论文检索

Jianwei Zheng, Xiaomin Yao, Guojiang Shen, Wei Li, Jiawei Jiang

Deep unfolding network (DUN) has shed new light on multi-sequence MRI reconstruction, providing both high interpretability and acceptable performance. However, current approaches still suffer from the plight of information isolation, i.e., learning features of multi-suquences individually and leaving the mask departed from model updating. In this work, we propose a new unfolding solution, namely Information-coupled MRI Acceleration (IMA), to address the isolation issue. Concretely, two specific mechanisms are presented. On the one hand, the latent connections across different sequences are explicitly molded via two auxiliary matrices. While the first matrix is meticulously engineered to assemble the spatial details, the second one hammers at capturing the depth information conditioned on the enriched channels. On the other hand, following a deep analysis on the non-uniform distribution in low- and high-frequency components of the given mask, we elaborate a new unfolding flow using a progressive masking scheme, featuring a dilation-contraction mechanism during forward propagation of successive stages. Massive experiments are conducted under various sampling patterns and acceleration rates, whose results demonstrate that, without any sophisticated architectures, our IMA outperforms the current cutting-edge methods both visually and numerically.

Xu Zhang, Jin Yuan, Hanwang Zhang, Guojin Zhong, Yongsheng Zang, Jiacheng Lin, Zhiyong Li

Controllable image semantic understanding tasks, such as captioning or segmentation, necessitate users to input a prompt (e.g., text or bounding boxes) to predict a unique outcome, presenting challenges such as high-cost prompt input or limited information output. This paper introduces a new task ``Image Collaborative Segmentation and Captioning'' (SegCaptioning), which aims to translate a straightforward prompt, like a bounding box around an object, into diverse semantic interpretations represented by (caption, masks) pairs, allowing flexible result selection by users. This task poses significant challenges, including accurately capturing a user's intention from a minimal prompt while simultaneously predicting multiple semantically aligned caption words and masks. Technically, we propose a novel Scene Graph Guided Diffusion Model that leverages structured scene graph features for correlated mask-caption prediction. Initially, we introduce a Prompt-Centric Scene Graph Adaptor to map a user's prompt to a scene graph, effectively capturing his intention. Subsequently, we employ a diffusion process incorporating a Scene Graph Guided Bimodal Transformer to predict correlated caption-mask pairs by uncovering intricate correlations between them. To ensure accurate alignment, we design a Multi-Entities Contrastive Learning loss to explicitly align visual and textual entities by considering inter-modal similarity, resulting in well-aligned caption-mask pairs. Extensive experiments conducted on two datasets demonstrate that SGDiff achieves superior performance in SegCaptioning, yielding promising results for both captioning and segmentation tasks with minimal prompt input.

Guosheng Zhang, Keyao Wang, Haixiao Yue, Ajian Liu, Gang Zhang, Kun Yao, Errui Ding, Jingdong Wang

Face Anti-Spoofing (FAS) is essential for ensuring the security and reliability of facial recognition systems. Most existing FAS methods are formulated as binary classification tasks, providing confidence scores without interpretation. They exhibit limited generalization in out-of-domain scenarios, such as new environments or unseen spoofing types. In this work, we introduce a multimodal large language model (MLLM) framework for FAS, termed Interpretable Face Anti-Spoofing (I-FAS), which transforms the FAS task into an interpretable visual question answering (VQA) paradigm. Specifically, we propose a Spoof-aware Captioning and Filtering (SCF) strategy to generate high-quality captions for FAS images, enriching the model's supervision with natural language interpretations. To mitigate the impact of noisy captions during training, we develop a Lopsided Language Model (L-LM) loss function that separates loss calculations for judgment and interpretation, prioritizing the optimization of the former. Furthermore, to enhance the model's perception of global visual features, we design a Globally Aware Connector (GAC) to align multi-level visual representations with the language model. Extensive experiments on standard and newly devised One to Eleven cross-domain benchmarks, comprising 12 public datasets, demonstrate that our method significantly outperforms state-of-the-art methods.

