Ziyong Feng

Xiang An, Jiankang Deng, Jia Guo et al.

Learning discriminative deep feature embeddings by using million-scale in-the-wild datasets and margin-based softmax loss is the current state-of-the-art approach for face recognition. However, the memory and computing cost of the Fully Connected (FC) layer linearly scales up to the number of identities in the training set. Besides, the large-scale training data inevitably suffers from inter-class conflict and long-tailed distribution. In this paper, we propose a sparsely updating variant of the FC layer, named Partial FC (PFC). In each iteration, positive class centers and a random subset of negative class centers are selected to compute the margin-based softmax loss. All class centers are still maintained throughout the whole training process, but only a subset is selected and updated in each iteration. Therefore, the computing requirement, the probability of inter-class conflict, and the frequency of passive update on tail class centers, are dramatically reduced. Extensive experiments across different training data and backbones (e.g. CNN and ViT) confirm the effectiveness, robustness and efficiency of the proposed PFC. The source code is available at \https://github.com/deepinsight/insightface/tree/master/recognition.

Kaicheng Yang, Jiankang Deng, Xiang An et al.

Contrastive Language-Image Pre-training (CLIP) has significantly boosted the performance of various vision-language tasks by scaling up the dataset with image-text pairs collected from the web. However, the presence of intrinsic noise and unmatched image-text pairs in web data can potentially affect the performance of representation learning. To address this issue, we first utilize the OFA model to generate synthetic captions that focus on the image content. The generated captions contain complementary information that is beneficial for pre-training. Then, we propose an Adaptive Language-Image Pre-training (ALIP), a bi-path model that integrates supervision from both raw text and synthetic caption. As the core components of ALIP, the Language Consistency Gate (LCG) and Description Consistency Gate (DCG) dynamically adjust the weights of samples and image-text/caption pairs during the training process. Meanwhile, the adaptive contrastive loss can effectively reduce the impact of noise data and enhances the efficiency of pre-training data. We validate ALIP with experiments on different scales of models and pre-training datasets. Experiments results show that ALIP achieves state-of-the-art performance on multiple downstream tasks including zero-shot image-text retrieval and linear probe. To facilitate future research, the code and pre-trained models are released at https://github.com/deepglint/ALIP.

Xiang An, Jiankang Deng, Kaicheng Yang et al.

Modern image retrieval methods typically rely on fine-tuning pre-trained encoders to extract image-level descriptors. However, the most widely used models are pre-trained on ImageNet-1K with limited classes. The pre-trained feature representation is therefore not universal enough to generalize well to the diverse open-world classes. In this paper, we first cluster the large-scale LAION400M into one million pseudo classes based on the joint textual and visual features extracted by the CLIP model. Due to the confusion of label granularity, the automatically clustered dataset inevitably contains heavy inter-class conflict. To alleviate such conflict, we randomly select partial inter-class prototypes to construct the margin-based softmax loss. To further enhance the low-dimensional feature representation, we randomly select partial feature dimensions when calculating the similarities between embeddings and class-wise prototypes. The dual random partial selections are with respect to the class dimension and the feature dimension of the prototype matrix, making the classification conflict-robust and the feature embedding compact. Our method significantly outperforms state-of-the-art unsupervised and supervised image retrieval approaches on multiple benchmarks. The code and pre-trained models are released to facilitate future research https://github.com/deepglint/unicom.

Qian Zhang, Xiangzi Dai, Ninghua Yang et al.

VAR is a new generation paradigm that employs 'next-scale prediction' as opposed to 'next-token prediction'. This innovative transformation enables auto-regressive (AR) transformers to rapidly learn visual distributions and achieve robust generalization. However, the original VAR model is constrained to class-conditioned synthesis, relying solely on textual captions for guidance. In this paper, we introduce VAR-CLIP, a novel text-to-image model that integrates Visual Auto-Regressive techniques with the capabilities of CLIP. The VAR-CLIP framework encodes captions into text embeddings, which are then utilized as textual conditions for image generation. To facilitate training on extensive datasets, such as ImageNet, we have constructed a substantial image-text dataset leveraging BLIP2. Furthermore, we delve into the significance of word positioning within CLIP for the purpose of caption guidance. Extensive experiments confirm VAR-CLIP's proficiency in generating fantasy images with high fidelity, textual congruence, and aesthetic excellence. Our project page are https://github.com/daixiangzi/VAR-CLIP

Xiang An, Kaicheng Yang, Xiangzi Dai et al.

