Large Vision-Language Models (LVLMs) can reason effectively over both textual and visual inputs, but they tend to hallucinate syntactically coherent yet visually ungrounded contents. In this paper, we investigate the internal dynamics of hallucination by examining the tokens logits rankings throughout the generation process, revealing three key patterns in how LVLMs process information: (1) gradual visual information loss – visually grounded tokens gradually become less favored throughout generation, and (2) early excitation – semantically meaningful tokens achieve peak activation in the layers earlier than the final layer. (3) hidden genuine information – visually grounded tokens though not being eventually decided still retain relatively high rankings at inference. Based on these insights, we propose VISTA (Visual Information Steering with Token-logit Augmentation), a training-free inference-time intervention framework that reduces hallucination while promoting genuine information. VISTA works by combining two complementary approaches: reinforcing visual information in activation space and leveraging early layer activations to promote semantically meaningful decoding. Compared to existing methods, VISTA requires no external supervision and is applicable to various decoding strategies. Extensive experiments show that VISTA on average reduces hallucination by about 40% on evaluated open-ended generation task, and it consistently outperforms existing methods on four benchmarks across four architectures under three decoding strategies. Code is available at https://github.com/LzVv123456/VISTA.

In generative modeling, tokenization simplifies complex data into compact, structured representations, creating a more efficient, learnable space. For high-dimensional visual data, it reduces redundancy and emphasizes key features for high-quality generation. Current visual tokenization methods rely on a traditional autoencoder framework, where the encoder compresses data into latent representations, and the decoder reconstructs the original input. In this work, we offer a new perspective by proposing denoising as decoding, shifting from single-step reconstruction to iterative refinement. Specifically, we replace the decoder with a diffusion process that iteratively refines noise to recover the original image, guided by the latents provided by the encoder. We evaluate our approach by assessing both reconstruction (rFID) and generation quality (FID), comparing it to state-of-the-art autoencoding approaches. By adopting iterative reconstruction through diffusion, our autoencoder, namely Epsilon-VAE, achieves high reconstruction quality, which in turn enhances downstream generation quality by 22% at the same compression rates or provides 2.3x inference speedup through increasing compression rates. We hope this work offers new insights into integrating iterative generation and autoencoding for improved compression and generation.

The recent advance in vision-language models is largely attributed to the abundance of image-text data. We aim to replicate this success for video-language models, but there simply is not enough human-curated video-text data available. We thus resort to fine-tuning a video-language model from a strong image-language baseline with syn-thesized instructional data. The resulting video model by video-instruction-tuning (VIIT) is then used to auto-label millions of videos to generate high-quality captions. We show the adapted video-language model performs well on a wide range of video-language benchmarks. For instance, it surpasses the best prior result on open-ended NExT-QA by 2.8%. Besides, our model generates detailed descriptions for previously unseen videos, which provide better textual supervision than existing methods. Experiments show that a video-language dual-encoder model contrastively trained on these auto-generated captions is 3.8% better than the strongest baseline that also leverages vision-language models. Our best model outperforms state-of-the-art methods on MSR-VTT zero-shot text-to-video retrieval by 6%. As a side product, we generate the largest video capation dataset to date.

The recent progress in language-based open-vocabulary object detection can be largely attributed to finding better ways of leveraging large-scale data with free-form text annotations. Training such models with a discriminative objective function has proven successful, but requires good positive and negative samples. However, the free-form nature and the open vocabulary of object descriptions make the space of negatives extremely large. Prior works randomly sample negatives or use rule-based techniques to build them. In contrast, we propose to leverage the vast knowledge built into modern generative models to automatically build negatives that are more relevant to the original data. Specifically, we use large-Language-models to generate negative text descriptions, and text-to-image diffusion models to also generate corresponding negative images. Our experimental analysis confirms the relevance of the generated negative data, and its use in language-based detectors improves performance on two complex benchmarks. Code is available at https://github.com/xiaofeng94/Gen-Enhanced-Negs.

