大模型
www.aibase.com ·
通义千问开源 Qwen-Image-2.1：7B 参数实现文生图与图像编辑一体化
据 AIbase 报道，Qwen 团队开源新一代图像模型 Qwen-Image-2.1，将文生图、透明图生成与多种图像编辑整合到单一框架。其视觉生成部分采用 32 层 Single-Stream DiT 结构，仅 7B 参数，通过混合粒度注意力和 KV Cache 复用提升多图输入场景的推理效率并控制显存。模型可根据提示自动输出普通图或带透明通道的 RGBA 图，支持在保留透明背景的前提下修改主体表情和文字，也可从真实照片中自动抠出主体生成透明层。图像编辑支持最多 10 张参考图输入，局部编辑提供框选、涂抹和独立 mask 三种方式，并称在人脸与商品一致性、文字渲染、人物光照和细节渲染上有改进。模型已在 GitHub、ModelScope 和 Hugging Face 上开放免费使用。上述能力与效果均来自该报道，尚无官方模型卡或评测数据佐证。
AI推荐理由 输入称 Qwen 团队开源 Qwen-Image-2.1，给出 7B 参数、32 层 Single-Stream DiT、混合粒度注意力与 KV Cache 复用等具体架构信息，并列出透明图生成、最多 10 张参考图输入、局部编辑三种方式等能力，对文生图与图像编辑开发者有实际参考价值。但来源为 AIbase 转述，非官方一手发布，正文未提供模型卡、评测数据或代码链接，'显著提升''高度真实'等效果描述缺乏可验证证据，需以官方仓库为准。
78
正文
Qwen-Image-2.1:7B Open Source - Achieving Text-to-Image and Image Editing in One
AIbase基地
Published inAI News · 4 min read · Sep 21, 2026
31.2k
On September 21, the Qwen team officially open-sourced the new generation image model Qwen-Image-2.1. This model integrates text-to-image generation, transparent image generation, and various image editing capabilities into a single framework, achieving a new balance between generation quality, inference efficiency, and usage cost with its lightweight visual generation module of only 7B parameters.
The visual generation part of Qwen-Image-2.1 adopts a 32-layer Single-Stream DiT structure with only 7B parameters. Through a mixed-granularity attention structure and KV Cache reuse mechanism, the model significantly improves inference efficiency in multi-image input scenarios and effectively controls memory consumption.
Transparent image generation is one of the core highlights of this update. The model can automatically determine whether to output a regular image or an RGBA image with a transparent channel based on the prompt, and supports operations such as modifying the subject's expression and editing text while keeping the transparent background. Users can also input real photos, and the model will automatically extract the subject and output a transparent layer, facilitating subsequent design and compositing.
You can modify the subject's expression while keeping the transparent background
Text within the transparent layer can also be edited, as shown in the example below, where the text "BLOOM" is changed to "Qwen-Image":
In terms of image editing, Qwen-Image-2.1 supports up to 10 reference images as input, enabling the generation of harmonious scenes by integrating multiple subjects and materials. Local editing supports three methods: selection, brushing, and independent masks, allowing more flexible specification of editing areas. The model has also significantly improved the consistency of human faces and products—after editing, facial features of people and the text and shape of products remain highly realistic.
In addition, the model has made significant improvements in text rendering, lighting on people, and detail rendering, covering various task scenarios such as panoramic images, infographics, and storyboards. Currently, Qwen-Image-2.1 is available for free use by developers and creators on platforms such as GitHub, ModelScope, and Hugging Face.
This article is from AIbase Daily
Scan to view
Welcome to the [AI Daily] column! This is your daily guide to exploring the world of artificial intelligence. Every day, we present you with hot topics in the AI field, focusing on developers, helping you understand technical trends, and learning about innovative AI product applications.
—— Created by the AIbase Daily Team
© Copyright AIbase Base 2024, Click to View Source -