Can we accurately identify the true correspondences from multimodal datasets containing mismatched data pairs? Existing methods primarily emphasize the similarity matching between the representations of objects across modalities, potentially neglecting the crucial relation consistency within modalities that are particularly important for distinguishing the true and false correspondences. Such an omission often runs the risk of misidentifying negatives as positives, thus leading to unanticipated performance degradation. To address this problem, we propose a general Relation Consistency learning framework, namely ReCon, to accurately discriminate the true correspondences among the multimodal data and thus effectively mitigate the adverse impact caused by mismatches. Specifically, ReCon leverages a novel relation consistency learning to ensure the dual-alignment, respectively of, the cross-modal relation consistency between different modalities and the intra-modal relation consistency within modalities. Thanks to such dual constrains on relations, ReCon significantly enhances its effectiveness for true correspondence discrimination and therefore reliably filters out the mismatched pairs to mitigate the risks of wrong supervisions. Extensive experiments on three widely-used benchmark datasets, including Flickr30K, MS-COCO, and Conceptual Captions, are conducted to demonstrate the effectiveness and superiority of ReCon compared with other SOTAs. The code is available at: https://github.com/qxzha/ReCon.
Local alignment between medical images and text is essential for accurate diagnosis, though it remains challenging due to the absence of natural local pairings and the limitations of rigid region recognition methods. Traditional approaches rely on hard boundaries, which introduce uncertainty, whereas medical imaging demands flexible soft region recognition to handle irregular structures. To overcome these challenges, we propose the Progressive Local Alignment Network (PLAN), which designs a novel contrastive learning-based approach for local alignment to establish meaningful word-pixel relationships and introduces a progressive learning strategy to iteratively refine these relationships, enhancing alignment precision and robustness. By combining these techniques, PLAN effectively improves soft region recognition while suppressing noise interference. Extensive experiments on multiple medical datasets demonstrate that PLAN surpasses state-of-the-art methods in phrase grounding, image-text retrieval, object detection, and zero-shot classification, setting a new benchmark for medical image-text alignment.
We introduce SigLIP 2, a family of new multilingual vision-language encoders that build on the success of the original SigLIP. In this second iteration, we extend the original image-text training objective with several prior, independently developed techniques into a unified recipe -- this includes captioning-based pretraining, self-supervised losses (self-distillation, masked prediction) and online data curation. With these changes, SigLIP 2 models outperform their SigLIP counterparts at all model scales in core capabilities, including zero-shot classification, image-text retrieval, and transfer performance when extracting visual representations for Vision-Language Models (VLMs). Furthermore, the new training recipe leads to significant improvements on localization and dense prediction tasks. We also train variants which support multiple resolutions and preserve the input's native aspect ratio. Finally, we train on a more diverse data-mixture that includes de-biasing techniques, leading to much better multilingual understanding and improved fairness. To allow users to trade off inference cost with performance, we release model checkpoints at four sizes: ViT-B (86M), L (303M), So400m (400M), and g (1B).
Multimodal transformers integrate diverse data types like images, audio, and text, advancing tasks such as audio-visual understanding and image-text retrieval; yet their high parameterization limits deployment on resource-constrained edge devices. Split Learning (SL), which partitions models at a designated cut-layer to offload compute-intensive operations to the server, offers a promising approach for distributed training of multimodal transformers, though its application remains underexplored. We present MPSL, a parallel SL approach for computational efficient fine-tuning of multimodal transformers in a distributed manner, while eliminating label sharing, client synchronization, and per-client sub-model management. MPSL employs lightweight client-side tokenizers and a unified modality-agnostic encoder, allowing flexible adaptation to task-specific needs. Our evaluation across 7 multimodal datasets demonstrates that MPSL matches or outperforms Federated Learning, reduces client-side computations by 250x, and achieves superior scalability in communication cost with model growth. Through extensive analysis, we highlight task suitability, trade-offs, and scenarios where MPSL excels, inspiring further exploration.