Recent studies have shown promising performance in open-vocabulary object detection (OVD) by utilizing pseudo labels (PLs) from pre-trained vision and language models (VLMs). However, teacher-student self-training, a powerful and widely used paradigm to leverage PLs, is rarely explored for OVD. This work identifies two challenges of using self-training in OVD: noisy PLs from VLMs and frequent distribution changes of PLs. To address these challenges, we propose SAS-Det that tames self-training for OVD from two key perspectives. First, we present a split-and-fusion (SAF) head that splits a standard detection into an open-branch and a closed-branch. This design can reduce noisy supervision from pseudo boxes. More-over, the two branches learn complementary knowledge from different training data, significantly enhancing performance when fused together. Second, in our view, un-like in closed-set tasks, the PL distributions in OVD are solely determined by the teacher model. We introduce a periodic update strategy to decrease the number of up-dates to the teacher, thereby decreasing the frequency of changes in PL distributions, which stabilizes the training process. Extensive experiments demonstrate SAS-Det is both efficient and effective. SAS-Det outperforms recent models of the same scale by a clear margin and achieves 37.4 AP50and 29.1 APron novel categories of the COCO and LVIS benchmarks, respectively. Code is available at https://github.com/xiaofeng94/SAS-Det.

Existing video-language pre-training methods primarily focus on instance-level alignment between video clips and captions via global contrastive learning but neglect rich fine-grained local information in both videos and text, which is of importance to downstream tasks requiring temporal localization and semantic reasoning. A powerful model is expected to be capable of capturing region-object correspondences and recognizing scene changes in a video clip, reflecting spatial and temporal granularity, respectively. To strengthen model's understanding into such fine-grained details, we propose a simple yet effective video-language modeling framework, S-ViLM, by exploiting the intrinsic structures of these two modalities. It includes two novel designs, inter-clip spatial grounding and intra-clip temporal grouping, to promote learning region-object alignment and temporal-aware features, simultaneously. Comprehensive evaluations demonstrate that S-ViLM performs favorably against existing approaches in learning more expressive representations. Specifically, S-ViLM surpasses the state-of-the-art methods substantially on four representative downstream tasks, covering text-video retrieval, video question answering, video action recognition, and temporal action localization.

We introduce VideoPrism, a general-purpose video encoder that tackles diverse video understanding tasks with a single frozen model. We pretrain VideoPrism on a heterogeneous corpus containing 36M high-quality video-caption pairs and 582M video clips with noisy parallel text (e.g., ASR transcripts). The pretraining approach improves upon masked autoencoding by global-local distillation of semantic video embeddings and a token shuffling scheme, enabling VideoPrism to focus primarily on the video modality while leveraging the invaluable text associated with videos. We extensively test VideoPrism on four broad groups of video understanding tasks, from web video question answering to CV for science, achieving state-of-the-art performance on 31 out of 33 video understanding benchmarks.

In the context of continual learning, prototypes—as representative class embeddings—offer advantages in memory conservation and the mitigation of catastrophic forgetting. However, challenges related to semantic drift and prototype interference persist. In this study, we introduce the Contrastive Prototypical Prompt (CPP) approach. Through task-specific prompt-tuning, underpinned by a contrastive learning objective, we effectively address both aforementioned challenges. Our evaluations on four challenging class-incremental benchmarks reveal that CPP achieves a significant 4% to 6% improvement over state-of-the-art methods. Importantly, CPP operates without a rehearsal buffer and narrows the performance divergence between continual and offline joint-learning, suggesting an innovative scheme for Transformer-based continual learning systems1.

We are concerned with a worst-case scenario in model generalization, in the sense that a model aims to perform well on many unseen domains while there is only one single domain available for training. We propose Meta-Learning based Adversarial Domain Augmentation to solve this Out-of-Domain generalization problem. The key idea is to leverage adversarial training to create "fictitious" yet "challenging" populations, from which a model can learn to generalize with theoretical guarantees. To facilitate fast and desirable domain augmentation, we cast the model training in a meta-learning scheme and use a Wasserstein Auto-Encoder to relax the widely used worst-case constraint. We further improve our method by integrating uncertainty quantification for efficient domain generalization. Extensive experiments on multiple benchmark datasets indicate its superior performance in tackling single domain generalization.

We are concerned with a challenging scenario in unpaired multiview video learning. In this case, the model aims to learn comprehensive multiview representations while the cross-view semantic information exhibits variations. We propose Semantics-based Unpaired Multiview Learning (SUM-L) to tackle this unpaired multiview learning problem. The key idea is to build cross-view pseudopairs and do view-invariant alignment by leveraging the semantic information of videos. To facilitate the data efficiency of multiview learning, we further perform video-text alignment for first-person and third-person videos, to fully leverage the semantic knowledge to improve video representations. Extensive experiments on multiple benchmark datasets verify the effectiveness of our framework. Our method also outperforms multiple existing view-alignment methods, under the more challenging scenario than typical paired or unpaired multimodal or multiview learning. Our code is available at https://github.com/wqtwjt1996/SUM-L.