Guoyuan Yang, Mufan Xue, Ziming Mao, Haofang Zheng, Jia Xu, Dabin Sheng, Ruotian Sun, Ruoqi Yang, Xuesong Li

Prior work employing deep neural networks (DNNs) with explainable techniques has identified human visual cortical selective representation to specific categories. However, constructing high-performing encoding models that accurately capture brain responses to coexisting multi-semantics remains elusive. Here, we used CLIP models combined with CLIP Dissection to establish a multi-semantic mapping framework (CLIP-MSM) for hypothesis-free analysis in human high-level visual cortex. First, we utilize CLIP models to construct voxel-wise encoding models for predicting visual cortical responses to natural scene images. Then, we apply CLIP Dissection and normalize the semantic mapping score to achieve the mapping of single brain voxels to multiple semantics. Our findings indicate that CLIP Dissection applied to DNNs modeling the human high-level visual cortex demonstrates better interpretability accuracy compared to Network Dissection. In addition, to demonstrate how our method enables fine-grained discovery in hypothesis-free analysis, we quantify the accuracy between CLIP-MSM’s reconstructed brain activation in response to categories of faces, bodies, places, words and food, and the ground truth of brain activation. We demonstrate that CLIP-MSM provides more accurate predictions of visual responses compared to CLIP Dissection. Our results have been validated using two large natural image datasets: the Natural Scenes Dataset (NSD) and the Natural Object Dataset (NOD).

Wenzhuang Wang, Mingcan Ma, Yong Chen, Changqun Xia, Zhenbao Liang, Jia Li

Text-to-image diffusion model has inspired research into text-to-data synthesis without human intervention, where spatial attentions correlated with semantic entities in text prompts are primarily interpreted as pseudo-masks. However, these vannila attentions often deliver visual-linguistic discrepancies, in which the associations between image features and entity-level tokens are unstable and divergent, yielding inferior masks for realistic applications, especially in more practical open-vocabulary settings. To tackle this issue, we propose a novel text-guided self-driven generative paradigm, termed FreeGen, which addresses the discrepancies by recalibrating intrinsic visual-linguistic correlations and serves as a real-data-free method to automatically synthesize open-vocabulary pixel-level data for arbitrary entities. Specifically, we first learn an Attention Self-Rectification mechanism to reproject the inherent attention matrices to achieve robust semantic alignment, thereby obtaining class-discriminative masks. A Temporal Fluctuation Factor is present to assess mask quality based on its variation over uniform sampling timesteps, enabling the selection of reliable masks. These masks are then employed as self-supervised signals to support the learning of an Entity-level Grounding Decoder in a self-training manner, thus producing open-vocabulary segmentation results. Extensive experiments show that the existing segmenters trained on FreeGen narrow the performance gap with real data counterparts and remarkably outperform the state-of-the-art methods.

Wenbin Wang, Liang Ding, Minyan Zeng, Xiabin Zhou, Li Shen, Yong Luo, Wei Yu, Dacheng Tao

Multimodal large language models (MLLMs) have experienced significant advancements recently, but still struggle to recognize and interpret intricate details in high-resolution (HR) images effectively. While state-of-the-art (SOTA) MLLMs claim to process images at 4K resolution, existing MLLM benchmarks only support up to 2K, leaving the capabilities of SOTA models on true HR images largely untested. Furthermore, existing methods for enhancing HR image perception in MLLMs rely on computationally expensive visual instruction tuning. To address these limitations, we introduce HR-Bench, the first deliberately designed benchmark to rigorously evaluate MLLM performance on 4K & 8K images. Through extensive experiments, we demonstrate that while downsampling HR images leads to vision information loss, leveraging complementary modalities, e.g., text, can effectively compensate for this loss. Building upon this insight, we propose Divide, Conquer and Combine, a novel training-free framework for enhancing MLLM perception of HR images. Our method follows a three-staged approach: 1) Divide: recursively partitioning the HR image into patches and merging similar patches to minimize computational overhead, 2) Conquer: leveraging the MLLM to generate accurate textual descriptions for each image patch, and 3) Combine: utilizing the generated text descriptions to enhance the MLLM's understanding of the overall HR image. Extensive experiments show that: 1) the SOTA MLLM achieves 63% accuracy, which is markedly lower than the 87% accuracy achieved by humans on HR-Bench; 2) our method brings consistent and significant improvements (a relative increase of +6% on HR-Bench and +8% on general multimodal benchmarks).

