Source-linked AI summary
Qwen-Image Technical Report
Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Xu, Yilei Chen, Yuxiang Chen, Zecheng Tang, Zekai Zhang, Zhengyi Wang, An Yang, Bowen Yu, Chen Cheng, Dayiheng Liu, Deqing Li, Hang Zhang, Hao Meng, Hu Wei, Jingyuan Ni, Kai Chen, Kuan Cao, Liang Peng, Lin Qu, Minggang Wu, Peng Wang, Shuting Yu, Tingkun Wen, Wensen Feng, Xiaoxiao Xu, Yi Wang, Yichang Zhang, Yongqiang Zhu, Yujia Wu, Yuxuan Cai, Zenan Liu
TL;DR
Qwen-Image addresses difficult text rendering and precise image editing, where complex prompts and faithful edits remain challenging. It combines engineered data, progressive training, multimodal objectives, and dual image representations, achieving state-of-the-art performance across generation and editing benchmarks.
Problem
Complex prompt alignment and precise image editing remain challenging, particularly for multi-line or Chinese text and preservation of unedited visual details.
Method
Qwen-Image combines comprehensive data engineering and curriculum learning with T2I, I2I, and TI2I training using Qwen2.5-VL semantic and VAE reconstructive representations.
Results
Qwen-Image achieves state-of-the-art performance across image generation and editing tasks, with strong consistency reported for novel-view synthesis and pose editing.
Takeaways & Limitations
Qwen-Image strengthens visual generation as a complement to Qwen2.5-VL’s visual understanding within multimodal foundation models.
Takeaways & Limitations
A qualitative comparison uses watermarked SeedEdit 3.0 images obtained through its web interface because no official API is provided.
Abstract
from arXiv · showhide
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.
1 Introduction
Qwen-Image targets complex text rendering and precise image editing through data engineering, progressive learning, and enhanced multimodal training. It reports strong generation and editing performance across benchmarks.
- Complex prompt alignment remains difficult, especially for multi-line text, Chinese rendering, localized insertions, and integrated text–visual compositions.
- Qwen-Image combines comprehensive data engineering, progressive curriculum learning, enhanced multi-task training, and infrastructure optimization.
- Its data pipeline progresses from basic text rendering to paragraph-level and layout-sensitive descriptions while improving multilingual following, particularly for Chinese.
- The editing framework integrates T2I, I2I, and TI2I objectives using semantic Qwen-VL features and reconstructive VAE features in a shared latent space.
- Qwen-Image delivers complex text rendering across English and Chinese and preserves semantic meaning and visual realism during editing.
- Across multiple benchmarks, Qwen-Image consistently outperforms existing models on diverse generation and editing tasks.
2 Model
The model combines Qwen2.5-VL, a VAE, and MMDiT, with MSRoPE providing joint image–text positional encoding. Its design supports multimodal conditioning and high-fidelity visual generation.
- Architecture: Qwen-Image uses an MLLM for text conditioning, a VAE for image tokenization, and MMDiT as the diffusion backbone.
- Multimodal Large Language Model: Qwen2.5-VL provides aligned language–visual features, retains language modeling ability, and accepts multimodal inputs for image editing.
- VAE: The decoder is trained on text-rich documents and synthetic paragraphs in English and Chinese to improve reconstruction of small text and fine details.
- Multimodal Diffusion Transformer: MSRoPE starts image encoding at the image center and places textual encodings along the grid diagonal to improve resolution scaling and text–image alignment.
3 Data
Qwen-Image’s data strategy prioritizes quality, balanced distributions, text coverage, and synthetic supplementation. A seven-stage filtering pipeline progressively refines data during training.
- Data Collection: The dataset contains Nature, Design, People, and Synthetic Data domains, collected and annotated at billion-scale with emphasis on quality and balance.
- Data Collection: Nature comprises approximately 55% of the dataset, while Design comprises around 27% and focuses on text-rich layouts and artistic content.
- Data Collection: People contributes about 13% of the dataset and covers portraits, sports, and human activities.
- Data Collection: Synthetic Data accounts for approximately 5% and uses controlled text-rendering synthesis rather than images generated by other AI models.
- Data Filtering: The seven-stage filtering pipeline progressively refines data distributions, introducing synthetic data from Stage 4 after foundational training stabilizes.
- Data Synthesis: Text-aware synthesis combines Pure, Compositional, and Complex Rendering to address scarce and imbalanced textual content, especially rare Chinese characters.
4 Training
Qwen-Image uses flow matching for pre-training and combines progressive data curricula with distributed infrastructure for scalable, stable optimization. Training progresses across resolution, text, quality, balance, and synthetic-data dimensions.
- Pre-training: Flow matching constructs intermediate latents and velocities from encoded image latents, sampled Gaussian noise, conditioning features, and a timestep.
- Pre-training: The model predicts the target velocity using an MSE loss between its output and the ground-truth velocity.
