Qwen-Image-Edit: Advanced AI for Semantic & Appearance Image Editing

In a significant advancement for multimodal artificial intelligence, Alibaba’s Qwen Team has unveiled Qwen-Image-Edit, an instruction-based image editing model that builds upon the robust 20-billion-parameter Qwen-Image foundation. Released in August 2025, this new iteration introduces sophisticated capabilities for both semantic and appearance editing, while retaining Qwen-Image’s notable strength in rendering complex text in both English and Chinese. Its integration with Qwen Chat and availability via Hugging Face aims to democratize professional content creation, from initial intellectual property design to intricate error correction in AI-generated artwork.

The technical backbone of Qwen-Image-Edit extends the Multimodal Diffusion Transformer (MMDiT) architecture. This framework incorporates a Qwen2.5-VL multimodal large language model (MLLM) for comprehensive text conditioning, a Variational AutoEncoder (VAE) for efficient image tokenization, and the MMDiT itself as the central processing unit for joint modeling. A key innovation for editing tasks is its dual encoding mechanism: an input image is simultaneously processed by the Qwen2.5-VL for high-level semantic understanding and by the VAE for capturing low-level reconstructive details. These distinct feature sets are then concatenated within the MMDiT’s image stream, enabling a delicate balance between maintaining semantic coherence—such as preserving object identity during a pose change—and ensuring visual fidelity, like leaving unmodified regions untouched. Further enhancing its adaptability, the Multimodal Scalable RoPE (MSRoPE) positional encoding has been augmented with a “frame dimension” to differentiate between pre- and post-edit images, a crucial feature for complex text-image-to-image (TI2I) editing tasks. The VAE, specifically fine-tuned on text-rich datasets, demonstrates superior reconstruction quality, achieving a Peak Signal-to-Noise Ratio (PSNR) of 33.42 on general images and an impressive 36.63 on text-heavy visuals, outperforming established models like FLUX-VAE and SD-3.5-VAE. These architectural refinements allow Qwen-Image-Edit to perform sophisticated bilingual text edits while meticulously preserving the original font, size, and style.

Qwen-Image-Edit excels in two primary domains of image manipulation. For appearance editing, it facilitates precise, low-level visual adjustments, enabling users to add, remove, or modify specific elements—such as realistically embedding signboards with reflections or subtly removing individual hair strands—without inadvertently altering surrounding regions. Concurrently, its semantic editing capabilities allow for high-level conceptual changes, supporting tasks like intellectual property creation, where a mascot can be adapted into various MBTI-themed emojis while maintaining character consistency. It can also perform advanced object rotation and style transfer, transforming a portrait into the distinctive aesthetic of a Studio Ghibli animation, all while ensuring semantic integrity and consistent pixel changes. A standout feature is its precise text editing, which supports both Chinese and English. Users can directly add, delete, or modify text within images, correcting errors in calligraphy via bounding boxes or changing words on a poster, always preserving the original typographical attributes. The model further supports “chained editing,” allowing for iterative corrections, such as step-by-step refinement of complex Chinese characters until perfect accuracy is achieved. Its ability to perform 180-degree novel view synthesis, rotating objects or entire scenes with high fidelity, is particularly noteworthy, achieving a PSNR of 15.11 on the GSO benchmark, a score that surpasses even specialized models like CRM.

The model’s robust performance is a direct result of an extensive training and data pipeline. Qwen-Image-Edit leverages Qwen-Image’s meticulously curated dataset, comprising billions of image-text pairs across diverse domains: Nature (55%), Design (27%), People (13%), and Synthetic (5%). It employs a multi-task training paradigm that unifies text-to-image (T2I), image-to-image (I2I), and text-image-to-image (TI2I) objectives. A rigorous seven-stage filtering pipeline refines this data for optimal quality and balance, incorporating innovative synthetic text rendering strategies (Pure, Compositional, Complex) to address long-tail issues prevalent in Chinese characters. The training process utilizes flow matching within a Producer-Consumer framework for scalability, followed by supervised fine-tuning and reinforcement learning techniques like DPO and GRPO to align the model with human preferences. For specific editing tasks, it integrates capabilities like novel view synthesis and depth estimation, employing DepthPro as a teacher model, which contributes to its strong performance in areas like correcting calligraphy errors through chained edits.

In benchmark evaluations, Qwen-Image-Edit has demonstrated state-of-the-art results across multiple public benchmarks for image editing. On GEdit-Bench-EN, it scored 7.56 overall, and on GEdit-Bench-CN, it achieved 7.52, outperforming competitors such as GPT Image 1 (7.53 EN, 7.30 CN) and FLUX.1 Kontext [Pro] (6.56 EN, 1.23 CN). Its performance on ImgEdit yielded an overall score of 4.27, with particular strengths in object replacement (4.66) and style changes (4.81). For depth estimation, it achieved an Absolute Relative error (AbsRel) of 0.078 on KITTI, a result competitive with leading models like DepthAnything v2. Human evaluations conducted on AI Arena further placed its base model third among available APIs, highlighting its superior instruction-following capabilities and multilingual fidelity, especially in text rendering.

For developers and creators, Qwen-Image-Edit is readily deployable via Hugging Face Diffusers, offering a streamlined integration process. Additionally, Alibaba Cloud’s Model Studio provides API access for scalable inference. Licensed under Apache 2.0, the training code is openly available on GitHub. This accessibility underscores a broader commitment to fostering innovation in AI-driven design. Qwen-Image-Edit represents a significant leap in vision-language interfaces, enabling more seamless and precise content manipulation for creators. Its unified approach to understanding and generating visual content suggests exciting potential for future extensions into video and 3D domains, promising to unlock new frontiers in AI-driven design applications.