Siwen Wang, Churan Wang, Fei Gao, Lixian Su, Fandong Zhang, Yizhou Wang, Yizhou Yu

Three-dimensional (3D) medical images, such as Computed Tomography (CT) and Magnetic Resonance Imaging (MRI), are essential for clinical applications. However, the need for diverse and comprehensive representations is particularly pronounced when considering the variability across different organs, diagnostic tasks, and imaging modalities. How to effectively interpret the intricate contextual information and extract meaningful insights from these images remains an open challenge to the community. While current self-supervised learning methods have shown potential, they often consider an image as a whole thereby overlooking the extensive, complex relationships among local regions from one or multiple images. In this work, we introduce a pioneering method for learning 3D medical image representations through an autoregressive pre-training framework. Our approach sequences various 3D medical images based on spatial, contrast, and semantic correlations, treating them as interconnected visual tokens within a token sequence. By employing an autoregressive sequence modeling task, we predict the next visual token in the sequence, which allows our model to deeply understand and integrate the contextual information inherent in 3D medical images. Additionally, we implement a random startup strategy to avoid overestimating token relationships and to enhance the robustness of learning. The effectiveness of our approach is demonstrated by the superior performance over others on nine downstream tasks in public datasets.

Yu Tian, Yixuan Li, Baoliang Chen, Hanwei Zhu, Shiqi Wang, Sam Kwong

Assessing the quality of artificial intelligence-generated images (AIGIs) plays a crucial role in their application in real-world scenarios. However, traditional image quality assessment (IQA) algorithms primarily focus on low-level visual perception, while existing IQA works on AIGIs overemphasize the generated content itself, neglecting its effectiveness in real-world applications. To bridge this gap, we propose AIGI-VC, a quality assessment database for AI-Generated Images in Visual Communication, which studies the communicability of AIGIs in the advertising field from the perspectives of information clarity and emotional interaction. The dataset consists of 2,500 images spanning 14 advertisement topics and 8 emotion types. It provides coarse-grained human preference annotations and fine-grained preference descriptions, benchmarking the abilities of IQA methods in preference prediction, interpretation, and reasoning. We conduct an empirical study of existing representative IQA methods and large multi-modal models on the AIGI-VC dataset, uncovering their strengths and weaknesses.

Yunlong Tang, Gen Zhan, Li Yang, Yiting Liao, Chenliang Xu

Video saliency prediction aims to identify the regions in a video that attract human attention and gaze, driven by bottom-up features from the video and top-down processes like memory and cognition. Among these top-down influences, language plays a crucial role in guiding attention by shaping how visual information is interpreted. Existing methods primarily focus on modeling perceptual information while neglecting the reasoning process facilitated by language, where ranking cues are crucial outcomes of this process and practical guidance for saliency prediction. In this paper, we propose CaRDiff (Caption, Rank, and generate with Diffusion), a framework that imitates the process by integrating multimodal large language model (MLLM), a grounding module, and a diffusion model, to enhance video saliency prediction. Specifically, we introduce a novel prompting method VSOR-CoT (Video Slient Object Ranking Chain of Thought), which utilizes an MLLM with a grounding module to caption video content and infer salient objects along with their rankings and positions. This process derives ranking maps that can be sufficiently leveraged by the diffusion model to accurately decode the saliency maps for the given video. Extensive experiments showcase the effectiveness of VSOR-CoT in improving the performance of video saliency prediction. The proposed CaRDiff performs better than state-of-the-art models on the MVS dataset and demonstrates cross-dataset capabilities on the DHF1k dataset through zero-shot evaluation.

Mengjie Qin, Yuchao Feng, Zongliang Wu, Yulun Zhang, Xin Yuan

In the coded aperture snapshot spectral imaging system, Deep Unfolding Networks (DUNs) have made impressive progress in recovering 3D hyperspectral images (HSIs) from a single 2D measurement. However, the inherent nonlinear and ill-posed characteristics of HSI reconstruction still pose challenges to existing methods in terms of accuracy and stability. To address this issue, we propose a Mamba-inspired Joint Unfolding Network (MiJUN), which integrates physics-embedded DUNs with learning-based HSI imaging. Firstly, leveraging the concept of trapezoid discretization to expand the representation space of unfolding networks, we introduce an accelerated unfolding network scheme. This approach can be interpreted as a generalized accelerated half-quadratic splitting with a second-order differential equation, which reduces the reliance on initial optimization stages and addresses challenges related to long-range interactions. Crucially, within the Mamba framework, we restructure the Mamba-inspired global-to-local attention mechanism by incorporating a selective state space model and an attention mechanism. This effectively reinterprets Mamba as a variant of the Transformer architecture, improving its adaptability and efficiency. Furthermore, we refine the scanning strategy with Mamba by integrating the tensor mode-k unfolding into the Mamba network. This approach emphasizes the low-rank properties of tensors along various modes, while conveniently facilitating 12 scanning directions. Numerical and visual comparisons on both simulation and real datasets demonstrate the superiority of our proposed MiJUN, and achieving overwhelming detail representation.