- Infrastructure: A Producer–Consumer framework decouples preprocessing from model training, enabling asynchronous operation and on-the-fly pipeline updates.
- Infrastructure: Megatron-LM, hybrid data and tensor parallelism, distributed optimizers, and activation-checkpointing experiments address large-model memory and scaling constraints.
- Training Curriculum: The curriculum increases resolution from 256×256 to 640×640 and 1328×1328 while progressing from non-text to text, massive to refined data, and unbalanced to balanced distributions.
4.2 Post-training
Qwen-Image’s post-training combines supervised fine-tuning with preference-based reinforcement learning. DPO handles large-scale offline preference learning, while GRPO provides smaller fine-grained refinement.
- Post-training consists of supervised fine-tuning followed by reinforcement learning.The supervised stage uses a hierarchically organized semantic dataset and human annotation.
- DPO is used for computationally efficient, large-scale offline preference learning, while GRPO is reserved for small fine-grained refinement.DPO targets one-step flow-matching preference modeling; GRPO performs on-policy sampling with reward-model evaluation.
- DPO training generates multiple images per prompt with different random seeds and uses human-selected best and worst samples.For prompts with reference images, annotators compare outputs against the reference; otherwise they rank generated samples or mark all unsatisfactory.
- The DPO objective uses preference differences from policy and reference models, scaled by β and passed through a sigmoid.The objective is built on the flow-matching training criterion.
(B) Group Relative Policy Optimization (GRPO)
GRPO refines the flow model using group-relative rewards and stochastic trajectory sampling. It reformulates deterministic flow-matching sampling as an SDE and discretizes the process with Euler–Maruyama.
- (B) Group Relative Policy Optimization (GRPO): GRPO computes within-group advantages from reward-model scores for groups of generated images.The advantage compares each reward with the mean reward across the group.
- (B) Group Relative Policy Optimization (GRPO): Because deterministic flow-matching sampling is unsuitable for exploration, GRPO reformulates trajectories as an SDE with added randomness.The randomness magnitude is denoted by σ_t.
- (B) Group Relative Policy Optimization (GRPO): Euler–Maruyama discretization is used to obtain the practical trajectory-sampling update.The resulting equation is used to sample trajectories during training.
4.3 Multi-task training
Qwen-Image extends text-to-image generation to multimodal image tasks by combining semantic visual features with VAE latents. This design targets both editing consistency and visual fidelity.
- 4.3 Multi-task training: The base model supports multimodal tasks including instruction-based editing, novel view synthesis, and depth estimation.User-image visual patches are encoded by a Vision Transformer and incorporated into the model.
- 4.3 Multi-task training: The model concatenates VAE-encoded input latents with noised image latents to help preserve character and scene consistency.The VAE representation is added along the sequence dimension of the image stream.
- 4.3 Multi-task training: Pixel-level VAE embeddings further enhance visual fidelity and structural consistency with the user-provided image.
5.1 Human Evaluation
AI Arena evaluates image-generation models through anonymous pairwise human comparisons and Elo ratings. Qwen-Image ranks third among the evaluated closed-source competitors and exceeds several leading APIs.
- 5.1 Human Evaluation: AI Arena randomly pairs model outputs for the same prompt and uses user votes to update Elo leaderboards.The platform is designed as an open, dynamic competition for holistic model assessment.
- 5.1 Human Evaluation: The arena uses about 5,000 diverse prompts and compares Qwen-Image with five state-of-the-art closed-source APIs.Each competitor had participated in at least 10,000 pairwise comparisons.
- 5.1 Human Evaluation: Qwen-Image ranks third, trails Imagen 4 Ultra Preview 0606 by approximately 30 Elo points, and leads GPT Image 1 [High] and FLUX.1 Kontext [Pro] by over 30 Elo points.
5.2 Quantitative Results
Qwen-Image is evaluated across reconstruction, text-to-image generation, text rendering, and image editing benchmarks. Results show strong prompt adherence, multilingual text rendering, and competitive or leading editing performance.
- Evaluation Setup: VAE reconstruction is evaluated as an upper-bound indicator of generation quality before assessing T2I and TI2I capabilities.The tokenizer comparison uses PSNR and SSIM with a shared 8x8 compression rate and 16 latent channels.
- Text-to-Image Generation: Qwen-Image achieves the highest overall score on DPG, indicating strong adherence to dense prompts and attributes.DPG contains 1K dense prompts designed for fine-grained prompt-adherence assessment.
- Text-to-Image Generation: 0.91 is Qwen-Image’s GenEval score after reinforcement-learning fine-tuning, making it the only foundation model above 0.9 on the leaderboard.The base model already surpasses Seedream 3.0 and GPT Image 1 [High].
