Skip to content
Search

Qwen-Image-Edit

lastmod 2025-08-22created 2025-08-222 edits

Release of a dedicated image editing model

About

  • Qwen-Image-Edit is a dedicated image editing extension model based on the Qwen-Image model
  • Its architecture feeds the input image to both Qwen2.5-VL and the VAE Encoder at once, supporting both semantic and appearance editing
  • Strong text editing: can directly edit Chinese and English text while preserving font, size, and style
  • Achieves state-of-the-art (SOTA) performance on various benchmarks, establishing itself as a strong foundation model for image editing
  • Released as open source under the Apache 2.0 license, so developers and researchers can use it freely

Introduction

  • Qwen-Image-Edit is a model that extends Qwen-Image's text rendering capability to support image editing
  • Its architecture feeds the image input to both semantic control (Qwen2.5-VL) and appearance control (VAE Encoder) at once
  • Supports both precise text editing and semantic/appearance editing

Key features

  • Semantic & appearance editing: supports semantic changes such as adding, removing, and rotating objects and style transfer, as well as appearance editing that changes only specific regions
  • Precise text editing: can directly edit English and Chinese while preserving the original font and style
  • Performance lead: achieves state-of-the-art performance on multiple public benchmarks

Quick start

  • Available through the Hugging Face diffusers library
  • The example code changes the rabbit's color to purple and changes the background to a flash light
  • Runs efficiently with CUDA acceleration and torch.bfloat16 support

Showcase

  • Semantic editing: character IP creation, object rotation (90 and 180 degrees), and style transfer (e.g., Ghibli style)
  • Appearance editing: precisely adds signboards, removes hair, changes the color of specific text, replaces backgrounds, changes clothing, and more
  • Text editing: accurately edits both large and small text on English and Chinese posters
  • Chained editing: demonstrates fixing character errors in a calligraphy work step by step to finally produce a complete version

Application scenarios

  • Brand IP extension: a case of creating MBTI emoticons based on a Capybara character is introduced
  • Art and creation: various style transfers of portraits open up the possibility of creating virtual avatars
  • Industrial use: supports fine-grained editing, such as naturally generating even reflection effects when inserting signs

License

See also

Favorite site

author Zer0 sitemap