Contrastive Language Image Pre-training (CLIP) has recently demonstrated success across various tasks due to superior feature representation empowered by image-text contrastive learning. However, the instance discrimination method used by CLIP can hardly encode the semantic structure of training data. To handle this limitation, cluster discrimination has been proposed through iterative cluster assignment and classification. Nevertheless, most cluster discrimination approaches only define a single pseudo-label for each image, neglecting multi-label signals in the image. In this paper, we propose a novel Multi-Label Cluster Discrimination method named MLCD to enhance representation learning. In the clustering step, we first cluster the large-scale LAION-400M dataset into one million centers based on off-the-shelf embedding features. Considering that natural images frequently contain multiple visual objects or attributes, we select the multiple closest centers as auxiliary class labels. In the discrimination step, we design a novel multi-label classification loss, which elegantly separates losses from positive classes and negative classes, and alleviates ambiguity on decision boundary. We validate the proposed multi-label cluster discrimination method with experiments on different scales of models and pre-training datasets. Experimental results show that our method achieves state-of-the-art performance on multiple downstream tasks including linear probe, zero-shot classification, and image-text retrieval. Code and models have been released at https://github.com/deepglint/unicom .

Tiancheng Gu, Kaicheng Yang, Chaoyi Zhang et al.

After pre-training on extensive image-text pairs, Contrastive Language-Image Pre-training (CLIP) demonstrates promising performance on a wide variety of benchmarks. However, a substantial volume of multimodal interleaved documents remains underutilized for contrastive vision-language representation learning. To fully leverage these unpaired documents, we initially establish a Real-World Data Extraction pipeline to extract high-quality images and texts. Then we design a hierarchical retrieval method to efficiently associate each image with multiple semantically relevant realistic texts. To further enhance fine-grained visual information, we propose an image semantic augmented generation module for synthetic text production. Furthermore, we employ a semantic balance sampling strategy to improve dataset diversity, enabling better learning of long-tail concepts. Based on these innovations, we construct RealSyn, a dataset combining realistic and synthetic texts, available in three scales: 15M, 30M, and 100M. We compare our dataset with other widely used datasets of equivalent scale for CLIP training. Models pre-trained on RealSyn consistently achieve state-of-the-art performance across various downstream tasks, including linear probe, zero-shot transfer, zero-shot robustness, and zero-shot retrieval. Furthermore, extensive experiments confirm that RealSyn significantly enhances contrastive vision-language representation learning and demonstrates robust scalability. To facilitate future research, the RealSyn dataset and pretrained model weights are released at https://github.com/deepglint/RealSyn.

Yin Xie, Kaicheng Yang, Xiang An et al.

Learning visual representations is foundational for a broad spectrum of downstream tasks. Although recent vision-language contrastive models, such as CLIP and SigLIP, have achieved impressive zero-shot performance via large-scale vision-language alignment, their reliance on global representations constrains their effectiveness for dense prediction tasks, such as grounding, OCR, and segmentation. To address this gap, we introduce Region-Aware Cluster Discrimination (RICE), a novel method that enhances region-level visual and OCR capabilities. We first construct a billion-scale candidate region dataset and propose a Region Transformer layer to extract rich regional semantics. We further design a unified region cluster discrimination loss that jointly supports object and OCR learning within a single classification framework, enabling efficient and scalable distributed training on large-scale data. Extensive experiments show that RICE consistently outperforms previous methods on tasks, including segmentation, dense detection, and visual perception for Multimodal Large Language Models (MLLMs). The pre-trained models have been released at https://github.com/deepglint/MVT.

Yin Xie, Kaicheng Yang, Peirou Liang et al.

Large Multimodal Models (LMMs) often face a modality representation gap during pretraining: while language embeddings remain stable, visual representations are highly sensitive to contextual noise (e.g., background clutter). To address this issue, we introduce a visual comprehension stage, which we call ViCToR (Visual Comprehension via Token Reconstruction), a novel pretraining framework for LMMs. ViCToR employs a learnable visual token pool and utilizes the Hungarian matching algorithm to select semantically relevant tokens from this pool for visual token replacement. Furthermore, by integrating a visual token reconstruction loss with dense semantic supervision, ViCToR can learn tokens which retain high visual detail, thereby enhancing the large language model's (LLM's) understanding of visual information. After pretraining on 3 million publicly accessible images and captions, ViCToR achieves state-of-the-art results, improving over LLaVA-NeXT-8B by 10.4%, 3.2%, and 7.2% on the MMStar, SEED$^I$, and RealWorldQA benchmarks, respectively. Code is available at https://github.com/deepglint/Victor.

Xiaoxing Hu, Kaicheng Yang, Ziyang Gong et al.

