Imagifly — Qwen Image Edit

Qwen-Image-Edit is the image-editing model the Qwen team open-sourced on 18 August 2025 on top of Qwen-Image. It extends Qwen-Image's bilingual text rendering from generation to editing, so a single instruction can change the picture or rewrite the words inside it.

  • Image Edit

From 2 credits per image

Image editing workflow preview for Qwen Image Edit
Sample output: Image editing workflow preview for Qwen Image Edit

About Qwen Image Edit

Qwen-Image-Edit is the image-editing model the Qwen team open-sourced on 18 August 2025 on top of Qwen-Image. It extends Qwen-Image's bilingual text rendering from generation to editing, so a single instruction can change the picture or rewrite the words inside it.

The backbone is the same 20B MMDiT used by Qwen-Image; the official model card lists 20.43B parameters in BF16, identical across releases. Each input image goes to Qwen2.5-VL for semantic control and to the VAE encoder for appearance, so edits stay local instead of repainting the whole frame.

The line advanced twice: Qwen-Image-Edit-2509 (22 September 2025) added multi-image inputs of one to three references plus native ControlNet support for depth, edge and keypoint maps, and Qwen-Image-Edit-2511 (23 December 2025) reduced drift after repeated edits, improved multi-person consistency and shipped built-in community LoRAs.

What it is good at

  • Officially an image-to-image editing model, not text-to-image
  • Rewrites Chinese and English text inside the image, the Qwen-Image signature
  • Multi-image references and ControlNet control since 2509
  • Weights released under Apache-2.0

Technical specs

Developer
Qwen Team, Alibaba
Released
2025-08-18
Latest version
Qwen-Image-Edit-2511
License
Apache-2.0
Architecture
20B MMDiT (Qwen2.5-VL semantic control + VAE appearance encoding)
Parameters
20.43B (BF16)
Task
Image editing (image-to-image)
Recommended settings
Documented defaults: 50 steps and true_cfg_scale 4.0 for the first release; 40 steps, true_cfg_scale 4.0 and guidance_scale 1.0 for 2509 / 2511. No recommended output resolution is published.

Source: Qwen-Image-Edit model card