Qwen-Image 2.0 API

Pricing

Cost = credits per image × n. Credits are reserved when you submit a job and refunded in full if it fails or you cancel it before it starts.

Sample outputs

Unretouched outputs from this model, generated through the API.

• 

  isometric data center clean vector style

At a glance

Quick start

Create a key in the dashboard, then submit a job and poll it. Replace the id in the second call with the one the first call returns.

curl -s -X POST https://api.genmux.tech/v1/generations \
  -H "Authorization: Bearer $API_KEY" \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{"model":"qwen-image-2","prompt":"a red fox in fresh snow, low winter sun","n":1,"resolution":"1K","ratio":"1:1"}'

{
  "id": "0d8b6a3e-2f1c-4a5b-9c1d-7e2f3a4b5c6d",
  "model": "qwen-image-2",
  "status": "queued",
  "params": {"prompt": "a red fox in fresh snow, low winter sun", "n": 1, "resolution": "1K", "ratio": "1:1"},
  "cost": 3,
  "created_at": "2026-08-30T20:01:12.345Z"
}

curl -s https://api.genmux.tech/v1/generations/0d8b6a3e-2f1c-4a5b-9c1d-7e2f3a4b5c6d \
  -H "Authorization: Bearer $API_KEY"

{
  "id": "0d8b6a3e-2f1c-4a5b-9c1d-7e2f3a4b5c6d",
  "model": "qwen-image-2",
  "status": "succeeded",
  "cost": 3,
  "images": [
    {"index": 0, "url": "https://…/0.png?X-Amz-Expires=900&…", "expires_at": "2026-08-30T20:16:30Z", "mime": "image/png", "width": 1024, "height": 1024, "bytes": 1843201, "sha256": "9f2c…"}
  ],
  "created_at": "2026-08-30T20:01:12.345Z",
  "finished_at": "2026-08-30T20:01:19.802Z"
}

import os, time, uuid, requests

BASE = "https://api.genmux.tech"
auth = {"Authorization": f"Bearer {os.environ['API_KEY']}"}

job = requests.post(
    f"{BASE}/v1/generations",
    headers={**auth, "Idempotency-Key": str(uuid.uuid4())},
    json={"model": "qwen-image-2", "prompt": "a red fox in fresh snow, low winter sun", "n": 1, "resolution": "1K", "ratio": "1:1"},
    timeout=30,
).json()

while job["status"] in ("queued", "running"):
    time.sleep(1)
    job = requests.get(f"{BASE}/v1/generations/{job['id']}", headers=auth, timeout=30).json()

print(job["images"][0]["url"])  # presigned, valid for 15 minutes

const BASE = "https://api.genmux.tech";
const auth = { Authorization: `Bearer ${process.env.API_KEY}` };

let job = await fetch(`${BASE}/v1/generations`, {
  method: "POST",
  headers: { ...auth, "Idempotency-Key": crypto.randomUUID(), "Content-Type": "application/json" },
  body: JSON.stringify({ model: "qwen-image-2", prompt: "a red fox in fresh snow, low winter sun", n: 1, resolution: "1K", ratio: "1:1" }),
}).then((r) => r.json());

while (job.status === "queued" || job.status === "running") {
  await new Promise((r) => setTimeout(r, 1000));
  job = await fetch(`${BASE}/v1/generations/${job.id}`, { headers: auth }).then((r) => r.json());
}

console.log(job.images[0].url); // presigned, valid for 15 minutes

Frequently asked questions

What is the Qwen-Image 2.0 API?

Qwen-Image 2.0 is Alibaba's image-generation model from the Qwen-Image line, a family the Qwen team introduced with an explicit focus on rendering legible text inside the picture — including Chinese characters — and on editing an image you supply. Through genmux API the Qwen-Image API is the model id "qwen-image-2": it renders at 1K, 2K, takes up to 4 reference images per request and costs 1K — 3 credits · 2K — 4 credits.

How much does one Qwen-Image 2.0 image cost?

From 3 credits per image, about $0.038 at the Scale pack rate. Price follows the resolution tier: 1K — 3 credits · 2K — 4 credits. A request for n images costs n times the per-image price. Credits are prepaid, so there is no subscription and no per-seat fee.

How is Qwen-Image 2.0 different from Qwen-Image Max?

Max is the flagship of the line and costs more per image here; 2.0 is the standard tier. This model renders at 1K, 2K and takes up to 4 reference images, where Max is restricted to a fixed, narrower set of aspect ratios. The request fields are otherwise the same, so comparing the two is a one-field change — put the same prompt and seed through both and read the two spec tables side by side.

When should I use Qwen-Image 2.0 instead of Nano Banana?

When the text in the image is the point: posters, packaging mockups, UI screenshots, slides, anything with a headline or CJK copy that has to come out readable. The Qwen-Image line was built around that case. Nano Banana is the general-purpose default — compare the two pricing tables and pick per job. Both accept reference images, share one API key and draw on the same credit balance.

