Generate campaign visuals.
Let users draft banners, editorial illustrations or campaign concepts from a written brief. Save prompt variations alongside their images so an approved direction can be reproduced.
Creative tools / content workflowsQwen-Image-2.1
Generate and edit images.
Qwen-Image-2.1 is an image generation and editing model from Qwen. Run its Python pipeline on AI Fabric to create visuals, edit product photos and generate transparent assets for your application.
€5 in free credits for every new customer.
Plan your deploymentAI Fabric access is opening gradually. Bring your use case to onboarding.
Product imagery, visual variations
and transparent assets.
Turn a brief or reference photo into an asset users can review, edit and save. Keep image creation inside your product’s existing workflow.
Edit product imagery.
Offer a workflow for changing a background or exploring a new scene from a reference photo. Keep a human review step for product details, colours and logos.
Commerce / visual variantsCreate reusable assets.
Prepare stickers, interface illustrations and presentation elements. Transparent outputs can go directly into a design composition without a separate background-removal step.
Design tools / asset pipelinesWhat can Qwen-Image-2.1 do?
Qwen combines text-to-image generation and reference-based editing in one pipeline, with native transparency. Choose the operation that matches your product workflow.
Model capabilities: official Qwen model card, reviewed 12 October 2026. The 7B figure covers the visual generation component, not the full runtime memory footprint.
Generate from a written brief. Describe composition, lighting and visible text. Qwen reports improvements in typography and texture; check lettering and brand details before publication.
Edit with visual references. The publisher documents up to ten reference images and local edits guided by annotations or masks. Build an interface that makes the intended change explicit.
Work with transparent layers. Native RGBA output supports assets that can be composited into another design. Preserve the alpha channel when saving or serving the image.
Python → AI Fabric
How to run Qwen-Image-2.1
with Python.
Start with the model’s Diffusers pipeline, then package the working configuration for AI Fabric. Keep loading outside the per-request function so a warm worker can reuse the model.
- Install a compatible PyTorch, Transformers, Diffusers, Accelerate and Pillow environment following Qwen’s setup instructions. Pin the versions you validate.
- Test a representative prompt on your chosen GPU and record memory usage, latency and output quality.
- Bring the Python entry point, dependencies and input/output contract to AI Fabric onboarding.
The example shows model inference, not an AI Fabric deployment command. CPU offloading reduces GPU memory pressure but uses host RAM and can increase latency.
Prepare your deploymentimport torch
from diffusers import QwenImage21Pipeline
pipe = QwenImage21Pipeline.from_pretrained(
"Qwen/Qwen-Image-2.1",
torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()
def predict(prompt: str):
return pipe(
prompt=prompt,
width=2048,
height=2048,
num_inference_steps=40,
generator=torch.Generator("cuda").manual_seed(42),
).images[0]
if __name__ == "__main__":
image = predict("A paper sculpture of a solar-powered city")
image.save("result.png")Adapted from the publisher’s implementation. Validate this environment and image size on your worker before exposing the function to requests.
Choose the GPU. AI Fabric offers RTX A4500 (20 GB) and RTX 2000 Ada (16 GB). Confirm which memory and offload settings suit your workload during onboarding.
Design the response. For a website, keep credentials in your backend. For longer jobs, design a queued workflow with a status response, timeouts and a retrievable image file.
Price the workload. Confirm AI Fabric billing and configuration with the team. Measure cost per accepted image, including retries; Models token rates do not price this deployment.
Use Qwen-Image-2.1
with Claude or GPT.
Connect your deployed image service as a custom tool in an API-based assistant. The assistant writes the brief; your application sends the job to Qwen-Image-2.1.
01 / BriefClaude or GPT
Request an image through a defined tool.
02 / GenerateQwen-Image-2.1
Your backend runs the job on AI Fabric.
03 / DeliverYour application
Return the image URL or job status.
Define one focused image tool.
Expose a function such as generate_image with a prompt and an approved aspect ratio. Your handler validates the arguments, submits the job and returns its result to the assistant.
Keep credentials and request limits in your backend. Return an explicit failure if generation fails, and use job IDs to avoid duplicate work when retrying.
{
"name": "generate_image",
"arguments": {
"prompt": "An editorial illustration of a solar-powered city",
"aspect_ratio": "1:1"
}
}Illustrative tool arguments, not a Blade endpoint schema. Implement the handler and provider-specific tool declaration using OpenAI function calling or Claude tool use. This does not switch the assistant’s model or add Qwen to the standard ChatGPT or Claude image menu.
Qwen-Image-2.1 FAQ.
What is Qwen-Image-2.1?
It is Qwen’s model for image generation and editing, including transparent images. On this page, deployment means bringing its Python workload to AI Fabric, with your own application controlling inputs and outputs.
Can I call it through the Models API?
This guide covers an AI Fabric deployment. It does not announce Qwen-Image-2.1 as a managed Models API endpoint. For currently listed hosted models, see the Models catalogue.
How fast does it generate an image?
Blade has confirmed a successful internal run, but there is no published latency benchmark or reproducible configuration for this deployment yet. Benchmark your chosen resolution, inference settings and concurrency on the target GPU before setting a response-time promise.
Can I use it from a Claude or GPT assistant?
Yes, with a custom tool integration in your application. The assistant requests an image, your backend calls your deployed workload, and the result returns through the tool response. See the integration flow above.
How much does Qwen-Image-2.1 cost on AI Fabric?
Pricing depends on the GPU and commercial terms confirmed during onboarding. There is no published per-image rate for this deployment. Estimate cost from your measured generation time, retries and agreed billing unit; hosted Models pricing does not apply here.
What should I check before a production launch?
Validate output quality, memory, timeouts, access controls and your expected traffic. The publisher lists a Qwen Research License; review its licence information for your intended use. AI Fabric infrastructure access does not grant additional model rights.