Qwen-Image-2.1 Released: Open-Weight Image Gen & Editing
Sample outputs are available on the model card.
At a Glance
| Item | Value |
|---|---|
| Repository | Qwen/Qwen-Image-2.1-PE-I2I |
| Published | 2026-09-20 |
| License | other |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code at collection time. Dates are JST.
Sample outputs are available on the model card.
Overview
The Qwen team has released “Qwen-Image-2.1," a new open-weight model that integrates text-to-image generation and image editing. This model combines a Text-to-Image function for generating high-quality images from text and an Image-Editing function for editing existing images according to instructions into a single model.
Along with this, dedicated prompt rewriting models—"Qwen-Image-2.1-PE-T2I" and “Qwen-Image-2.1-PE-I2I"—have also been released as open source to rewrite short instructions and vague requests from users into detailed prompts that image generation and editing models can easily understand. These are fine-tuned models based on Qwen3.5-VL 9B, responsible for expanding short requests entered in any language into detailed English prompts and outputting them in JSON format along with recommendation information such as the optimal aspect ratio.
Specifications
The main specifications of the released model group and the base “Qwen-Image-2.1" are as follows.
Qwen-Image-2.1 (Image Generation and Editing Model)
- Number of Parameters: 7B parameters in the visual generation component
- Architecture: 32-layer single-stream DiT (Diffusion Transformer)
- Key Features:
- Native Transparency (RGBA Support): In addition to standard text-to-image generation, it can handle generating RGBA images with transparent backgrounds, editing transparent layers, and extracting subjects from photos (background removal) with a single model. – Advanced Image Editing: Supports up to 10 reference images. It supports local editing (partial editing) specifications using circles, painted annotations, or individual masks, enabling editing while preserving the identity (features) of people or products. – Efficiency Technologies: Achieves high image quality at low computational cost through a lightweight architecture adopting mixed-granularity attention and prefix KV cache reuse.
Qwen-Image-2.1-PE-T2I (Prompt Rewriting Model for Text Generation)
- Base Architecture: Qwen3.5-VL 9B (fine-tuned version)
- Input: Short image generation requests in any language
- Output: Outputs the following JSON format after the reasoning process via `")
result = json.loads(answer.strip())
Output example: {“rewritten_prompt": “…", “wh_ratio": “", “ratio_follow": ““}
“`
ratio_follow included in this output indicates inheriting the aspect ratio of the editing source image as-is (performing in-place editing). By passing the obtained rewritten_prompt along with the original image input_image to QwenImage21Pipeline, an image with the specified edits applied is output.
In addition, when inputting multiple images, you can place multiple image objects within the list of messages and specify <image1> or <image2> in the text instructions to have the PE model interpret complex instructions targeting specific images (e.g., “place the subject of <image1> into the background of <image2>“).