Multimodal foundation models (MFMs) have demonstrated significant success in tasks such as visual captioning, question answering, and image-text retrieval. However, these models face inherent limitations due to their finite internal capacity, which restricts their ability to process extended temporal sequences, a crucial requirement for comprehensive video and audio analysis. To overcome these challenges, we introduce a specialized cognitive module, temporal working memory (TWM), which aims to enhance the temporal modeling capabilities of MFMs. It selectively retains task-relevant information across temporal dimensions, ensuring that critical details are preserved throughout the processing of video and audio content. The TWM uses a query-guided attention approach to focus on the most informative multimodal segments within temporal sequences. By retaining only the most relevant content, TWM optimizes the use of the model's limited capacity, enhancing its temporal modeling ability. This plug-and-play module can be easily integrated into existing MFMs. With our TWM, nine state-of-the-art models exhibit significant performance improvements across tasks such as video captioning, question answering, and video-text retrieval. By enhancing temporal modeling, TWM extends the capability of MFMs to handle complex, time-sensitive data effectively. Our code is available at https://github.com/xid32/NAACL_2025_TWM.
Vision-language models (VLMs) align visual and textual representations, enabling high-performance zero-shot classification and image-text retrieval in 2D medical imaging. However, extending VLMs to 3D medical imaging remains computationally challenging. Existing 3D VLMs rely on Vision Transformers (ViTs), which are computationally expensive due to self-attention's quadratic complexity, or 3D convolutions, which demand excessive parameters and FLOPs as kernel size increases. We introduce DCFormer, an efficient 3D medical image encoder that factorizes 3D convolutions into three parallel 1D convolutions along depth, height, and width. This design preserves spatial information while significantly reducing computational cost. Integrated into a CLIP-based vision-language framework, DCFormer is evaluated on CT-RATE, a dataset of 50,188 paired 3D chest CT volumes and radiology reports, for zero-shot multi-abnormality detection across 18 pathologies. Compared to ViT, ConvNeXt, PoolFormer, and TransUNet, DCFormer achieves superior efficiency and accuracy, with DCFormer-Tiny reaching 62.0% accuracy and a 46.3% F1-score while using significantly fewer parameters. These results highlight DCFormer's potential for scalable, clinically deployable 3D medical VLMs. Our codes will be publicly available.
Pretrained visual-language models have made significant advancements in multimodal tasks, including image-text retrieval. However, a major challenge in image-text matching lies in language bias, where models predominantly rely on language priors and neglect to adequately consider the visual content. We thus present Multimodal ASsociation Score (MASS), a framework that reduces the reliance on language priors for better visual accuracy in image-text matching problems. It can be seamlessly incorporated into existing visual-language models without necessitating additional training. Our experiments have shown that MASS effectively lessens language bias without losing an understanding of linguistic compositionality. Overall, MASS offers a promising solution for enhancing image-text matching performance in visual-language models.
Cross-modal retrieval maps data under different modality via semantic relevance. Existing approaches implicitly assume that data pairs are well-aligned and ignore the widely existing annotation noise, i.e., noisy correspondence (NC). Consequently, it inevitably causes performance degradation. Despite attempts that employ the co-teaching paradigm with identical architectures to provide distinct data perspectives, the differences between these architectures are primarily stemmed from random initialization. Thus, the model becomes increasingly homogeneous along with the training process. Consequently, the additional information brought by this paradigm is severely limited. In order to resolve this problem, we introduce a Tripartite learning with Semantic Variation Consistency (TSVC) for robust image-text retrieval. We design a tripartite cooperative learning mechanism comprising a Coordinator, a Master, and an Assistant model. The Coordinator distributes data, and the Assistant model supports the Master model's noisy label prediction with diverse data. Moreover, we introduce a soft label estimation method based on mutual information variation, which quantifies the noise in new samples and assigns corresponding soft labels. We also present a new loss function to enhance robustness and optimize training effectiveness. Extensive experiments on three widely used datasets demonstrate that, even at increasing noise ratios, TSVC exhibits significant advantages in retrieval accuracy and maintains stable training performance.
