Shuai Bai

Jinze Bai, Shuai Bai, Yunfei Chu et al. · pku, tsinghua

Large language models (LLMs) have revolutionized the field of artificial intelligence, enabling natural language processing tasks that were previously thought to be exclusive to humans. In this work, we introduce Qwen, the first installment of our large language model series. Qwen is a comprehensive language model series that encompasses distinct models with varying parameter counts. It includes Qwen, the base pretrained language models, and Qwen-Chat, the chat models finetuned with human alignment techniques. The base language models consistently demonstrate superior performance across a multitude of downstream tasks, and the chat models, particularly those trained using Reinforcement Learning from Human Feedback (RLHF), are highly competitive. The chat models possess advanced tool-use and planning capabilities for creating agent applications, showcasing impressive performance even when compared to bigger models on complex tasks like utilizing a code interpreter. Furthermore, we have developed coding-specialized models, Code-Qwen and Code-Qwen-Chat, as well as mathematics-focused models, Math-Qwen-Chat, which are built upon base language models. These models demonstrate significantly improved performance in comparison with open-source models, and slightly fall behind the proprietary models.

Jinze Bai, Shuai Bai, Shusheng Yang et al.

In this work, we introduce the Qwen-VL series, a set of large-scale vision-language models (LVLMs) designed to perceive and understand both texts and images. Starting from the Qwen-LM as a foundation, we endow it with visual capacity by the meticulously designed (i) visual receptor, (ii) input-output interface, (iii) 3-stage training pipeline, and (iv) multilingual multimodal cleaned corpus. Beyond the conventional image description and question-answering, we implement the grounding and text-reading ability of Qwen-VLs by aligning image-caption-box tuples. The resulting models, including Qwen-VL and Qwen-VL-Chat, set new records for generalist models under similar model scales on a broad range of visual-centric benchmarks (e.g., image captioning, question answering, visual grounding) and different settings (e.g., zero-shot, few-shot). Moreover, on real-world dialog benchmarks, our instruction-tuned Qwen-VL-Chat also demonstrates superiority compared to existing vision-language chatbots. Code, demo and models are available at https://github.com/QwenLM/Qwen-VL.

Jinze Bai, Rui Men, Hao Yang et al. · pku

Generalist models, which are capable of performing diverse multi-modal tasks in a task-agnostic way within a single model, have been explored recently. Being, hopefully, an alternative to approaching general-purpose AI, existing generalist models are still at an early stage, where modality and task coverage is limited. To empower multi-modal task-scaling and speed up this line of research, we release a generalist model learning system, OFASys, built on top of a declarative task interface named multi-modal instruction. At the core of OFASys is the idea of decoupling multi-modal task representations from the underlying model implementations. In OFASys, a task involving multiple modalities can be defined declaratively even with just a single line of code. The system automatically generates task plans from such instructions for training and inference. It also facilitates multi-task training for diverse multi-modal workloads. As a starting point, we provide presets of 7 different modalities and 23 highly-diverse example tasks in OFASys, with which we also develop a first-in-kind, single model, OFA+, that can handle text, image, speech, video, and motion data. The single OFA+ model achieves 95% performance in average with only 16% parameters of 15 task-finetuned models, showcasing the performance reliability of multi-modal task-scaling provided by OFASys. Available at https://github.com/OFA-Sys/OFASys

Peng Wang, Shuai Bai, Sinan Tan et al.

We present the Qwen2-VL Series, an advanced upgrade of the previous Qwen-VL models that redefines the conventional predetermined-resolution approach in visual processing. Qwen2-VL introduces the Naive Dynamic Resolution mechanism, which enables the model to dynamically process images of varying resolutions into different numbers of visual tokens. This approach allows the model to generate more efficient and accurate visual representations, closely aligning with human perceptual processes. The model also integrates Multimodal Rotary Position Embedding (M-RoPE), facilitating the effective fusion of positional information across text, images, and videos. We employ a unified paradigm for processing both images and videos, enhancing the model's visual perception capabilities. To explore the potential of large multimodal models, Qwen2-VL investigates the scaling laws for large vision-language models (LVLMs). By scaling both the model size-with versions at 2B, 8B, and 72B parameters-and the amount of training data, the Qwen2-VL Series achieves highly competitive performance. Notably, the Qwen2-VL-72B model achieves results comparable to leading models such as GPT-4o and Claude3.5-Sonnet across various multimodal benchmarks, outperforming other generalist models. Code is available at https://github.com/QwenLM/Qwen2-VL .

