Nano Banana

Nano Banana is Google’s image generation and editing model. It is the model that powers the conversational, prompt-driven image editing in the Gemini app — you describe what you want changed in natural language, and the model edits the image while preserving everything else. On Prodia, both modes are exposed as flat-rate jobs that complete in around 8 seconds.

Architecture

Nano Banana is a multimodal model built on Google’s Gemini family. It accepts text and images in the same context window, which is what makes its editing behavior different from a typical diffusion img2img: rather than running noise through a pre-conditioned latent, the model reads the input image as visual tokens and reasons about which regions the prompt describes. In practice this means edits stay tightly localized — change “the puppy’s collar” and the rest of the scene, lighting, and pose stay frozen.

The text-to-image variant uses the same model with no input images, generating output from prompt and aspect ratio alone.

When to use Nano Banana

• Localized edits to existing images — adding, removing, or modifying a specific element while keeping the rest of the image unchanged. Identity preservation, prop swaps, expression edits, clothing changes
• Multi-image composition — img2img.v2 accepts up to 3 input images, useful for combining a subject from one photo with a setting from another
• Conversational prompt style — the model responds well to natural-language instructions (“add a red bow tie”, “make it nighttime”) rather than the keyword-heavy prompts used for diffusion models
• Flat per-job pricing — every job is $0.039 regardless of resolution or aspect ratio, so cost is predictable

For photorealistic generation with style presets and high-resolution output up to 4096px, FLUX.2 is a stronger choice. For native text rendering inside images, see Recraft V4. For higher-fidelity Google models with selectable resolution, see the Gemini 3 Pro and Gemini 3.1 Flash job types listed below.

Job types

Job type	Description	ETA
inference.nano-banana.txt2img.v2	Generate an image from a text prompt	~8s
inference.nano-banana.img2img.v2	Edit one or more input images (up to 3) with a text prompt	~8s
inference.nano-banana.img2img.v1	Single-image editing (deprecated — use v2)	~8s

The nano-banana processor on Prodia also serves Google’s higher-tier Gemini image models with selectable resolution up to 4K:

Job type	Description	ETA
inference.gemini-3-pro.txt2img.v1	Gemini 3 Pro text-to-image	~10s
inference.gemini-3-pro.img2img.v1	Gemini 3 Pro image-to-image (up to 3 inputs)	~12s
inference.gemini-3-1-flash.txt2img.v1	Gemini 3.1 Flash text-to-image, optional Google Search grounding	~30s
inference.gemini-3-1-flash.img2img.v1	Gemini 3.1 Flash image-to-image (up to 14 inputs)	~35s

Note

The deprecated inference.nano-banana.img2img.v1 accepts a single image filename and is kept for backwards compatibility. New integrations should use img2img.v2, which accepts an images array and exposes the aspect_ratio parameter.

Parameters

inference.nano-banana.txt2img.v2:

• prompt (required) — text description of the desired output, up to 2,500 characters
• aspect_ratio — one of 1:1 (default), 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9
• include_messages — when true, the response also returns message.txt parts containing the model’s natural-language reasoning

inference.nano-banana.img2img.v2:

• prompt (required) — describe the edit you want, up to 2,500 characters
• images — array of 1–3 input image filenames sent as multipart input parts
• aspect_ratio — one of auto (default — match the first input), 1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, 21:9
• include_messages — see above

Prompting tips

• Describe the change, not the whole scene. For img2img edits, write what should be different (“add a red bow tie”) rather than re-describing the entire image — the model already sees the input
• Anchor preservation explicitly. Phrases like “keep everything else exactly the same” reduce drift in untouched regions, especially for subtle edits
• Use natural language. Nano Banana is a multimodal LLM, not a CLIP-conditioned diffusion model. Conversational instructions outperform comma-separated keyword prompts
• Reference inputs by position. When passing multiple images to img2img.v2, refer to them in order — “the subject from the first image, in the setting from the second image”
• Pick aspect_ratio deliberately for txt2img. The default is 1:1. For social, web, or phone use cases, set 9:16 or 16:9 rather than upscaling/cropping after the fact

Examples

Text-to-image at 16:9:

{
  "type": "inference.nano-banana.txt2img.v2",
  "config": {
    "prompt": "a cute corgi puppy in a sunny meadow with wildflowers, soft natural light, photorealistic",
    "aspect_ratio": "16:9"
  }
}

Image-to-image edit (using the previous output as input):

{
  "type": "inference.nano-banana.img2img.v2",
  "config": {
    "prompt": "Add a small red bow tie to the puppy. Keep everything else exactly the same.",
    "aspect_ratio": "16:9"
  }
}

Notice how the puppy’s pose, fur, the wildflowers, the lighting, and the background are all preserved — only the bow tie is added.

Multi-image composition (up to 3 inputs):

{
  "type": "inference.nano-banana.img2img.v2",
  "config": {
    "prompt": "Place the product from the first image onto the wooden surface in the second image, matching the warm lighting.",
    "images": ["product.jpg", "surface.jpg"],
    "aspect_ratio": "3:2"
  }
}

Guides

Generating Images Step-by-step guide for generating images with Prodia.

Transforming Images Guide for image-to-image transformation.