We present Qwen-Image, an image generation foundation model in the Qwen series that achieves significant advances in complex text rendering and precise image editing. To address the challenges of complex text rendering, we design a comprehensive data pipeline that includes large-scale data collection, filtering, annotation, synthesis, and balancing. Moreover, we adopt a progressive training strategy that starts with non-text-to-text rendering, evolves from simple to complex textual inputs, and gradually scales up to paragraph-level descriptions. This curriculum learning approach substantially enhances the model's native text rendering capabilities. As a result, Qwen-Image not only performs exceptionally well in alphabetic languages such as English, but also achieves remarkable progress on more challenging logographic languages like Chinese. To enhance image editing consistency, we introduce an improved multi-task training paradigm that incorporates not only traditional text-to-image (T2I) and text-image-to-image (TI2I) tasks but also image-to-image (I2I) reconstruction, effectively aligning the latent representations between Qwen2.5-VL and MMDiT. Furthermore, we separately feed the original image into Qwen2.5-VL and the VAE encoder to obtain semantic and reconstructive representations, respectively. This dual-encoding mechanism enables the editing module to strike a balance between preserving semantic consistency and maintaining visual fidelity. Qwen-Image achieves state-of-the-art performance, demonstrating its strong capabilities in both image generation and editing across multiple benchmarks.
核心贡献 · Key contributions
提出 Qwen-Image，一个在复杂文本渲染和精确图像编辑上取得进展的基础模型。 Proposes Qwen-Image, a foundation model achieving advances in complex text rendering and precise image editing.
设计了全面的数据管道和从非文本到段落级描述的渐进训练，增强了字母和表意文字的文本渲染。 Designs a comprehensive data pipeline and progressive training from non-text to paragraph-level descriptions, enhancing text rendering for alphabetic and logographic languages.
引入改进的多任务训练范式，集成 T2I、TI2I 和 I2I 重建，对齐 Qwen2.5-VL 和 MMDiT 的潜在空间。 Introduces an improved multi-task training paradigm integrating T2I, TI2I, and I2I reconstruction, aligning latent spaces of Qwen2.5-VL and MMDiT.
采用双重编码机制（来自 Qwen2.5-VL 的语义和来自 VAE 的重建特征），在编辑中平衡语义一致性和视觉保真度。 Employs a dual-encoding mechanism (semantic from Qwen2.5-VL, reconstructive from VAE) for balanced semantic consistency and visual fidelity in editing.
在 GenEval、GEdit、ChineseWord 和 LongText-Bench 等多个基准上达到最先进，在文本渲染和编辑上超越现有模型。 Achieves state-of-the-art on multiple benchmarks including GenEval, GEdit, ChineseWord, and LongText-Bench, surpassing existing models in text rendering and editing.
开源模型和基础设施，包括用于可扩展训练的 Producer-Consumer 框架。 Open-sources the model and infrastructure, including a Producer-Consumer framework for scalable training.
局限 · Limitations
在 AI Arena 中落后于 Imagen4 等领先闭源模型约 30 Elo 分，表明通用图像质量仍有提升空间。 Trails leading closed-source models like Imagen4 by ~30 Elo points in AI Arena, indicating room for improvement in general image quality.
严重依赖合成数据和人工标注，可能引入偏见或限制对罕见真实场景的覆盖。 Relies heavily on synthetic data and human annotation, which may introduce biases or limit coverage of rare real-world scenarios.
训练和推理的高计算成本可能阻碍在资源受限环境中的部署。 High computational cost for training and inference may hinder deployment in resource-constrained environments.
图像编辑评估限于特定基准；对多样化真实编辑任务的泛化性尚未验证。 Evaluation on image editing is limited to specific benchmarks; generalization to diverse real-world editing tasks remains unverified.
未解决用于生成欺骗性或有害图像的潜在滥用，缺乏对齐保证或滥用缓解措施。 Potential misuse for generating deceptive or harmful images is not addressed, lacking alignment guarantees or misuse mitigation.