The original CLIP text encoder is limited by a maximum input length of 77 tokens, which hampers its ability to effectively process long texts and perform fine-grained semantic understanding. In addition, the CLIP text encoder lacks support for multilingual inputs. All these limitations significantly restrict its applicability across a broader range of tasks. Recent studies have attempted to replace the CLIP text encoder with an LLM-based embedder to enhance its ability in processing long texts, multilingual understanding, and fine-grained semantic comprehension. However, because the representation spaces of LLMs and the vision-language space of CLIP are pretrained independently without alignment priors, direct alignment using contrastive learning can disrupt the intrinsic vision-language alignment in the CLIP image encoder, leading to an underutilization of the knowledge acquired during pre-training. To address this challenge, we propose ProCLIP, a curriculum learning-based progressive vision-language alignment framework to effectively align the CLIP image encoder with an LLM-based embedder. Specifically, ProCLIP first distills knowledge from CLIP's text encoder into the LLM-based embedder to leverage CLIP's rich pretrained knowledge while establishing initial alignment between the LLM embedder and CLIP image encoder. Subsequently, ProCLIP further aligns the CLIP image encoder with the LLM-based embedder through image-text contrastive tuning, employing self-distillation regularization to avoid overfitting. To achieve a more effective alignment, instance semantic alignment loss and embedding structure alignment loss are employed during representation inheritance and contrastive tuning. The Code is available at https://github.com/VisionXLab/ProCLIP.

Zhumei Wang, Zechen Hu, Ruoxi Guo et al.

Recovering absolute human motion from monocular inputs is challenging due to two main issues. First, existing methods depend on 3D training data collected from limited environments, constraining out-of-distribution generalization. The second issue is the difficulty of estimating metric-scale poses from monocular input. To address these challenges, we introduce Mocap-2-to-3, a novel framework that performs multi-view lifting from monocular input by leveraging 2D data pre-training, enabling the reconstruction of metrically accurate 3D motions with absolute positions. To leverage abundant 2D data, we decompose complex 3D motion into multi-view syntheses. We first pretrain a single-view diffusion model on extensive 2D datasets, then fine-tune a multi-view model using public 3D data to enable view-consistent motion generation from monocular input, allowing the model to acquire action priors and diversity through 2D data. Furthermore, to recover absolute poses, we propose a novel human motion representation that decouples the learning of local pose and global movements, while encoding geometric priors of the ground to accelerate convergence. This enables progressive recovery of motion in absolute space during inference. Experimental results on in-the-wild benchmarks demonstrate that our method surpasses state-of-the-art approaches in both camera-space motion realism and world-grounded human positioning, while exhibiting superior generalization capability. Our code will be made publicly available.

Yiwen Huang, Rihui Wu, Pinglai Ou et al.

Modern deep networks generally implement a certain form of shortcut connections to alleviate optimization difficulties. However, we observe that such network topology alters the nature of deep networks. In many ways, these networks behave similarly to aggregated wide networks. We thus exploit the aggregation nature of shortcut connections at a finer architectural level and place them within wide convolutional layers. We end up with a sequentially aggregated convolutional (SeqConv) layer that combines the benefits of both wide and deep representations by aggregating features of various depths in sequence. The proposed SeqConv serves as a drop-in replacement of regular wide convolutional layers and thus could be handily integrated into any backbone network. We apply SeqConv to widely adopted backbones including ResNet and ResNeXt, and conduct experiments for image classification on public benchmark datasets. Our ResNet based network with a model size of ResNet-50 easily surpasses the performance of the 2.35$\times$ larger ResNet-152, while our ResNeXt based model sets a new state-of-the-art accuracy on ImageNet classification for networks with similar model complexity. The code and pre-trained models of our work are publicly available at https://github.com/GroupOfAlchemists/SeqConv.

Tiancheng Gu, Kaicheng Yang, Ziyong Feng et al.

The Contrastive Language-Image Pre-training (CLIP) framework has become a widely used approach for multimodal representation learning, particularly in image-text retrieval and clustering. However, its efficacy is constrained by three key limitations: (1) text token truncation, (2) isolated image-text encoding, and (3) deficient compositionality due to bag-of-words behavior. While recent Multimodal Large Language Models (MLLMs) have demonstrated significant advances in generalized vision-language understanding, their potential for learning transferable multimodal representations remains underexplored.In this work, we present UniME (Universal Multimodal Embedding), a novel two-stage framework that leverages MLLMs to learn discriminative representations for diverse downstream tasks. In the first stage, we perform textual discriminative knowledge distillation from a powerful LLM-based teacher model to enhance the embedding capability of the MLLMś language component. In the second stage, we introduce hard negative enhanced instruction tuning to further advance discriminative representation learning. Specifically, we initially mitigate false negative contamination and then sample multiple hard negatives per instance within each batch, forcing the model to focus on challenging samples. This approach not only improves discriminative power but also enhances instruction-following ability in downstream tasks. We conduct extensive experiments on the MMEB benchmark and multiple retrieval tasks, including short and long caption retrieval and compositional retrieval. Results demonstrate that UniME achieves consistent performance improvement across all tasks, exhibiting superior discriminative and compositional capabilities.