Shuai Bai, Shusheng Yang, Jinze Bai et al.

Large vision-language models (LVLMs) have recently witnessed rapid advancements, exhibiting a remarkable capacity for perceiving, understanding, and processing visual information by connecting visual receptor with large language models (LLMs). However, current assessments mainly focus on recognizing and reasoning abilities, lacking direct evaluation of conversational skills and neglecting visual storytelling abilities. In this paper, we propose an evaluation method that uses strong LLMs as judges to comprehensively evaluate the various abilities of LVLMs. Firstly, we construct a comprehensive visual dialogue dataset TouchStone, consisting of open-world images and questions, covering five major categories of abilities and 27 subtasks. This dataset not only covers fundamental recognition and comprehension but also extends to literary creation. Secondly, by integrating detailed image annotations we effectively transform the multimodal input content into a form understandable by LLMs. This enables us to employ advanced LLMs for directly evaluating the quality of the multimodal dialogue without requiring human intervention. Through validation, we demonstrate that powerful LVLMs, such as GPT-4, can effectively score dialogue quality by leveraging their textual capabilities alone, aligning with human preferences. We hope our work can serve as a touchstone for LVLMs' evaluation and pave the way for building stronger LVLMs. The evaluation code is available at https://github.com/OFA-Sys/TouchStone.

An Yang, Baosong Yang, Binyuan Hui et al.

This report introduces the Qwen2 series, the latest addition to our large language models and large multimodal models. We release a comprehensive suite of foundational and instruction-tuned language models, encompassing a parameter range from 0.5 to 72 billion, featuring dense models and a Mixture-of-Experts model. Qwen2 surpasses most prior open-weight models, including its predecessor Qwen1.5, and exhibits competitive performance relative to proprietary models across diverse benchmarks on language understanding, generation, multilingual proficiency, coding, mathematics, and reasoning. The flagship model, Qwen2-72B, showcases remarkable performance: 84.2 on MMLU, 37.9 on GPQA, 64.6 on HumanEval, 89.5 on GSM8K, and 82.4 on BBH as a base language model. The instruction-tuned variant, Qwen2-72B-Instruct, attains 9.1 on MT-Bench, 48.1 on Arena-Hard, and 35.7 on LiveCodeBench. Moreover, Qwen2 demonstrates robust multilingual capabilities, proficient in approximately 30 languages, spanning English, Chinese, Spanish, French, German, Arabic, Russian, Korean, Japanese, Thai, Vietnamese, and more, underscoring its versatility and global reach. To foster community innovation and accessibility, we have made the Qwen2 model weights openly available on Hugging Face and ModelScope, and the supplementary materials including example code on GitHub. These platforms also include resources for quantization, fine-tuning, and deployment, facilitating a wide range of applications and research endeavors.

Shuai Bai, Huiling Zhou, Zhikang Li et al.

Virtual try-on aims to generate a photo-realistic fitting result given an in-shop garment and a reference person image. Existing methods usually build up multi-stage frameworks to deal with clothes warping and body blending respectively, or rely heavily on intermediate parser-based labels which may be noisy or even inaccurate. To solve the above challenges, we propose a single-stage try-on framework by developing a novel Deformable Attention Flow (DAFlow), which applies the deformable attention scheme to multi-flow estimation. With pose keypoints as the guidance only, the self- and cross-deformable attention flows are estimated for the reference person and the garment images, respectively. By sampling multiple flow fields, the feature-level and pixel-level information from different semantic areas are simultaneously extracted and merged through the attention mechanism. It enables clothes warping and body synthesizing at the same time which leads to photo-realistic results in an end-to-end manner. Extensive experiments on two try-on datasets demonstrate that our proposed method achieves state-of-the-art performance both qualitatively and quantitatively. Furthermore, additional experiments on the other two image editing tasks illustrate the versatility of our method for multi-view synthesis and image animation.