Hao Li, Hao Fei, Zechao Hu, Zhengwei Yang, Zheng Wang

Social Intelligence Queries (Social-IQ) serve as the primary multimodal benchmark for evaluating a model’s social intelligence level. While impressive multiple-choice question (MCQ) accuracy is achieved by current solutions, increasing evidence shows that they are largely, and in some cases entirely, dependent on language modality, overlooking visual context. Additionally, the closed-set nature further prevents the exploration of whether and to what extent the reasoning path behind selection is correct. To address these limitations, we propose the Visually Explainable and Grounded Artificial Social Intelligence (VEGAS) model. As a generative multimodal model, VEGAS leverages open-ended answering to provide explainable responses, which enhances the clarity and evaluation of reasoning paths. To enable visually grounded answering, we propose a novel sampling strategy to provide the model with more relevant visual frames. We then enhance the model’s interpretation of these frames through Generalist Instruction Fine-Tuning (GIFT), which aims to: i) learn multimodal language transformations for fundamental emotional social traits, and ii) establish multimodal joint reasoning capabilities. Extensive experiments, comprising modality ablation, open-ended assessments, and supervised MCQ evaluations, consistently show that VEGAS effectively utilizes visual information in reasoning to produce correct and also credible answers. We expect this work to offer a new perspective on Social-IQ and advance the development of human-like social AI.

Chenxin Li, Xinyu Liu, Wuyang Li, Cheng Wang, Hengyu Liu, Yifan Liu, Zhen Chen, Yixuan Yuan

U-Net has become a cornerstone in various visual applications such as image segmentation and diffusion probability models. While numerous innovative designs and improvements have been introduced by incorporating transformers or MLPs, the networks are still limited to linearly modeling patterns as well as the deficient interpretability. To address these challenges, our intuition is inspired by the impressive results of the Kolmogorov-Arnold Networks (KANs) in terms of accuracy and interpretability, which reshape the neural network learning via the stack of non-linear learnable activation functions derived from the Kolmogorov-Anold representation theorem. Specifically, in this paper, we explore the untapped potential of KANs in improving backbones for vision tasks. We investigate, modify and re-design the established U-Net pipeline by integrating the dedicated KAN layers on the tokenized intermediate representation, termed U-KAN. Rigorous medical image segmentation benchmarks verify the superiority of UKAN by higher accuracy even with less computation cost. We further delved into the potential of U-KAN as an alternative U-Net noise predictor in diffusion models, demonstrating its applicability in generating task-oriented model architectures.

Er Jin, Qihui Feng, Yongli Mou, Gerhard Lakemeyer, Stefan Decker, Oliver Simons, Johannes Stegmaier

Logical image understanding involves interpreting and reasoning about the relationships and consistency within an image's visual content. This capability is essential in applications such as industrial inspection, where logical anomaly detection is critical for maintaining high-quality standards and minimizing costly recalls. Previous research in anomaly detection (AD) has relied on prior knowledge for designing algorithms, which often requires extensive manual annotations, significant computing power, and large amounts of data for training. Autoregressive, multimodal Vision Language Models (AVLMs) offer a promising alternative due to their exceptional performance in visual reasoning across various domains. Despite this, their application to logical AD remains unexplored. In this work, we investigate using AVLMs for logical AD and demonstrate that they are well-suited to the task. Combining AVLMs with format embedding and a logic reasoner, we achieve SOTA performance on public benchmarks, MVTec LOCO AD, with an AUROC of 86.0% and an F1-max of 83.7% along with explanations of the anomalies. This significantly outperforms the existing SOTA method by 18.1% in AUROC and 4.6% in F1-max score.

Muye Huang, Lingling Zhang, Han Lai, Wenjun Wu, Xinyu Zhang, Jun Liu

Charts are widely used for data visualization across various fields, including education, research, and business. Chart Question Answering (CQA) is an emerging task focused on the automatic interpretation and reasoning of data presented in charts. However, chart images are inherently difficult to interpret, and chart-related questions often involve complex logical and numerical reasoning, which hinders the performance of existing models. This paper introduces VProChart, a novel framework designed to address these challenges in CQA by integrating a lightweight Visual Perception Alignment Agent (VPAgent) and a Programmatic Solution Reasoning approach. VPAgent aligns and models chart elements based on principles of human visual perception, enhancing the understanding of chart context. The Programmatic Solution Reasoning approach leverages large language models (LLMs) to transform natural language reasoning questions into structured solution programs, facilitating precise numerical and logical reasoning. Extensive experiments on benchmark datasets such as ChartQA and PlotQA demonstrate that VProChart significantly outperforms existing methods, highlighting its capability in understanding and reasoning with charts.