Siying Cui, Jia Guo, Xiang An et al.

Leveraging Stable Diffusion for the generation of personalized portraits has emerged as a powerful and noteworthy tool, enabling users to create high-fidelity, custom character avatars based on their specific prompts. However, existing personalization methods face challenges, including test-time fine-tuning, the requirement of multiple input images, low preservation of identity, and limited diversity in generated outcomes. To overcome these challenges, we introduce IDAdapter, a tuning-free approach that enhances the diversity and identity preservation in personalized image generation from a single face image. IDAdapter integrates a personalized concept into the generation process through a combination of textual and visual injections and a face identity loss. During the training phase, we incorporate mixed features from multiple reference images of a specific identity to enrich identity-related content details, guiding the model to generate images with more diverse styles, expressions, and angles compared to previous works. Extensive evaluations demonstrate the effectiveness of our method, achieving both diversity and identity fidelity in generated images.

Jiaxing Chen, Yuxuan Liu, Dehu Li et al.

The rise of Multimodal Large Language Models (MLLMs), renowned for their advanced instruction-following and reasoning capabilities, has significantly propelled the field of visual reasoning. However, due to limitations in their image tokenization processes, most MLLMs struggle to capture fine details of text and objects in images, especially in high-resolution samples. To overcome this limitation, we introduce P2G, a novel framework for plug-and-play grounding in MLLMs. P2G utilizes the tool-usage potential of MLLMs to employ expert agents for on-the-fly grounding of reasoning into critical visual and textual elements in images, thereby enabling deliberate reasoning through multimodal prompting. Additionally, we develop P2GB, a benchmark designed to evaluate MLLMs' proficiency in understanding inter-object relationships and textual content in challenging high-resolution images. Extensive experiments on visual reasoning tasks demonstrate the superiority of P2G, achieving performance comparable to GPT-4V on P2GB with a 7B backbone. Our work underscores the potential of grounding reasoning with external agents in MLLMs, presenting a promising alternative to mere model scaling.

Tiancheng Gu, Kaicheng Yang, Kaichen Zhang et al.

Universal multimodal embedding models are foundational to various tasks. Existing approaches typically employ in-batch negative mining by measuring the similarity of query-candidate pairs. However, these methods often struggle to capture subtle semantic differences among candidates and lack diversity in negative samples. Moreover, the embeddings exhibit limited discriminative ability in distinguishing false and hard negatives. In this paper, we leverage the advanced understanding capabilities of MLLMs to enhance representation learning and present a novel Universal Multimodal Embedding (UniME-V2) model. Our approach first constructs a potential hard negative set through global retrieval. We then introduce the MLLM-as-a-Judge mechanism, which utilizes MLLMs to assess the semantic alignment of query-candidate pairs and generate soft semantic matching scores. These scores serve as a foundation for hard negative mining, mitigating the impact of false negatives and enabling the identification of diverse, high-quality hard negatives. Furthermore, the semantic matching scores are used as soft labels to mitigate the rigid one-to-one mapping constraint. By aligning the similarity matrix with the soft semantic matching score matrix, the model learns semantic distinctions among candidates, significantly enhancing its discriminative capacity. To further improve performance, we propose UniME-V2-Reranker, a reranking model trained on our mined hard negatives through a joint pairwise and listwise optimization approach. We conduct comprehensive experiments on the MMEB benchmark and multiple retrieval tasks, demonstrating that our method achieves state-of-the-art performance on average across all tasks.

Zhuoyao Zhong, Lianwen Jin, Shuye Zhang et al.

In this paper, we develop a novel unified framework called DeepText for text region proposal generation and text detection in natural images via a fully convolutional neural network (CNN). First, we propose the inception region proposal network (Inception-RPN) and design a set of text characteristic prior bounding boxes to achieve high word recall with only hundred level candidate proposals. Next, we present a powerful textdetection network that embeds ambiguous text category (ATC) information and multilevel region-of-interest pooling (MLRP) for text and non-text classification and accurate localization. Finally, we apply an iterative bounding box voting scheme to pursue high recall in a complementary manner and introduce a filtering algorithm to retain the most suitable bounding box, while removing redundant inner and outer boxes for each text instance. Our approach achieves an F-measure of 0.83 and 0.85 on the ICDAR 2011 and 2013 robust text detection benchmarks, outperforming previous state-of-the-art results.

