Qwen-Image 2.1 brings transparent images and multi-reference editing to ComfyUI

01The background is optional

Qwen released Qwen-Image 2.1 on 20 September, bringing image generation and editing into one model. Its standout addition is native RGBA output: the image can include an alpha channel, which controls transparency, rather than a permanently painted background.

The model can generate transparent assets, edit them and extract subjects. That changes the useful output from a finished rectangle to something a designer can place inside a poster, interface or motion composition. Those are possible applications, not results from a XANTRANS test.

02Bring the references into the edit

Qwen documents support for up to ten reference images, alongside local editing guided by painted marks, circles or separate masks. It also supports native 2K generation.

For an experimental collage, a reference could define the subject while another suggests the setting. A marked area can specify where the change belongs. The practical question is whether the result preserves the details that matter; this release has not been independently tested by XANTRANS.

03Three ComfyUI templates to start with

ComfyUI provides separate templates for text-to-image generation, image editing and background removal. Its model repository includes the diffusion model, a Qwen3-VL text encoder and a VAE, the component that converts between image pixels and the model’s internal representation.

BF16 and compressed INT8 ConvRot model options are listed. The text encoder also has a W4A8 option. Pick the files specified by the chosen template; do not assume every variant has the same memory requirements.

An optional prompt-enhancement model uses ComfyUI’s TextGenerate node. It is a separate step from the normal text encoder, even though both sets of files are stored in the text_encoders folder.

04Downloadable does not mean commercially unrestricted

The published Qwen Research License limits use of the model materials to research and evaluation unless a separate commercial licence is obtained. That is a material restriction for anyone building a client workflow or a paid product around the weights.

Local execution still needs suitable hardware and setup. This article does not promise a minimum VRAM figure or a particular generation speed. Start with the official template and licence before budgeting a production workflow around it.