Image Generation

Generate images from text prompts using Nano Banana 2 Lite (Gemini 3.1 Flash Lite, default), Nano Banana 2 (Gemini 3.1 Flash), Go Bananas! Pro (Gemini 3 Pro Image), OpenAI GPT Image 2, or OpenAI GPT Image 2.5 (Flare and Sunburst).

Basic Generation

The simplest way to generate an image:

Generate an image of a mountain landscape at sunrise

This creates a single 1024x1024 image with default settings.

Parameters

Prompt (Required)

The text description of what you want to generate.

Tips for effective prompts:

Aspect	Example
Subject	"a red fox"
Action	"running through snow"
Setting	"in a winter forest"
Lighting	"golden hour sunlight"
Style	"oil painting style"
Mood	"peaceful, serene"

Good prompt example:

A red fox running through fresh snow in a winter pine forest,
golden hour sunlight filtering through the trees, photorealistic,
detailed fur texture, breath visible in cold air

Negative Prompt (Optional)

What to avoid in the image:

Generate a portrait photo of a business professional
Avoid: cartoon, anime, illustration, blurry, low quality

Common negative prompt additions:

• blurry, low quality, pixelated
• text, watermark, signature
• deformed, distorted
• multiple subjects (for single-subject images)

Aspect Ratio

Value	Dimensions	Best For
square	1:1	Social media, profiles
portrait	3:4	Character art, posters
landscape	4:3	Scenes, environments
16:9	Widescreen	Desktop wallpapers, videos
9:16	Vertical	Mobile, stories

Default: with no aspect ratio, the image is square (1:1). The request's own aspect_ratio always wins; if you leave it out, a default saved on a scene, style preset or character you use applies (checked in that order), and only then square. Edits keep the source image's shape unless you pass aspect_ratio or a scene, style preset or character you use supplies one.

Generate a cinematic landscape in 16:9 aspect ratio

Number of Images

Generate 1-4 variations in a single request:

Generate 4 variations of a fantasy castle

Model Selection

Default Model

Nano Banana 2 Lite (gemini-flash-lite-image) is the current default — fastest, lowest-cost, and optimized for 1K generation/editing. Pass model_id to override per request, subject to your tenant's allowed-models list.

Model	Provider	Use Case
gemini-flash-lite-image	Gemini 3.1 Flash Lite	Fastest, lowest cost, 1K output (default)
gemini-flash-image	Gemini 3.1 Flash	Strong all-around model, 2K/4K tiers, broader ratios
gemini-pro-image	Gemini 3 Pro	Best Gemini quality, Google Search grounding, 1K/2K/4K tiers
openai-gpt-image-2	OpenAI GPT Image 2	Text rendering, prompt adherence, quality tiers, JPEG/WebP output
openai-gpt-image-2.5-flare	OpenAI GPT Image 2.5 Flare	Fast everyday generation; xhigh/max quality, transparent backgrounds
openai-gpt-image-2.5-sunburst	OpenAI GPT Image 2.5 Sunburst	Precise generation and editing; same controls as Flare

model_id is validated against your tenant's allowedModels allowlist. Omit it to use the tenant default (Nano Banana 2 Lite); unknown or disallowed values are rejected. See the OpenAI gpt-image-2 deep-dive and the GPT Image 2.5 page for provider-specific parameters.

Generate a detailed architectural rendering

To use the general-purpose Flash model explicitly:

Generate a quick sketch using the Flash model

OpenAI-Specific Parameters

When model_id is an OpenAI model (openai-gpt-image-2, openai-gpt-image-2.5-flare or openai-gpt-image-2.5-sunburst), these extra parameters are honored:

Parameter	Values	Purpose
output_format	png, jpeg, webp	Native compression formats
output_compression	0–100	JPEG/WebP only; rejected for PNG
quality	low, medium, high, auto (+ xhigh, max on GPT Image 2.5)	Trade cost vs. detail
size	auto or WIDTHxHEIGHT	Custom GPT Image 2 size under OpenAI constraints
background	auto, opaque (+ transparent on GPT Image 2.5, png/webp only)	Transparent is not supported by GPT Image 2
moderation	auto, low	low is less restrictive

{
  "prompt": "Marketing poster with the headline 'Summer Sale' in a bold serif",
  "model_id": "openai-gpt-image-2",
  "aspect_ratio": "4:5",
  "size": "auto",
  "output_format": "webp",
  "output_compression": 85,
  "quality": "high"
}

