Images with the google-genai SDK

Three things: the base URL, the key, and the model name.

from google import genai
from google.genai import types

client = genai.Client(
    api_key=os.environ["TOKENIFY_API_KEY"],
    http_options=types.HttpOptions(base_url="https://api.tokenify.dev"),
)

res = client.models.generate_content(
    model="seedream-5.0-pro",
    contents="A still life of three pears, soft window light",
    config=types.GenerateContentConfig(
        response_modalities=["IMAGE"],
        image_config=types.ImageConfig(aspect_ratio="16:9", image_size="1K"),
    ),
)
for part in res.candidates[0].content.parts:
    if part.inline_data:
        open("out.jpg", "wb").write(part.inline_data.data)

import { GoogleGenAI } from "@google/genai"

const ai = new GoogleGenAI({
  apiKey: process.env.TOKENIFY_API_KEY,
  httpOptions: { baseUrl: "https://api.tokenify.dev" },
})

const res = await ai.models.generateContent({
  model: "seedream-5.0-pro",
  contents: "A still life of three pears, soft window light",
  config: { responseModalities: ["IMAGE"], imageConfig: { aspectRatio: "16:9", imageSize: "1K" } },
})

The SDK reads a base URL from GOOGLE_GEMINI_BASE_URL too, so an environment variable is enough if you would rather not touch the client. The key goes in x-goog-api-key, which is where the SDK puts it; ?key= and Authorization: Bearer are accepted as well, and a header wins over a query parameter.

The path puts the model before a colon, so it cannot contain a slash — use the plain name. Both of these spellings work for each model:

imageConfig.aspectRatio and imageConfig.imageSize become one exact width and height — exact because you named the shape, so the model is not left to choose a different one.

The ratios published for each tier on the image generation page resolve to exactly those sizes. The other ratios this SDK accepts — 1:4, 4:1, 1:8, 8:1, 4:5 and 5:4 — are computed: the tier’s own area at the ratio you asked for, each side rounded to a multiple of 16 (which is the grid every published size sits on), then clamped into the model’s pixel range. The reply tells you what you got, and the width and height we sent are what you are charged for.

A ratio outside the model’s range — beyond 1:16 either way — is a 400 rather than a squashed picture, and so is one that cannot be produced between 0.92 megapixels and 4.62 megapixels.

A content refusal comes back the way yours does: a 200 with finishReason: "IMAGE_SAFETY" and no image part. Nothing is charged for it.

Layer decomposition is not available through this shape — it has no equivalent concept here. Use either of the other two.

google.rpc.Status, which is what the SDK raises from:

{ "error": { "code": 400, "message": "…", "status": "INVALID_ARGUMENT" } }

UNAUTHENTICATED for a key problem, PERMISSION_DENIED for a key that may not use that model, NOT_FOUND for a model we do not serve, RESOURCE_EXHAUSTED for a rate limit or an empty balance, UNAVAILABLE and INTERNAL for our side.