Jianxin Ma, Shuai Bai, Chang Zhou

Generative modeling of human motion has broad applications in computer animation, virtual reality, and robotics. Conventional approaches develop separate models for different motion synthesis tasks, and typically use a model of a small size to avoid overfitting the scarce data available in each setting. It remains an open question whether developing a single unified model is feasible, which may 1) benefit the acquirement of novel skills by combining skills learned from multiple tasks, and 2) help in increasing the model capacity without overfitting by combining multiple data sources. Unification is challenging because 1) it involves diverse control signals as well as targets of varying granularity, and 2) motion datasets may use different skeletons and default poses. In this paper, we present MoFusion, a framework for unified motion synthesis. MoFusion employs a Transformer backbone to ease the inclusion of diverse control signals via cross attention, and pretrains the backbone as a diffusion model to support multi-granularity synthesis ranging from motion completion of a body part to whole-body motion generation. It uses a learnable adapter to accommodate the differences between the default skeletons used by the pretraining and the fine-tuning data. Empirical results show that pretraining is vital for scaling the model size without overfitting, and demonstrate MoFusion's potential in various tasks, e.g., text-to-motion, motion completion, and zero-shot mixing of multiple control signals. Project page: \url{https://ofa-sys.github.io/MoFusion/}.

Zhikang Li, Huiling Zhou, Shuai Bai et al.

The fashion industry has diverse applications in multi-modal image generation and editing. It aims to create a desired high-fidelity image with the multi-modal conditional signal as guidance. Most existing methods learn different condition guidance controls by introducing extra models or ignoring the style prior knowledge, which is difficult to handle multiple signal combinations and faces a low-fidelity problem. In this paper, we adapt both style prior knowledge and flexibility of multi-modal control into one unified two-stage framework, M6-Fashion, focusing on the practical AI-aided Fashion design. It decouples style codes in both spatial and semantic dimensions to guarantee high-fidelity image generation in the first stage. M6-Fashion utilizes self-correction for the non-autoregressive generation to improve inference speed, enhance holistic consistency, and support various signal controls. Extensive experiments on a large-scale clothing dataset M2C-Fashion demonstrate superior performances on various image generation and editing tasks. M6-Fashion model serves as a highly potential AI designer for the fashion industry.

Liang Chen, Haozhe Zhao, Tianyu Liu et al. · pku

In this study, we identify the inefficient attention phenomena in Large Vision-Language Models (LVLMs), notably within prominent models like LLaVA-1.5, QwenVL-Chat and Video-LLaVA. We find out that the attention computation over visual tokens is of extreme inefficiency in the deep layers of popular LVLMs, suggesting a need for a sparser approach compared to textual data handling. To this end, we introduce FastV, a versatile plug-and-play method designed to optimize computational efficiency by learning adaptive attention patterns in early layers and pruning visual tokens in subsequent ones. Our evaluations demonstrate FastV's ability to dramatically reduce computational costs (e.g., a 45 reduction in FLOPs for LLaVA-1.5-13B) without sacrificing performance in a wide range of image and video understanding tasks. The computational efficiency and performance trade-off of FastV are highly customizable and pareto-efficient. It can compress the FLOPs of a 13B-parameter model to achieve a lower budget than that of a 7B-parameter model, while still maintaining superior performance. We believe FastV has practical values for deployment of LVLMs in edge devices and commercial models. Code is released at https://github.com/pkunlp-icler/FastV.