The development of vision-language models (VLMs) is driven by large-scale and diverse multimodal datasets. However, progress toward generalist biomedical VLMs is limited by the lack of annotated, publicly accessible datasets across biology and medicine. Existing efforts are restricted to narrow domains, missing the full diversity of biomedical knowledge encoded in scientific literature. To address this gap, we introduce BIOMEDICA, a scalable, open-source framework to extract, annotate, and serialize the entirety of the PubMed Central Open Access subset into an easy-to-use, publicly accessible dataset. Our framework produces a comprehensive archive with over 24 million unique image-text pairs from over 6 million articles. Metadata and expert-guided annotations are also provided. We demonstrate the utility and accessibility of our resource by releasing BMCA-CLIP, a suite of CLIP-style models continuously pre-trained on the BIOMEDICA dataset via streaming, eliminating the need to download 27 TB of data locally. On average, our models achieve state-of-the-art performance across 40 tasks - spanning pathology, radiology, ophthalmology, dermatology, surgery, molecular biology, parasitology, and cell biology - excelling in zero-shot classification with a 6.56% average improvement (as high as 29.8% and 17.5% in dermatology and ophthalmology, respectively), and stronger image-text retrieval, all while using 10x less compute. To foster reproducibility and collaboration, we release our codebase and dataset for the broader research community.
Cross-modal (e.g. image-text, video-text) retrieval is an important task in information retrieval and multimodal vision-language understanding field. Temporal understanding makes video-text retrieval more challenging than image-text retrieval. However, we find that the widely used video-text benchmarks have shortcomings in comprehensively assessing abilities of models, especially in temporal understanding, causing large-scale image-text pre-trained models can already achieve comparable zero-shot performance with video-text pre-trained models. In this paper, we introduce RTime, a novel temporal-emphasized video-text retrieval dataset. We first obtain videos of actions or events with significant temporality, and then reverse these videos to create harder negative samples. We then recruit annotators to judge the significance and reversibility of candidate videos, and write captions for qualified videos. We further adopt GPT-4 to extend more captions based on human-written captions. Our RTime dataset currently consists of 21k videos with 10 captions per video, totalling about 122 hours. Based on RTime, we propose three retrieval benchmark tasks: RTime-Origin, RTime-Hard, and RTime-Binary. We further enhance the use of harder-negatives in model training, and benchmark a variety of video-text models on RTime. Extensive experiment analysis proves that RTime indeed poses new and higher challenges to video-text retrieval. We release our RTime dataset\footnote{\url{https://github.com/qyr0403/Reversed-in-Time}} to further advance video-text retrieval and multimodal understanding research.
Contrastive Language-Image Pretraining (CLIP) enables zero-shot inference in downstream tasks such as image-text retrieval and classification. However, recent works extending CLIP suffer from the issue of modality gap, which arises when the image and text embeddings are projected to disparate manifolds, deviating from the intended objective of image-text contrastive learning. We discover that this phenomenon is linked to the modality-specific characteristic that each image/text encoder independently possesses and propose two methods to address the modality gap: (1) a post-hoc embedding standardization method, $\text{I0T}_{\text{post}}$ that reduces the modality gap approximately to zero and (2) a trainable method, $\text{I0T}_{\text{async}}$, to alleviate the modality gap problem by adding two normalization layers for each encoder. Our I0T framework can significantly reduce the modality gap while preserving the original embedding representations of trained models with their locked parameters. In practice, $\text{I0T}_{\text{post}}$ can serve as an alternative explainable automatic evaluation metric of widely used CLIPScore (CLIP-S).
Vision-Language Models (VLMs) implicitly learn to associate image regions with words from large-scale training data, demonstrating an emergent capability for grounding concepts without dense annotations[14,18,51]. However, the coarse-grained supervision from image-caption pairs is often insufficient to resolve ambiguities in object-concept correspondence, even with enormous data volume. Rich semantic and syntactic structures within the text modality have been overlooked as sources of supervision. Starting from contrastive architectures (BLIP and ALBEF) that show strong intrinsic grounding abilities, we propose HIerarchically STructured Learning (HIST). HIST enhances spatial vision-language alignment without using additional human annotations, by hierarchically decomposing captions into the constituent Subjects, Phrases, and Composite Phrases, and enforcing entailment relation between a parent and its children in the hierarchy. Specifically, we introduce two novel loss functions: (1) Subject Loss, which aligns image content with the subject of the corresponding phrase, acting as an entailment of standard contrastive/matching losses at the Phrase level; (2) Composition Loss, to balance attention across multiple objects. HIST is general, and can be applied to any VLM for which attention between vision and language can be computed. Compared to baseline VLMs, HIST achieves up to +9.8% improvement in visual grounding and +6.3% in multi-object referring segmentation. Surprisingly, the improved spatial grounding leads to improvements in other downstream VLM tasks: +1.1% in image-text retrieval, and +0.2% in visual question answering.
