Use Nano Banana 2 with Comfy Router

Python, TypeScript and cURL snippets for generating images with Nano Banana 2 (Gemini 3.1 Flash Image) over HTTP through Comfy Router, plus the request fields and the result shape

API Reference for Nano Banana 2. Nano Banana 2 (Gemini 3.1 Flash Image) is Google’s image generation and editing model, balancing quality and speed.

Comfy Router is not generally available yet. POST /v2/models/{provider}/{model} and its catalog and schema siblings are not serving requests yet: an authenticated call answers 404 today. The snippets on this page document the contract those routes will serve, published ahead of the rollout so your integration is ready to write against.

Quick start

Create a key at platform.comfy.org/profile/api-keys and export it as COMFY_API_KEY. The Python and TypeScript snippets use the Comfy SDKs (pip install comfy-sdk, npm install @comfyorg/sdk); the cURL snippet is the same call over raw HTTP.

Model ID: vertexai/gemini-3.1-flash-image

Endpoint: POST https://api.comfy.org/v2/models/vertexai/gemini-3.1-flash-image

Python

from comfy_sdk import Comfy

# Reads COMFY_API_KEY from the environment. Each call sends a fresh
# Idempotency-Key and waits up to 10 minutes for the finished result.
with Comfy() as client:
    result = client.models.run(
        "vertexai/gemini-3.1-flash-image",
        {
            "contents": [
                {
                    "role": "user",
                    "parts": [
                        {
                            "text": "a single red maple leaf on a plain white background, studio lighting",
                        },
                    ],
                },
            ],
            "generationConfig": {
                "responseModalities": ["IMAGE"],
                "imageConfig": {
                    "aspectRatio": "1:1",
                },
            },
        },
    )

print("image (base64):", result["candidates"][0]["content"]["parts"][0]["inlineData"]["data"])

Schema

Input

Fields follow Google’s published API specification and are checked against it in CI. Router’s own schema for this model is not published yet, so requests are forwarded to the provider unvalidated.

contents · object[] · required

The conversation so far. For a single image, one user turn with a text part; add an inlineData image part to edit an existing image.

contents[].role · string · required

Who authored the turn.

Possible values: user, model

contents[].parts · object[] · required

The turn’s content parts.

contents[].parts[].text · string

A text part.

contents[].parts[].inlineData · object

An inline media part.

contents[].parts[].inlineData.mimeType · string

Media type, for example image/png.

contents[].parts[].inlineData.data · string

Base64-encoded media bytes.

generationConfig · object

Generation settings.

generationConfig.responseModalities · `TEXT`, `IMAGE`[]

Ask for an image with ["IMAGE"], or ["TEXT", "IMAGE"] to also get a caption.

generationConfig.imageConfig · object

Image output settings.

generationConfig.imageConfig.aspectRatio · string

Output aspect ratio, for example 1:1, 16:9, 9:16.

generationConfig.seed · integer

Seed for reproducible results.

safetySettings · object[]

Per-category harm thresholds.

safetySettings[].category · string

Possible values: HARM_CATEGORY_SEXUALLY_EXPLICIT, HARM_CATEGORY_HATE_SPEECH, HARM_CATEGORY_HARASSMENT, HARM_CATEGORY_DANGEROUS_CONTENT

safetySettings[].threshold · string

Possible values: BLOCK_NONE, BLOCK_LOW_AND_ABOVE, BLOCK_MEDIUM_AND_ABOVE, BLOCK_ONLY_HIGH

Output

Router returns Google’s native response unchanged. The image (base64) is at candidates[0].content.parts[0].inlineData.data.

candidates · object[]

Generated candidates; one unless you asked for more.

candidates[].content · object

candidates[].content.role · string

Possible values: model

candidates[].content.parts · object[]

candidates[].content.parts[].inlineData · object

The generated image, inline.

candidates[].content.parts[].inlineData.mimeType · string

Media type of the image, typically image/png.

candidates[].content.parts[].inlineData.data · string

Base64-encoded image bytes. Decode and write to a file; there is no URL.

candidates[].content.parts[].text · string

Present when TEXT was among the requested modalities.

candidates[].finishReason · string

Why generation stopped, for example STOP.

usageMetadata · object

Token accounting for the call.

usageMetadata.promptTokenCount · integer

Tokens in the prompt.

usageMetadata.candidatesTokenCount · integer

Tokens in the generated candidates.

promptFeedback · object

Present when the prompt itself was blocked. It is the only field returned in that case, so check for it before reading the result.

promptFeedback.blockReason · string

Why the prompt was blocked. No candidates are returned; rephrase the prompt and retry.

Possible values: SAFETY, OTHER, BLOCKLIST, PROHIBITED_CONTENT, IMAGE_SAFETY

Examples

Input

{
  "contents": [
    {
      "role": "user",
      "parts": [
        {
          "text": "a single red maple leaf on a plain white background, studio lighting"
        }
      ]
    }
  ],
  "generationConfig": {
    "responseModalities": [
      "IMAGE"
    ],
    "imageConfig": {
      "aspectRatio": "1:1"
    }
  }
}

Output

{
  "candidates": [
    {
      "content": {
        "role": "model",
        "parts": [
          {
            "inlineData": {
              "mimeType": "image/png",
              "data": "iVBORw0KGgoAAAANSUhEUgAA..."
            }
          }
        ]
      },
      "finishReason": "STOP"
    }
  ],
  "usageMetadata": {
    "promptTokenCount": 12,
    "candidatesTokenCount": 1290
  }
}

The image comes back inline as base64 in inlineData.data, with its mimeType beside it. Decode it and write it to a file; there is no URL to download.

Before you ship

The snippets above are the shortest working call. Three things are the same for every model and are documented once on the Comfy Router headers page: send an Idempotency-Key on every paid call and reuse it when you retry, expect the connection to be held up to Router’s 10 minute deadline, and keep X-Comfy-Request-Id from every response. The SDKs do all three for you; the cURL tab does none of them. On failure, X-Comfy-Error-Type names the bucket, and a 422 means the body failed the model’s schema and was never billed.

Headers

Authentication, idempotency, request IDs, error buckets, retry pacing, spend limits.

Quickstart

Typed error handling in Python and TypeScript, reading the 422, walking the catalog.

Limitations

What Router does not do today, and what to use instead.