Jin Xu, Zhifang Guo, Hangrui Hu et al. · pku

We present Qwen3-Omni, a single multimodal model that, for the first time, maintains state-of-the-art performance across text, image, audio, and video without any degradation relative to single-modal counterparts. Qwen3-Omni matches the performance of same-sized single-modal models within the Qwen series and excels particularly on audio tasks. Across 36 audio and audio-visual benchmarks, Qwen3-Omni achieves open-source SOTA on 32 benchmarks and overall SOTA on 22, outperforming strong closed-source models such as Gemini-2.5-Pro, Seed-ASR, and GPT-4o-Transcribe. Qwen3-Omni adopts a Thinker-Talker MoE architecture that unifies perception and generation across text, images, audio, and video, yielding fluent text and natural real-time speech. It supports text interaction in 119 languages, speech understanding in 19 languages, and speech generation in 10 languages. To reduce first-packet latency in streaming synthesis, Talker autoregressively predicts discrete speech codecs using a multi-codebook scheme. Leveraging the representational capacity of these codebooks, we replace computationally intensive block-wise diffusion with a lightweight causal ConvNet, enabling streaming from the first codec frame. In cold-start settings, Qwen3-Omni achieves a theoretical end-to-end first-packet latency of 234 ms. To further strengthen multimodal reasoning, we introduce a Thinking model that explicitly reasons over inputs from any modality. Since the research community currently lacks a general-purpose audio captioning model, we fine-tuned Qwen3-Omni-30B-A3B to obtain Qwen3-Omni-30B-A3B-Captioner, which produces detailed, low-hallucination captions for arbitrary audio inputs. Qwen3-Omni-30B-A3B, Qwen3-Omni-30B-A3B-Thinking, and Qwen3-Omni-30B-A3B-Captioner are publicly released under the Apache 2.0 license.

Jin Xu, Zhifang Guo, Jinzheng He et al.

In this report, we present Qwen2.5-Omni, an end-to-end multimodal model designed to perceive diverse modalities, including text, images, audio, and video, while simultaneously generating text and natural speech responses in a streaming manner. To enable the streaming of multimodal information inputs, both audio and visual encoders utilize a block-wise processing approach. To synchronize the timestamps of video inputs with audio, we organize the audio and video sequentially in an interleaved manner and propose a novel position embedding approach, named TMRoPE(Time-aligned Multimodal RoPE). To concurrently generate text and speech while avoiding interference between the two modalities, we propose \textbf{Thinker-Talker} architecture. In this framework, Thinker functions as a large language model tasked with text generation, while Talker is a dual-track autoregressive model that directly utilizes the hidden representations from the Thinker to produce audio tokens as output. Both the Thinker and Talker models are designed to be trained and inferred in an end-to-end manner. For decoding audio tokens in a streaming manner, we introduce a sliding-window DiT that restricts the receptive field, aiming to reduce the initial package delay. Qwen2.5-Omni is comparable with the similarly sized Qwen2.5-VL and outperforms Qwen2-Audio. Furthermore, Qwen2.5-Omni achieves state-of-the-art performance on multimodal benchmarks like Omni-Bench. Notably, Qwen2.5-Omni's performance in end-to-end speech instruction following is comparable to its capabilities with text inputs, as evidenced by benchmarks such as MMLU and GSM8K. As for speech generation, Qwen2.5-Omni's streaming Talker outperforms most existing streaming and non-streaming alternatives in robustness and naturalness.

Peng Wang, Shijie Wang, Junyang Lin et al.

In this work, we explore a scalable way for building a general representation model toward unlimited modalities. We release ONE-PEACE, a highly extensible model with 4B parameters that can seamlessly align and integrate representations across vision, audio, and language modalities. The architecture of ONE-PEACE comprises modality adapters, shared self-attention layers, and modality FFNs. This design allows for the easy extension of new modalities by adding adapters and FFNs, while also enabling multi-modal fusion through self-attention layers. To pretrain ONE-PEACE, we develop two modality-agnostic pretraining tasks, cross-modal aligning contrast and intra-modal denoising contrast, which align the semantic space of different modalities and capture fine-grained details within modalities concurrently. With the scaling-friendly architecture and pretraining tasks, ONE-PEACE has the potential to expand to unlimited modalities. Without using any vision or language pretrained model for initialization, ONE-PEACE achieves leading results on a wide range of uni-modal and multi-modal tasks, including image classification (ImageNet), semantic segmentation (ADE20K), audio-text retrieval (AudioCaps, Clotho), audio classification (ESC-50, FSD50K, VGGSound), audio question answering (AVQA), image-text retrieval (MSCOCO, Flickr30K), and visual grounding (RefCOCO/+/g). Code is available at https://github.com/OFA-Sys/ONE-PEACE.