Multimodal learning has recently gained significant popularity, demonstrating impressive performance across various zero-shot classification tasks and a range of perceptive and generative applications. Models such as Contrastive Language-Image Pretraining (CLIP) are designed to bridge different modalities, such as images and text, by learning a shared representation space through contrastive learning. Despite their success, the working mechanisms underlying multimodal learning are not yet well understood. Notably, these models often exhibit a modality gap, where different modalities occupy distinct regions within the shared representation space. In this work, we conduct an in-depth analysis of the emergence of modality gap by characterizing the gradient flow learning dynamics. Specifically, we identify the critical roles of mismatched data pairs and a learnable temperature parameter in causing and perpetuating the modality gap during training. Furthermore, our theoretical insights are validated through experiments on practical CLIP models. These findings provide principled guidance for mitigating the modality gap, including strategies such as appropriate temperature scheduling and modality swapping. Additionally, we demonstrate that closing the modality gap leads to improved performance on tasks such as image-text retrieval.
Contrastively-trained Vision-Language Models (VLMs) like CLIP have become the de facto approach for discriminative vision-language representation learning. However, these models have limited language understanding, often exhibiting a "bag of words" behavior. At the same time, Large Vision-Language Models (LVLMs), which combine vision encoders with LLMs, have been shown to be capable of detailed vision-language reasoning, yet their autoregressive nature renders them less suitable for discriminative tasks. In this work, we propose to combine "the best of both worlds": a new training approach for discriminative fine-tuning of LVLMs that results in strong discriminative and compositional capabilities. Essentially, our approach converts a generative LVLM into a discriminative one, unlocking its capability for powerful image-text discrimination combined with enhanced language understanding. Our contributions include (1) a carefully designed training/optimization framework that utilizes image-text pairs of variable length and granularity for training the model with both contrastive and next-token prediction losses. This is accompanied by ablation studies that justify the necessity of our framework's components; (2) a parameter-efficient adaptation method using a combination of soft prompting and LoRA adapters; (3) significant improvements over state-of-the-art CLIP-like models of similar size, including standard image-text retrieval benchmarks and notable gains in compositionality.
Multimodal tasks, such as image-text retrieval and generation, require embedding data from diverse modalities into a shared representation space. Aligning embeddings from heterogeneous sources while preserving shared and modality-specific information is a fundamental challenge. This paper provides an initial attempt to integrate algebraic geometry into multimodal representation learning, offering a foundational perspective for further exploration. We model image and text data as polynomials over discrete rings, \( \mathbb{Z}_{256}[x] \) and \( \mathbb{Z}_{|V|}[x] \), respectively, enabling the use of algebraic tools like fiber products to analyze alignment properties. To accommodate real-world variability, we extend the classical fiber product to an approximate fiber product with a tolerance parameter \( \epsilon \), balancing precision and noise tolerance. We study its dependence on \( \epsilon \), revealing asymptotic behavior, robustness to perturbations, and sensitivity to embedding dimensionality. Additionally, we propose a decomposition of the shared embedding space into orthogonal subspaces, \( Z = Z_s \oplus Z_I \oplus Z_T \), where \( Z_s \) captures shared semantics, and \( Z_I \), \( Z_T \) encode modality-specific features. This decomposition is geometrically interpreted via manifolds and fiber bundles, offering insights into embedding structure and optimization. This framework establishes a principled foundation for analyzing multimodal alignment, uncovering connections between robustness, dimensionality allocation, and algebraic structure. It lays the groundwork for further research on embedding spaces in multimodal learning using algebraic geometry.
Recent advancements in vision-language pre-training via contrastive learning have significantly improved performance across computer vision tasks. However, in the medical domain, obtaining multimodal data is often costly and challenging due to privacy, sensitivity, and annotation complexity. To mitigate data scarcity while boosting model performance, we introduce \textbf{Uni-Mlip}, a unified self-supervision framework specifically designed to enhance medical vision-language pre-training. Uni-Mlip seamlessly integrates cross-modality, uni-modality, and fused-modality self-supervision techniques at the data-level and the feature-level. Additionally, Uni-Mlip tailors uni-modal image self-supervision to accommodate the unique characteristics of medical images. Our experiments across datasets of varying scales demonstrate that Uni-Mlip significantly surpasses current state-of-the-art methods in three key downstream tasks: image-text retrieval, image classification, and visual question answering (VQA).