Shiyu Hou, Tianfei Zhou, Shuai Zhang, Ye Yuan, Guoren Wang

Prompt tuning (PT) has emerged as a key to unlocking the power of visual-language models like CLIP for various downstream tasks. Predominant approaches learn a small set of task-relevant soft prompts by solving an image-class matching problem. Nevertheless, by optimizing merely with respect to class names, they face challenges in learning high performant prompts capable of capturing fine-grained, diverse characteristics of each class, and tends to overfit potentially biased distribution of base classes. In this work, we propose PTinCAS to tackle prompt tuning in a compact attribute space, driven by the premise that attributes offer detailed class interpretations and can facilitate transfer across related categories. Particularly, PTinCAS is grounded in two innovative designs. First, we create a compact attribute space by properly prompting large language models to generate factual descriptions about categories, which are subsequently clustered to form a concise attribute vocabulary. Second, we leverage attributes as a source of supervision in PT to transfer the inherent common sense knowledge in attributes to soft prompts. An object-aware visual prompting mechanism is developed to effortlessly highlight intended regions in the original image, which guides the model towards learning visual attributes associated with object regions rather than the background. We show that PTinCAS not only improves few-shot generalizability compared to existing PT methods, but also provides some level of inherent explainability that helps us understand why a class name is determined based on the attributes activated in an image.

Hangzhou He, Lei Zhu, Xinliang Zhang, Shuang Zeng, Qian Chen, Yanye Lu

Concept Bottleneck Models (CBMs) offer inherent interpretability by initially translating images into human-comprehensible concepts, followed by a linear combination of these concepts for classification. However, the annotation of concepts for visual recognition tasks requires extensive expert knowledge and labor, constraining the broad adoption of CBMs. Recent approaches have leveraged the knowledge of large language models to construct concept bottlenecks, with multimodal models like CLIP subsequently mapping image features into the concept feature space for classification. Despite this, the concepts produced by language models can be verbose and may introduce non-visual attributes, which hurts accuracy and interpretability. In this study, we investigate to avoid these issues by constructing CBMs directly from multimodal models. To this end, we adopt common words as base concept vocabulary and leverage auxiliary unlabeled images to construct a Vision-to-Concept (V2C) tokenizer that can explicitly quantize images into their most relevant visual concepts, thus creating a vision-oriented concept bottleneck tightly coupled with the multimodal model. This leads to our V2C-CBM which is training efficient and interpretable with high accuracy. Our V2C-CBM has matched or outperformed LLM-supervised CBMs on various visual classification benchmarks, validating the efficacy of our approach.

Xingrui Wang, Wufei Ma, Tiezheng Zhang, Celso M de Melo, Jieneng Chen, Alan Yuille

Although large multimodal models (LMMs) have demonstrated remarkable capabilities in visual scene interpretation and reasoning, their capacity for complex and precise 3-dimensional spatial reasoning remains uncertain. Existing benchmarks focus predominantly on 2D spatial understanding and lack a framework to comprehensively evaluate 6D spatial reasoning across varying complexities.To address this limitation, we present Spatial457, a scalable and unbiased synthetic dataset designed with 4 key spatial components: multi-object recognition, 2D and 3D spatial relationships, and 3D orientation. Spatial457 supports a cascading evaluation structure, offering 7 question types across 5 difficulty levels that progress from basic single-object recognition to our newly proposed complex 6D spatial reasoning tasks.We evaluated various large multimodal models (LMMs) on Spatial457, observing a general decline in performance as task complexity increases, particularly in 3D reasoning and 6D spatial tasks. To quantify these challenges, we introduce the Relative Performance Dropping Rate (RPDR), highlighting key weaknesses in 3D reasoning capabilities. Leveraging the unbiased attribute design of our dataset, we also uncover prediction biases across different attributes, with similar patterns observed in real-world image settings.