Peng Wang, An Yang, Rui Men et al.

In this work, we pursue a unified paradigm for multimodal pretraining to break the scaffolds of complex task/modality-specific customization. We propose OFA, a Task-Agnostic and Modality-Agnostic framework that supports Task Comprehensiveness. OFA unifies a diverse set of cross-modal and unimodal tasks, including image generation, visual grounding, image captioning, image classification, language modeling, etc., in a simple sequence-to-sequence learning framework. OFA follows the instruction-based learning in both pretraining and finetuning stages, requiring no extra task-specific layers for downstream tasks. In comparison with the recent state-of-the-art vision & language models that rely on extremely large cross-modal datasets, OFA is pretrained on only 20M publicly available image-text pairs. Despite its simplicity and relatively small-scale training data, OFA achieves new SOTAs in a series of cross-modal tasks while attaining highly competitive performances on uni-modal tasks. Our further analysis indicates that OFA can also effectively transfer to unseen tasks and unseen domains. Our code and models are publicly available at https://github.com/OFA-Sys/OFA.

Shuai Bai, Zhedong Zheng, Xiaohan Wang et al.

Vehicle search is one basic task for the efficient traffic management in terms of the AI City. Most existing practices focus on the image-based vehicle matching, including vehicle re-identification and vehicle tracking. In this paper, we apply one new modality, i.e., the language description, to search the vehicle of interest and explore the potential of this task in the real-world scenario. The natural language-based vehicle search poses one new challenge of fine-grained understanding of both vision and language modalities. To connect language and vision, we propose to jointly train the state-of-the-art vision models with the transformer-based language model in an end-to-end manner. Except for the network structure design and the training strategy, several optimization objectives are also re-visited in this work. The qualitative and quantitative experiments verify the effectiveness of the proposed method. Our proposed method has achieved the 1st place on the 5th AI City Challenge, yielding competitive performance 18.69% MRR accuracy on the private test set. We hope this work can pave the way for the future study on using language description effectively and efficiently for real-world vehicle retrieval systems. The code will be available at https://github.com/ShuaiBai623/AIC2021-T5-CLV.

Hanzhe Hu, Shuai Bai, Aoxue Li et al.

Conventional deep learning based methods for object detection require a large amount of bounding box annotations for training, which is expensive to obtain such high quality annotated data. Few-shot object detection, which learns to adapt to novel classes with only a few annotated examples, is very challenging since the fine-grained feature of novel object can be easily overlooked with only a few data available. In this work, aiming to fully exploit features of annotated novel object and capture fine-grained features of query object, we propose Dense Relation Distillation with Context-aware Aggregation (DCNet) to tackle the few-shot detection problem. Built on the meta-learning based framework, Dense Relation Distillation module targets at fully exploiting support features, where support features and query feature are densely matched, covering all spatial locations in a feed-forward fashion. The abundant usage of the guidance information endows model the capability to handle common challenges such as appearance changes and occlusions. Moreover, to better capture scale-aware features, Context-aware Aggregation module adaptively harnesses features from different scales for a more comprehensive feature representation. Extensive experiments illustrate that our proposed approach achieves state-of-the-art results on PASCAL VOC and MS COCO datasets. Code will be made available at https://github.com/hzhupku/DCNet.

Liang Chen, Shuai Bai, Wenhao Chai et al.

The field of advanced text-to-image generation is witnessing the emergence of unified frameworks that integrate powerful text encoders, such as CLIP and T5, with Diffusion Transformer backbones. Although there have been efforts to control output images with additional conditions, like canny and depth map, a comprehensive framework for arbitrary text-image interleaved control is still lacking. This gap is especially evident when attempting to merge concepts or visual elements from multiple images in the generation process. To mitigate the gap, we conducted preliminary experiments showing that large multimodal models (LMMs) offer an effective shared representation space, where image and text can be well-aligned to serve as a condition for external diffusion models. Based on this discovery, we propose Dream Engine, an efficient and unified framework designed for arbitrary text-image interleaved control in image generation models. Building on powerful text-to-image models like SD3.5, we replace the original text-only encoders by incorporating versatile multimodal information encoders such as QwenVL. Our approach utilizes a two-stage training paradigm, consisting of joint text-image alignment and multimodal interleaved instruction tuning. Our experiments demonstrate that this training method is effective, achieving a 0.69 overall score on the GenEval benchmark, and matching the performance of state-of-the-art text-to-image models like SD3.5 and FLUX.