Zecheng Xie, Zenghui Sun, Lianwen Jin et al.

This paper proposes an end-to-end framework, namely fully convolutional recurrent network (FCRN) for handwritten Chinese text recognition (HCTR). Unlike traditional methods that rely heavily on segmentation, our FCRN is trained with online text data directly and learns to associate the pen-tip trajectory with a sequence of characters. FCRN consists of four parts: a path-signature layer to extract signature features from the input pen-tip trajectory, a fully convolutional network to learn informative representation, a sequence modeling layer to make per-frame predictions on the input sequence and a transcription layer to translate the predictions into a label sequence. The FCRN is end-to-end trainable in contrast to conventional methods whose components are separately trained and tuned. We also present a refined beam search method that efficiently integrates the language model to decode the FCRN and significantly improve the recognition results. We evaluate the performance of the proposed method on the test sets from the databases CASIA-OLHWDB and ICDAR 2013 Chinese handwriting recognition competition, and both achieve state-of-the-art performance with correct rates of 96.40% and 95.00%, respectively.

Jie Xu, Lianwen Jin, Lingyu Liang et al.

This paper proposes a deep leaning method to address the challenging facial attractiveness prediction problem. The method constructs a convolutional neural network of facial beauty prediction using a new deep cascaded fine-turning scheme with various face inputting channels, such as the original RGB face image, the detail layer image, and the lighting layer image. With a carefully designed CNN model of deep structure, large input size and small convolutional kernels, we have achieved a high prediction correlation of 0.88. This result convinces us that the problem of facial attractiveness prediction can be solved by deep learning approach, and it also shows the important roles of the facial smoothness, lightness, and color information that were involved in facial beauty perception, which is consistent with the result of recent psychology studies. Furthermore, we analyze the high-level features learnt by CNN through visualization of its hidden layers, and some interesting phenomena were observed. It is found that the contours and appearance of facial features, especially eyes and moth, are the most significant facial attributes for facial attractiveness prediction, which is also consistent with the visual perception intuition of human.

Weixin Yang, Lianwen Jin, Zecheng Xie et al.

Deep convolutional neural networks (DCNNs) have achieved great success in various computer vision and pattern recognition applications, including those for handwritten Chinese character recognition (HCCR). However, most current DCNN-based HCCR approaches treat the handwritten sample simply as an image bitmap, ignoring some vital domain-specific information that may be useful but that cannot be learnt by traditional networks. In this paper, we propose an enhancement of the DCNN approach to online HCCR by incorporating a variety of domain-specific knowledge, including deformation, non-linear normalization, imaginary strokes, path signature, and 8-directional features. Our contribution is twofold. First, these domain-specific technologies are investigated and integrated with a DCNN to form a composite network to achieve improved performance. Second, the resulting DCNNs with diversity in their domain knowledge are combined using a hybrid serial-parallel (HSP) strategy. Consequently, we achieve a promising accuracy of 97.20% and 96.87% on CASIA-OLHWDB1.0 and CASIA-OLHWDB1.1, respectively, outperforming the best results previously reported in the literature.

Weixin Yang, Lianwen Jin, Dacheng Tao et al.

Inspired by the theory of Leitners learning box from the field of psychology, we propose DropSample, a new method for training deep convolutional neural networks (DCNNs), and apply it to large-scale online handwritten Chinese character recognition (HCCR). According to the principle of DropSample, each training sample is associated with a quota function that is dynamically adjusted on the basis of the classification confidence given by the DCNN softmax output. After a learning iteration, samples with low confidence will have a higher probability of being selected as training data in the next iteration; in contrast, well-trained and well-recognized samples with very high confidence will have a lower probability of being involved in the next training iteration and can be gradually eliminated. As a result, the learning process becomes more efficient as it progresses. Furthermore, we investigate the use of domain-specific knowledge to enhance the performance of DCNN by adding a domain knowledge layer before the traditional CNN. By adopting DropSample together with different types of domain-specific knowledge, the accuracy of HCCR can be improved efficiently. Experiments on the CASIA-OLHDWB 1.0, CASIA-OLHWDB 1.1, and ICDAR 2013 online HCCR competition datasets yield outstanding recognition rates of 97.33%, 97.06%, and 97.51% respectively, all of which are significantly better than the previous best results reported in the literature.