Yanbang Li, Ziyang Gong, Haoyang Li, Xiaoqi Huang, Haolan Kang, Guangping Bai, Xianzheng Ma

Recently, natural language has been the primary medium for human-robot interaction. However, its inherent lack of spatial precision for robotic control introduces challenges such as ambiguity and verbosity. To address these limitations, we introduce the ***Robotic Visual Instruction (RoVI)***, a novel paradigm to guide robotic tasks through an object-centric, hand-drawn symbolic representation. RoVI effectively encodes spatial-temporal information into human-interpretable visual instructions through 2D sketches, utilizing arrows, circles, colors, and numbers to direct 3D robotic manipulation. To enable robots to understand RoVI better and generate precise actions based on RoVI, we present ***Visual Instruction Embodied Workflow (VIEW)***, a pipeline formulated for RoVI-conditioned policies. This approach leverages Vision-Language Models (VLMs) to interpret RoVI inputs, decode spatial and temporal constraints from 2D pixel space via keypoint extraction, and then transform them into executable 3D action sequences. We additionally curate a specialized dataset of 15K instances to fine-tune small VLMs for edge deployment, enabling them to effectively learn RoVI capabilities. Our approach is rigorously validated across 11 novel tasks in both real and simulated environments, demonstrating significant generalization capability. Notably, VIEW achieves an 87.5% success rate in real-world scenarios involving unseen tasks that feature multi-step actions, with disturbances, and trajectory-following requirements. Code and Datasets in this paper will be released soon.

Qi Zhu, Jiangwei Lao, Deyi Ji, Junwei Luo, Kang Wu, Yingying Zhang, Lixiang Ru, Jian Wang, Jingdong Chen, Ming Yang 等

Open-world interpretation aims to accurately localize and recognize all objects within images by vision-language models (VLMs). While substantial progress has been made in this task for natural images, the advancements for remote sensing (RS) images still remain limited, primarily due to these two challenges. 1) Existing RS semantic categories are limited, particularly for pixel-level interpretation datasets. 2) Distinguishing among diverse RS spatial regions solely by language space is challenging due to the dense and intricate spatial distribution in open-world RS imagery. To address the first issue, we develop a fine-grained RS interpretation dataset, Sky-SA, which contains 183,375 high-quality local image-text pairs with full-pixel manual annotations, covering 1,763 category labels, exhibiting richer semantics and higher density than previous datasets. Afterwards, to solve the second issue, we introduce the vision-centric principle for vision-language modeling. Specifically, in the pre-training stage, the visual self-supervised paradigm is incorporated into image-text alignment, reducing the degradation of general visual representation capabilities of existing paradigms. Then, we construct a visual-relevance knowledge graph across open-category texts and further develop a novel vision-centric image-text contrastive loss for fine-tuning with text prompts. This new model, denoted as SkySense-O, demonstrates impressive zero-shot capabilities on a thorough evaluation encompassing 14 datasets over 4 tasks, from recognizing to reasoning and classification to localization. Specifically, it outperforms the latest models such as SegEarth-OV, GeoRSCLIP, and VHM by a large margin, i.e., 11.95%, 8.04% and 3.55% on average respectively.

Chenxin Tao, Shiqian Su, Xizhou Zhu, Chenyu Zhang, Zhe Chen, Jiawen Liu, Wenhai Wang, Lewei Lu, Gao Huang, Yu Qiao 等

The rapid advance of Large Language Models (LLMs) has catalyzed the development of Vision-Language Models (VLMs). Monolithic VLMs, which avoid modality-specific encoders, offer a promising alternative to the compositional ones but face the challenge of inferior performance. Most existing monolithic VLMs require tuning pre-trained LLMs to acquire vision abilities, which may degrade their language capabilities. To address this dilemma, this paper presents a novel high-performance monolithic VLM named HoVLE. We note that LLMs have been shown to be capable of interpreting images when image embeddings are aligned with text embeddings. The challenge for current monolithic VLMs actually lies in the lack of a holistic embedding module for both vision and language inputs. Therefore, HoVLE introduces a holistic embedding module that converts visual and textual inputs into a shared space, allowing LLMs to process images in the same way as texts. Furthermore, a multi-stage training strategy is carefully designed to empower the holistic embedding module. It is first trained to distill visual features from a pre-trained vision encoder and text embeddings from the LLM, enabling large-scale training with unpaired random images and text tokens. The whole model further undergoes next-token prediction on multi-modal data to align the embeddings. Finally, an instruction-tuning stage is incorporated. Our experiments show that HoVLE achieves performance close to leading compositional models on various benchmarks, outperforming previous monolithic models by a large margin.