Qwen Image 2.1
Generate and edit images.
Create transparent layers, too.
Give one model text and reference images to compose scenes, make local changes and generate transparent PNGs, with a focus on preserving people and product details.
Open weights · Research license · Separate commercial permission
- Visual generation backbone
- 7Bparameters
- Reference images
- 10maximum
- Native transparency
- RGBAfour channels
- Native resolution
- 2Kofficially supported
Online demo
Try Qwen Image 2.1 here
Use the embedded demo to generate images or edit your references without leaving this page.
Powered by Qwen’s Hugging Face Space. Inputs and generation are handled by that service; its interface language, queue and availability are controlled by the Space.
Core capabilities
From reference images to editable assets
Official Qwen examples. Select an image to inspect the details.
Generate transparent PNGs directly
The model natively produces images with an alpha channel, edits transparent layers and extracts subjects from photos. Stickers, characters and design elements can be placed on other backgrounds.
Transparency is generated by the model. Switch the preview background to inspect transparent areas and edges; a white background alone is not transparency.
Official transparency examples
Compose a scene from several references
Supply up to 10 images for people, objects and settings. Refer to their input order to describe the role of each image.
Official example: six individual portraits combined into one group photo
Select a region and describe the change
Use circles, painted annotations or a separate mask to select regions, then describe the edit. Useful for hair color, clothing and individual objects.
State both what should change and what should remain. This gives a clearer boundary than asking to improve the whole image.
Official example: remove a watch and change hair color and clothing
Change the setting, preserve the product
Use the original product as a reference while adjusting the setting, composition and lighting. Check logos, packaging text and structure before using a product image.
Official example: a handheld product placed on a table by a window
Keep identity while changing the composition
Use a portrait to guide a new framing, outfit or pose. Identity preservation is a focus of the model, but exact consistency is not guaranteed in every result.
Official example: a close-up portrait becomes an outdoor full-body photo
Text and images in one layout
The model continues the Qwen-Image series’ focus on text rendering for posters, visual explanations and infographics. This official example combines Chinese text, architectural illustrations and several layout regions.
For factual content, provide the exact wording and check generated spelling, numbers and labels.
Official text-rendering examplesImage source: Qwen-Image-2.1 official showcase. Examples illustrate capabilities; results vary with the input.
Specifications and architecture
7B describes the visual backbone, not the whole pipeline
Text, reference images and transparency are processed in one generation pipeline.
| Component | Confirmed information | What it means |
|---|---|---|
| Visual backbone | About 7B · 32 layers | Single-stream block-causal DiT; 7B excludes the full pipeline. |
| Vision and text encoder | Qwen3-VL | Encodes text instructions and reference images together. |
| Image autoencoder | 64-channel RGBA VAE | 16× spatial compression with native transparency. |
| Reference images | Up to 10 | Input order corresponds to image 1, image 2 and so on. |
| Resolution | Native 2K support | Official sizes include 2048 × 2048; Diffusers defaults to a target edge of 1024. |
| Steps / CFG | 40 steps / 1.0 | Diffusers defaults; negative prompts require CFG greater than 1. |
| Prefix KV cache | Enabled by default | Reuses text and reference-image K/V across denoising steps. |
| Weight license | Research license | Commercial use requires separate permission. |
Verified on 2026-09-20. Native 2K support does not mean every tool defaults to 2K. Specifications and inference defaults: Official repository · Diffusers documentation
Getting started
Choose how to run the model
Start with the default settings, then adjust steps, resolution and memory management.
Embedded demo
Try reference fidelity and transparency in the demo above. Processing and queues are handled by Qwen’s Hugging Face Space.
ComfyUI workflow
Update ComfyUI, download the compatible model files and load an official text-to-image or image-editing workflow.
Python / Diffusers
Generate and edit through one pipeline. Suitable for batch processing, parameter experiments and existing image-processing workflows.
Minimal Diffusers exampleGeneration + multi-reference editingHide code
Use a compatible PyTorch / CUDA environment and a Diffusers main-branch version with QwenImage21Pipeline. This example uses CPU offloading and still needs sufficient system RAM and GPU memory. Official installation instructions
import torch
from PIL import Image
from diffusers import QwenImage21Pipeline
pipe = QwenImage21Pipeline.from_pretrained(
"Qwen/Qwen-Image-2.1", dtype=torch.bfloat16
)
pipe.enable_model_cpu_offload()
# Set dimensions explicitly to avoid differences between implementation defaults
result = pipe(
prompt="A capybara wearing an astronaut helmet, watercolor illustration",
width=1024, height=1024,
num_inference_steps=40,
true_cfg_scale=1.0,
).images[0]
result.save("result.png")
# Reference order must match the image list
edited = pipe(
prompt="Put the flowers from image 1 into the scene in image 2. Preserve the lighting of image 2.",
image=[Image.open("flowers.png"), Image.open("scene.png")],
num_inference_steps=40,
true_cfg_scale=1.0,
).images[0]
edited.save("edited.png")Negative prompts require true_cfg_scale > 1. Compile the model if using Flex Attention; the default attention implementation is a suitable starting point.
Write specific instructions
Image Toolkit wrote these templates. They are not the original prompts for the official images above.
Transparent sticker
Create a full-body capybara sticker wearing an astronaut helmet, with a clear outline. Output an RGBA image with an alpha channel and a fully transparent background. Do not include a floor, scenery or background shadows.
Product background
Place the product from image 1 on the table in image 2, matching the direction of light in image 2. Preserve the product’s shape, color, logo and packaging text. Change only its setting.
Multi-image outfit
Dress the person from image 1 in the jacket from image 2 and shoes from image 3, holding the bag from image 4. Preserve facial features and the colors and designs of the clothing, shoes and bag. Create a full-body photo.
FAQ
Before you start
Understand the model’s settings, outputs and usage terms.
Can I generate images on this page?
Yes, through the embedded Qwen Hugging Face Space above. Its inputs and generation are handled by that service. The embedded interface retains its own language and availability; Image Toolkit does not run this model’s inference.
Does 7B mean I only need 14 GB of GPU memory?
No. The 7B figure covers the visual backbone only. The full pipeline also includes Qwen3-VL, the RGBA VAE and intermediate data. Memory use depends on precision, resolution, reference count and offloading. CPU offloading adds system RAM use and waiting time.
How does transparency differ from a white background?
Transparent images contain an alpha channel and can be composited over another background. White backgrounds still contain opaque pixels. Ask for RGBA and a transparent background, then save as PNG. The three preview examples preserve their original alpha channels.
Is the default output 1024 or 2048?
The official repository describes native 2K support and recommended sizes such as 2048 × 2048. At verification, Diffusers used output_resolution=1024 by default. Specify width and height explicitly to distinguish model capability from tool defaults.
Can I use the model in a paid product?
Section 2 of the Qwen Research License Agreement limits the grant to non-commercial research or evaluation. Commercial use requires a separate license. Contact model-business@notice.qwencloud.com and follow the formal agreement and the permission you obtain.
Are 10–15 steps and 7 GB VRAM recommended settings?
No. Early community tests at low step counts do not replace the Diffusers default of 40 steps. Low-memory configurations depend on offloading and tiled decoding. Evaluate your own images and record peak memory, elapsed time and detail changes.