Gemini Image Generation

Generate and edit images via Google Gemini's native multimodal image generation.

Model Selection

Model ID	Codename	Best for	Max resolution
`gemini-2.5-flash-image`	Nano Banana	Fast drafts, high-volume, low-latency	1K
`gemini-3-pro-image-preview`	Nano Banana Pro	Studio-quality, text rendering, complex prompts	4K

Default: gemini-3-pro-image-preview (Pro) unless speed/cost is a concern.

Setup

python

# Install (once)
# pip install google-genai

from google import genai
import os, base64

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])

If GEMINI_API_KEY is missing, instruct the user to set it as an environment variable. Never ask the user to paste the key in chat.

Text-to-Image

python

from google import genai
from google.genai import types
import os

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])

response = client.models.generate_content(
    model="gemini-3-pro-image-preview",
    contents="A photorealistic cat on a rainbow sofa",
    config=types.GenerateContentConfig(
        response_modalities=["TEXT", "IMAGE"],
    ),
)

# Extract and save
for part in response.candidates[0].content.parts:
    if part.inline_data is not None:
        with open("output.png", "wb") as f:
            f.write(part.inline_data.data)
        break

Aspect Ratio

Set via image_config:

python

config=types.GenerateContentConfig(
    response_modalities=["TEXT", "IMAGE"],
    image_config=types.ImageConfig(
        aspect_ratio="16:9",  # for slides / widescreen
    ),
)

Supported ratios: 1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9

Common choices:

•Slides / presentations → 16:9
•Social media / portraits → 9:16 or 4:5
•Square thumbnails → 1:1

Image Editing (with reference image)

python

from google.genai import types
from pathlib import Path
import base64

ref_bytes = Path("input.jpg").read_bytes()

response = client.models.generate_content(
    model="gemini-3-pro-image-preview",
    contents=[
        types.Part(inline_data=types.Blob(mime_type="image/jpeg", data=base64.b64encode(ref_bytes).decode())),
        types.Part(text="Remove the background and replace with a sunset gradient"),
    ],
    config=types.GenerateContentConfig(
        response_modalities=["TEXT", "IMAGE"],
    ),
)

Pro supports up to 14 reference images for multi-image composition and up to 5 human reference images for character/identity consistency.

Batch Generation (for slides)

When generating multiple images (e.g. one per slide), loop sequentially and save with numbered filenames:

python

import os, time
from google import genai
from google.genai import types

client = genai.Client(api_key=os.environ["GEMINI_API_KEY"])

prompts = [...]  # list of prompt strings

for i, prompt in enumerate(prompts, 1):
    response = client.models.generate_content(
        model="gemini-3-pro-image-preview",
        contents=prompt,
        config=types.GenerateContentConfig(
            response_modalities=["TEXT", "IMAGE"],
            image_config=types.ImageConfig(aspect_ratio="16:9"),
        ),
    )
    for part in response.candidates[0].content.parts:
        if part.inline_data is not None:
            with open(f"slide_{i}.png", "wb") as f:
                f.write(part.inline_data.data)
            break
    time.sleep(1)  # rate limit courtesy

Error Handling

•Safety filter block: The model may refuse prompts it deems unsafe. Adjust the prompt to be less ambiguous (remove violent/adult/medical imagery language) and retry.
•Empty response: If response.candidates is empty or has no image parts, the prompt may be too vague. Add concrete scene details and retry.
•Rate limit (429): Back off with exponential delay. Default: time.sleep(2 ** attempt).
•Timeout: Set a reasonable timeout; Pro model may take 10–30s for complex prompts.

Prompt Best Practices

•Structure: scene → subject → style → composition → constraints
•Always specify art style: "flat vector illustration", "watercolor painting", "3D render", "photorealistic photograph"
•Include lighting and mood: "soft diffused lighting", "dramatic rim light", "golden hour"
•For text in images: quote exact text, specify font style and placement
•For slide illustrations: add "negative space on [side]" to leave room for text overlay
•Use English prompts even for non-English content (better generation quality)
•Keep prompts under 500 words; be specific but not verbose

Style Consistency for Multi-Image Sets

When generating a series (e.g. slide deck), prepend a style prefix to every prompt:

code

Style prefix: "flat vector illustration, soft pastel color palette, clean lines, minimal detail, 16:9 widescreen"

Slide 1 prompt: "{style_prefix}, a wide establishing shot of a modern office building at sunrise"
Slide 2 prompt: "{style_prefix}, a close-up of hands typing on a laptop keyboard"

This ensures visual coherence across all generated images.