E 2/3 as OpenAI's only image generation model. This article covers the prompt formula, style controls, API parameters, and a pre-publish Usage Policy checklist.

1. The 5-part prompt formula

Describe the picture you want in five parts:

[Subject]       who or what
[Scene / Background]  where
[Light]         how it's lit
[Style]         artistic style
[Composition / Camera]  how it's framed

Example 1 (product shot):

"A deep-blue ceramic coffee mug, on a wooden table, soft side light from a morning window, natural photography style, medium shot, shallow depth of field."

Example 2 (marketing poster):

"A young woman standing on a neon-lit city street, giant billboards in the background, rainy night with reflections, cyberpunk style, wide-angle lens, high contrast."

One or two keywords per part — no adjective pile-up. "Very extremely particularly beautiful" loses to "minimal, elegant".

2. natural vs vivid styles

When in doubt, pick natural — it won't go wrong. vivid can over-saturate.

3. Images API parameters

from openai import OpenAI

client = OpenAI()

result = client.images.generate(
    model="gpt-image-2",
    prompt="...",
    size="1024x1024",       # 1024x1024 / 1024x1792 / 1792x1024
    quality="medium",       # low / / medium / / high / auto
    style="natural",        # natural / vivid
    n=1,                    # images per call; 2-4 is usually right
)

image_bytes = result.data[0].b64_json  # or .url

Parameter picks:

• size: square (1024×1024) is the social default; portrait (1024×1792) fits phone screens; landscape (1792×1024) for website banners
• quality: auto lets the model decide; explicit usage → medium (social) or high (print / large display)
• n: one call with n=4 is 4× cheaper than 4 calls — pricing is per call, not per image

4. Image editing (inpainting / outpainting)

Inpainting: pass original image + mask (transparent region = what to change) + prompt, the model only redraws the masked area.

from openai import OpenAI
import base64

client = OpenAI()

with open("original.png", "rb") as f:
    image_b64 = base64.b64encode(f.read()).decode()
with open("mask.png", "rb") as f:
    mask_b64 = base64.b64encode(f.read()).decode()

result = client.images.edit(
    model="gpt-image-2",
    image=image_b64,
    mask=mask_b64,
    prompt="a cat sitting on a chair",
    size="1024x1024",
)

Outpainting: extends the image beyond its current canvas — useful for turning a square into a banner, or adding borders.

Common uses: swap backgrounds, repair details, add elements, expand framing.

5. Commercial use boundaries (read the Usage Policy)

OpenAI allows commercial use of GPT Image 2 output, but with hard lines:

Pre-publish self-checklist:

• Subject is not on a celebrity/politician list
• No medical diagnostic implications
• Clearly labeled "AI-generated" (some jurisdictions require it)
• If mimicking a style, the original artist's work is still in copyright
• Prompt has no misleading elements

6. Common errors and troubleshooting

• Output doesn't match the prompt → prompt too abstract; rewrite with the 5-part formula
• Hands / eyes / text look off → GPT Image 2 is significantly better than DALL·E 3, but rare edge cases happen; switch style to natural or retry with a new prompt
• Style too over-the-top → switch style to natural or quality to medium
• A generated portrait looks like a real person → don't publish; it usually means the prompt included a real name
• Post-publish copyright warning → self-check the Usage Policy; if you mimicked a living artist's style, remove

7. What's Next

• GPT Image 2 complete guide: text-to-image, editing, style control
• OpenAI API Beginner: Your First GPT-5.6 Call Explained
• Function Calling with the OpenAI API: A Complete Guide to Tool Use in the Responses API