What resolutions does Qwen-Image 2.0 support?

Resolution is a tier, not a pixel size: this model renders at 1K, 2K, and the request defaults to 1K when you omit the resolution field. A tier outside that list is rejected with a 400 validation_error rather than quietly downscaled, so you always know which price you are paying. The legacy size field ("1024x1024") still works for existing integrations and is mapped to the nearest tier and ratio.

What aspect ratios can I ask for?

1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 5:4, 4:5, 21:9. Pass it as the ratio field; it defaults to 1:1. A ratio outside the list is a 400 validation_error. Ratio never changes the price — only the resolution tier does.

Can I send reference images (image-to-image)?

Yes. Pass images as an array of public https URLs — up to 4 per request — and the prompt then describes the edit or the variation rather than the whole scene; with reference images the prompt itself is optional. URLs must be reachable from the public internet: private hosts and non-https URLs are rejected with a 400 validation_error.

Is there a negative prompt?

Yes. Pass negative_prompt alongside prompt to describe what you do not want in the frame — artefacts, text, watermarks, a colour you are steering away from. It costs nothing extra.

How many images can I generate in one request?

Up to 4 per call, via the n field. Each image is billed separately, so n=4 at 1K costs 4 times the per-image price. For larger batches submit several jobs — throughput is bounded by your account's in-flight job limit (20 by default), not by n.

Are generations reproducible? Does it support seeds?

Yes. Pass an integer seed alongside the prompt, and the same seed with the same prompt, resolution and ratio gives you a reproducible result — which is also how you compare two models fairly. Leave it out and each call is sampled independently.

How long does a generation take, and do I have to poll?

Typically 1–20 seconds from submit to succeeded, depending on the resolution tier and queue depth. The API is asynchronous: POST /v1/generations returns a job id in milliseconds, and from there you either poll GET /v1/generations/{id} or pass webhook_url and let us POST you once the job finishes, retrying with backoff for up to 8 attempts until your endpoint returns 2xx. Every delivery carries an HMAC-SHA256 signature you verify with your account webhook secret.

What happens if a generation fails?

You are not charged. Credits are reserved when the job is created and refunded in full if the job fails, is cancelled while still queued, or the model is temporarily unavailable (503 model_unavailable). The failed job carries an error object with a code such as content_policy so you can branch on it.

What stops me from being double-charged on a retry?

Every POST /v1/generations requires an Idempotency-Key header. Resending the same key returns the original job with 200 and charges nothing extra, so a network timeout can be retried safely. Reusing a key with a different body is rejected with 422 idempotency_key_mismatch.

Can I use the generated images commercially?

Yes — under our terms of service you own the rights we can pass on to the output of your generations and may use them commercially, provided the prompt and the use comply with the terms and the upstream model provider's content policy. Read the terms before shipping anything customer-facing.

Keep reading

• genmux MCP server: images from Claude, Cursor and OpenCode

  Install @genmux/mcp with one npx command and let Claude, Cursor, OpenCode or Continue list models, estimate cost and generate Nano Banana images.

  7 September 2026

• Rate limits and queueing for image generation pipelines

  Designing an image generation pipeline that survives traffic spikes: per-key RPS, max in-flight jobs, client-side queueing, backoff and capacity math.

  30 August 2026

• Nano Banana vs Nano Banana Pro: when 2× is worth it

  A practical comparison of the two Nano Banana models: quality, prompt adherence, supported resolutions, latency and cost — with concrete guidance by use case.

  30 August 2026

Other models

Same key, same request shape — switch by changing the model field.

• Nano Banana API

  nanobanana

  Fast, general-purpose text-to-image and image-to-image generation at 1K, behind a single HTTP endpoint. Billed per image in credits, refunded automatically when a job fails.

  from 2 credits ≈ $0.025 per image

• Nano Banana Pro API

  nanobanana-pro

  Higher-fidelity generation for hero shots and print-sized crops, with 1K, 2K and 4K output — roughly twice the base price per image at 1K.

  from 4 credits ≈ $0.05 per image

• Qwen-Image Max API

  qwen-image-max

  Qwen-Image Max — the flagship of Alibaba's Qwen-Image line, with reference-image input, 1K and 2K output and a fixed set of five aspect ratios.

  from 8 credits ≈ $0.10 per image

• Seedream 5.0 API

  seedream-5

  ByteDance Seedream 5.0 — a high-resolution model for photoreal frames and in-image text, served at 2K and 4K only. Takes reference images, billed per image in credits.

  from 16 credits ≈ $0.20 per image

• Wan 2.7 Image API

  wan-2-7-image

  wan2.7-image from Alibaba's Tongyi Wanxiang line — text-to-image and reference-image generation at 1K or 2K, on the same endpoint and API key as every other model here.

  from 2 credits ≈ $0.025 per image