We present a way to learn novel concepts by only using their textual description. We call this method Knowledge Transfer. Similarly to human perception, we leverage cross-modal interaction to introduce new concepts. We hypothesize that in a pre-trained visual encoder there are enough low-level features already learned (e.g. shape, appearance, color) that can be used to describe previously unknown high-level concepts. Provided with a textual description of the novel concept, our method works by aligning the known low-level features of the visual encoder to its high-level textual description. We show that Knowledge Transfer can successfully introduce novel concepts in multimodal models, in a very efficient manner, by only requiring a single description of the target concept. Our approach is compatible with both separate textual and visual encoders (e.g. CLIP) and shared parameters across modalities. We also show that, following the same principle, Knowledge Transfer can improve concepts already known by the model. Leveraging Knowledge Transfer we improve zero-shot performance across different tasks such as classification, segmentation, image-text retrieval, and captioning.
Medical vision-and-language models (MVLMs) have attracted substantial interest due to their capability to offer a natural language interface for interpreting complex medical data. Their applications are versatile and have the potential to improve diagnostic accuracy and decision-making for individual patients while also contributing to enhanced public health monitoring, disease surveillance, and policy-making through more efficient analysis of large data sets. MVLMS integrate natural language processing with medical images to enable a more comprehensive and contextual understanding of medical images alongside their corresponding textual information. Unlike general vision-and-language models trained on diverse, non-specialized datasets, MVLMs are purpose-built for the medical domain, automatically extracting and interpreting critical information from medical images and textual reports to support clinical decision-making. Popular clinical applications of MVLMs include automated medical report generation, medical visual question answering, medical multimodal segmentation, diagnosis and prognosis and medical image-text retrieval. Here, we provide a comprehensive overview of MVLMs and the various medical tasks to which they have been applied. We conduct a detailed analysis of various vision-and-language model architectures, focusing on their distinct strategies for cross-modal integration/exploitation of medical visual and textual features. We also examine the datasets used for these tasks and compare the performance of different models based on standardized evaluation metrics. Furthermore, we highlight potential challenges and summarize future research trends and directions. The full collection of papers and codes is available at: https://github.com/YtongXie/Medical-Vision-and-Language-Tasks-and-Methodologies-A-Survey.
Methods based on Contrastive Language-Image Pre-training (CLIP) are nowadays extensively used in support of vision-and-language tasks involving remote sensing data, such as cross-modal retrieval. The adaptation of CLIP to this specific domain has relied on model fine-tuning with the standard contrastive objective, using existing human-labeled image-caption datasets, or using synthetic data corresponding to image-caption pairs derived from other annotations over remote sensing images (e.g., object classes). The use of different pre-training mechanisms has received less attention, and only a few exceptions have considered multilingual inputs. This work proposes a novel vision-and-language model for the remote sensing domain, exploring the fine-tuning of a multilingual CLIP model and testing the use of a self-supervised method based on aligning local and global representations from individual input images, together with the standard CLIP objective. Model training relied on assembling pre-existing datasets of remote sensing images paired with English captions, followed by the use of automated machine translation into nine additional languages. We show that translated data is indeed helpful, e.g. improving performance also on English. Our resulting model, which we named Remote Sensing Multilingual CLIP (RS-M-CLIP), obtains state-of-the-art results in a variety of vision-and-language tasks, including cross-modal and multilingual image-text retrieval, or zero-shot image classification.
Multimodal models leverage large-scale pre-training to achieve strong but still imperfect performance on tasks such as image captioning, visual question answering, and cross-modal retrieval. In this paper, we present a simple and efficient method for correcting errors in trained contrastive image-text retrieval models with no additional training, called Nearest Neighbor Normalization (NNN). We show an improvement on retrieval metrics in both text retrieval and image retrieval for all of the contrastive models that we tested (CLIP, BLIP, ALBEF, SigLIP, BEiT) and for both of the datasets that we used (MS-COCO and Flickr30k). NNN requires a reference database, but does not require any training on this database, and can even increase the retrieval accuracy of a model after finetuning.