System Instruction

Guide the model's overall approach:

System instruction: Professional product photography style
Generate a smartwatch on a marble surface

Reference Images

Use existing images as style or composition guides:

Generate an image in the style of [reference image URL]

Limits (total reference images per request):

• Gemini models (Nano Banana 2 Lite, Nano Banana 2, Go Bananas! Pro): Up to 14 reference images
• OpenAI GPT Image 2 and 2.5: Up to 16 reference images (routed through /v1/images/edits)

Character references are capped separately: when generating with a saved character, up to 4 of its reference images are used on the standard-tier models (Nano Banana 2 Lite and Nano Banana 2) and up to 5 on the pro-tier models (Go Bananas! Pro and the OpenAI models). With several characters, each contributes up to 3, within the same total.

Reference images must be at least 512x512 pixels for reliable character consistency.

Style Presets

Apply saved style templates:

Generate a coffee shop scene using the "Vintage Film" style preset

Learn about style presets →

Advanced Options (Pro Model)

Resolution Tiers

Tier	Quality	Speed
1024	Standard	Fastest
768	Medium	Medium
512	Lower	Fast

Google Search Grounding

Enable real-world accuracy for specific subjects:

Generate the Eiffel Tower at sunset with Google Search grounding enabled

Thinking Mode

Enable step-by-step reasoning for complex compositions:

Generate a detailed steampunk cityscape with thinking mode enabled

Output

Each generated image returns:

{
  "images": [
    {
      "id": 42,
      "full_url": "https://cdn.example.com/image.png",
      "thumbnail_url": "https://cdn.example.com/image-thumb.jpg",
      "width": 1024,
      "height": 1024,
      "size_bytes": 524288,
      "format": "png"
    }
  ],
  "session_id": "sess_abc123",
  "model": "gemini-flash-image"
}

Generation Workflow

Prompt → Style preset → References → Gemini or OpenAI API → R2 storage → Metadata

Tips for Better Results

Be Specific

❌ "a cat" ✅ "a fluffy orange tabby cat lounging on a velvet cushion, soft window light"

Use Art Terminology

• Photography: "shallow depth of field", "golden hour", "macro shot"
• Painting: "impasto technique", "chiaroscuro lighting", "watercolor wash"
• Digital: "3D render", "vector illustration", "low poly"

Iterate

1. Start with a basic prompt
2. Review the result
3. Add details to improve
4. Use editing for refinements

Combine Techniques

Generate a portrait in the style of [reference image],
with dramatic Rembrandt lighting, shallow depth of field,
using the "Cinematic" style preset

Combined Character & Product Generation

Generate images with multiple characters AND products in a single request for marketing and branding content.

Multi-Character Scenes

Generate scenes with up to 5 saved characters:

Generate an image of characters 19 and 20 having a team meeting in a modern office

The system automatically:

• Fetches character reference images from your library
• Builds optimal prompts for character consistency
• Maintains exact likeness across scenes

Character + Product Placement

Combine characters with product references:

Generate characters 19 and 20 at a product photoshoot with product 2

Example API call:

{
  "prompt": "at a professional photoshoot",
  "character_ids": [19, 20],
  "product_ids": [2],
  "aspect_ratio": "16:9"
}

Automatic Prompt Construction:

• Characters only: "The characters from the reference images {scene}"
• Products only: "Using the products shown, {scene}. Maintain exact design..."
• Both: "The characters with the products shown, {scene}. Maintain exact product design..."

Limits

Resource	Maximum
Characters per scene	5
Products per scene	6
Total reference images	14

WARNING

Cannot mix singular and array parameters (e.g., character_id with character_ids). Use one form consistently.

Common Use Cases

Product Shots

Professional product photography of a luxury watch on a marble
surface, soft studio lighting, reflection, minimalist background

Character Concept Art

Fantasy character concept art, elven ranger with silver hair,
green cloak, bow and arrows, forest background, digital painting style

Marketing Graphics

Modern social media banner for a tech startup, abstract geometric
shapes in blue and purple, clean minimalist design, 16:9 aspect ratio

Architectural Visualization

Photorealistic architectural render of a modern beach house,
floor-to-ceiling windows, infinity pool, sunset view, 4:3 landscape

Rate Limits & Quotas

• Rate limit: 60 requests/minute (configurable per tenant)
• Storage quota: Tracked per image stored
• Tokens: Prompt length counts against model limits

Next Steps

• Learn about editing images →
• Create consistent characters →
• Set up style presets →