Xiao Pan, Zongxin Yang, Shuai Bai et al.

In this paper, we focus on the One-shot Novel View Synthesis (O-NVS) task which targets synthesizing photo-realistic novel views given only one reference image per scene. Previous One-shot Generalizable Neural Radiance Fields (OG-NeRF) methods solve this task in an inference-time finetuning-free manner, yet suffer the blurry issue due to the encoder-only architecture that highly relies on the limited reference image. On the other hand, recent diffusion-based image-to-3d methods show vivid plausible results via distilling pre-trained 2D diffusion models into a 3D representation, yet require tedious per-scene optimization. Targeting these issues, we propose the GD$^2$-NeRF, a Generative Detail compensation framework via GAN and Diffusion that is both inference-time finetuning-free and with vivid plausible details. In detail, following a coarse-to-fine strategy, GD$^2$-NeRF is mainly composed of a One-stage Parallel Pipeline (OPP) and a 3D-consistent Detail Enhancer (Diff3DE). At the coarse stage, OPP first efficiently inserts the GAN model into the existing OG-NeRF pipeline for primarily relieving the blurry issue with in-distribution priors captured from the training dataset, achieving a good balance between sharpness (LPIPS, FID) and fidelity (PSNR, SSIM). Then, at the fine stage, Diff3DE further leverages the pre-trained image diffusion models to complement rich out-distribution details while maintaining decent 3D consistency. Extensive experiments on both the synthetic and real-world datasets show that GD$^2$-NeRF noticeably improves the details while without per-scene finetuning.

Hanzhe Hu, Deyi Ji, Weihao Gan et al.

Recent works have made great progress in semantic segmentation by exploiting contextual information in a local or global manner with dilated convolutions, pyramid pooling or self-attention mechanism. In order to avoid potential misleading contextual information aggregation in previous works, we propose a class-wise dynamic graph convolution (CDGC) module to adaptively propagate information. The graph reasoning is performed among pixels in the same class. Based on the proposed CDGC module, we further introduce the Class-wise Dynamic Graph Convolution Network(CDGCNet), which consists of two main parts including the CDGC module and a basic segmentation network, forming a coarse-to-fine paradigm. Specifically, the CDGC module takes the coarse segmentation result as class mask to extract node features for graph construction and performs dynamic graph convolutions on the constructed graph to learn the feature aggregation and weight allocation. Then the refined feature and the original feature are fused to get the final prediction. We conduct extensive experiments on three popular semantic segmentation benchmarks including Cityscapes, PASCAL VOC 2012 and COCO Stuff, and achieve state-of-the-art performance on all three benchmarks.

Shuai Bai, Zhiqun He, Ting-Bing Xu et al.

For visual tracking, most of the traditional correlation filters (CF) based methods suffer from the bottleneck of feature redundancy and lack of motion information. In this paper, we design a novel tracking framework, called multi-hierarchical independent correlation filters (MHIT). The framework consists of motion estimation module, hierarchical features selection, independent CF online learning, and adaptive multi-branch CF fusion. Specifically, the motion estimation module is introduced to capture motion information, which effectively alleviates the object partial occlusion in the temporal video. The multi-hierarchical deep features of CNN representing different semantic information can be fully excavated to track multi-scale objects. To better overcome the deep feature redundancy, each hierarchical features are independently fed into a single branch to implement the online learning of parameters. Finally, an adaptive weight scheme is integrated into the framework to fuse these independent multi-branch CFs for the better and more robust visual object tracking. Extensive experiments on OTB and VOT datasets show that the proposed MHIT tracker can significantly improve the tracking performance. Especially, it obtains a 20.1% relative performance gain compared to the top trackers on the VOT2017 challenge, and also achieves new state-of-the-art performance on the VOT2018 challenge.