- Text-to-Image Generation: Qwen-Image ranks first overall on both Chinese and English OneIG-Bench tracks, while ranking second on TIIF Bench behind GPT Image 1.OneIG-Bench averages scores across five dimensions; TIIF Bench evaluates intricate textual instruction following.
- Text Rendering: Qwen-Image shows strong text rendering across English, Chinese, and long-text benchmarks, including the highest accuracy on long Chinese text.CVTG-2K finds performance comparable to leading models for English rendering, while LongText-Bench reports second-highest accuracy on long English text.
- Image Editing: Qwen-Image ranks first on both English and Chinese GEdit leaderboards and highest overall on ImgEdit, demonstrating strong multilingual editing performance.GEdit covers 11 instruction categories, while ImgEdit covers nine tasks and 734 real-world test cases.
- Image Editing: Qwen-Image achieves state-of-the-art novel view synthesis among the reported baselines.The evaluation uses prompts that instruct view changes, such as rotating a dog 90 degrees.
5.3 Qualitative Results
Qualitative comparisons examine reconstruction, text-to-image generation, and image editing across text, layout, object, pose, chained-editing, and viewpoint tasks. Qwen-Image consistently preserves requested content, visual structure, and fine details in challenging examples.
- VAE Reconstruction: Qwen-Image-VAE reconstructs small text in dense images more precisely than the other compared VAEs.In a PDF example, “double-aspect” remains legible in Qwen-Image-VAE’s reconstruction but not in the others.
- English Text Rendering: Qwen-Image renders long English paragraphs, multiple text locations, and slide layouts while avoiding missing, incorrect, duplicate, or distorted characters.The qualitative comparisons include long paragraphs, seven text locations, and structured slide content.
- Chinese Text Rendering: Qwen-Image accurately renders Chinese couplets, multiple characters, store plaques, spatial layouts, and structured handwritten text in complex scenes.Competing models miss or distort characters and struggle with complex Chinese text and spatial instructions.
- Multi-Object Generation: Qwen-Image follows multi-object prompts by generating the required animals, preserving their positions, applying the requested plush style, and rendering mixed-language text.The passage also compares failures involving incorrect animals and style mismatches.
- Spatial Relationship Generation: Qwen-Image captures specified interactions and spatial relationships more reliably than most compared models.Only Qwen-Image and GPT Image 1 accurately depict the remaining character-object and object-object relationships.
- Text and Material Editing: Qwen-Image edits complex text and materials while preserving style, and uniquely generates the requested enamel-colored glass art.It is compared with Seedream 3.0, GPT Image 1 [High], and FLUX.1 Kontext [Pro].
- Pose Manipulation: During pose editing, Qwen-Image preserves hair, clothing, backgrounds, character details, and scene stability across examples.It also infers hidden clothing structure, such as silk trousers beneath a side-slit dress.
- Chained Editing: Qwen-Image preserves fine textures and completes chained editing sequences that include structural preservation and object addition.GPT Image 1 [High] also extracts paired birds, but Qwen-Image better preserves their fine texture details.
6 Conclusion
Qwen-Image extends image generation toward precise text rendering, image understanding, and broader multimodal integration. Its conclusion emphasizes a foundation model that connects perception and creation across visual modalities.
- Conclusion: Qwen-Image combines major advances in complex text rendering and precise image editing with state-of-the-art performance across image generation and editing tasks.The paper attributes these advances to comprehensive data engineering, progressive curriculum learning, improved multi-task training, and dual encoding.
- Text-image alignment: Its text-image alignment emphasis supports text-rich, visually coherent imagery intended for structured visual explanations and knowledge externalization.The paper positions this capability as a step toward vision-language user interfaces that integrate language and vision.
- Understanding: Qwen-Image applies generative frameworks to classical understanding tasks, using distributional reasoning rather than only direct discriminative inference.For depth estimation, it performs close to specialized discriminative models without surpassing them.
- 3D and video: The model generalizes beyond 2D synthesis to novel view synthesis and pose editing, maintaining consistency across views, subject identity, and background structure.The paper reports that it outperforms dedicated 3D models in several challenging rendering scenarios and uses a video VAE for visual representation.
- Integrated multimodal AI: Qwen-Image complements Qwen2.5-VL’s visual understanding by strengthening the generation pillar of integrated understanding-generation systems.Together, they are presented as a balanced foundation toward Visual-Language Omni systems producing readable and semantically faithful imagery.
- Broader implications: The paper frames Qwen-Image as a paradigm shift toward multimodal agents that both understand images through generative processes and produce interactive visual outputs.This conclusion extends beyond technical benchmarks to implications for perception, interface design, and cognitive modeling.
7 Authors
The paper lists core contributors and additional contributors, with both groups presented in alphabetical order.
- Contributors: The core contributors are listed separately from the additional contributors.The contributor lists contain named individuals in each group.
- Ordering: Both contributor lists are explicitly ordered alphabetically.The footnotes identify the ordering convention for both groups.