This work focuses on training a single visual relationship detector predicting over the union of label spaces from multiple datasets. Merging labels spanning different datasets could be challenging due to inconsistent taxonomies. The issue is exacerbated in visual relationship detection when second-order visual semantics are introduced between pairs of objects. To address this challenge, we propose UniVRD, a novel bottom-up method for Unified Visual Relationship Detection by leveraging vision and language models (VLMs). VLMs provide well-aligned image and text embeddings, where similar relationships are optimized to be close to each other for semantic unification. Our bottom-up design enables the model to enjoy the benefit of training with both object detection and visual relationship datasets. Empirical results on both human-object interaction detection and scene-graph generation demonstrate the competitive performance of our model. UniVRD achieves 38.07 mAP on HICO-DET, outperforming the current best bottom-up HOI detector by 14.26 mAP. More importantly, we show that our unified detector performs as well as dataset-specific models in mAP, and achieves further improvements when we scale up the model. Our code will be made publicly available on GitHub1.

Cross-modal attention mechanisms have been widely applied to the image-text matching task. They have achieved remarkable improvements thanks to their capability of learning fine-grained relevance across different modalities. However, the cross-modal attention models of existing methods could be sub-optimal and inaccurate because there is no direct supervision provided during the training process. In this work, we propose two novel training strategies, namely Contrastive Content Resourcing (CCR) and Contrastive Content Swapping (CCS) constraints, to address such limitations. These constraints supervise the training of cross-modal attention models in a contrastive learning manner without requiring explicit attention annotations. They are plug-in training strategies and can be generally integrated into existing cross-modal attention models. Additionally, we introduce three metrics, including Attention Precision, Recall, and F1-Score, to quantitatively measure the quality of learned attention models. We evaluate the proposed constraints by incorporating them into four state- of-the-art cross-modal attention-based image-text matching models. Experimental results on both Flickr30k and MS-COCO datasets demonstrate that integrating these constraints generally improves the model performance in terms of both retrieval performance and attention metrics.

Hierarchical structures are popular in recent vision transformers, however, they require sophisticated designs and massive datasets to work well. In this paper, we explore the idea of nesting basic local transformers on non-overlapping image blocks and aggregating them in a hierarchical way. We find that the block aggregation function plays a critical role in enabling cross-block non-local information communication. This observation leads us to design a simplified architecture that requires minor code changes upon the original vision transformer. The benefits of the proposed judiciously-selected design are threefold: (1) NesT converges faster and requires much less training data to achieve good generalization on both ImageNet and small datasets like CIFAR; (2) when extending our key ideas to image generation, NesT leads to a strong decoder that is 8 times faster than previous transformer-based generators; and (3) we show that decoupling the feature learning and abstraction processes via this nested hierarchy in our design enables constructing a novel method (named GradCAT) for visually interpreting the learned model. Source code is available https://github.com/google-research/nested-transformer.

Multimodal data collected from the real world are often imperfect due to missing modalities. Therefore multimodal models that are robust against modal-incomplete data are highly preferred. Recently, Transformer models have shown great success in processing multimodal data. However, existing work has been limited to either architecture designs or pre-training strategies; whether Transformer models are naturally robust against missing-modal data has rarely been investigated. In this paper, we present the first-of-its-kind work to comprehensively investigate the behavior of Transformers in the presence of modal-incomplete data. Unsurprising, we find Transformer models are sensitive to missing modalities while different modal fusion strategies will significantly affect the robustness. What surprised us is that the optimal fusion strategy is dataset dependent even for the same Transformer model; there does not exist a universal strategy that works in general cases. Based on these findings, we propose a principle method to improve the robustness of Transformer models byautomatically searching for an optimal fusion strategy regarding input data. Experimental validations on three benchmarks support the superior performance of the proposed method.

Optical flow estimation is a fundamental task in computer vision. Recent direct-regression methods using deep neural networks achieve remarkable performance improvement. However, they do not explicitly capture long-term motion correspondences and thus cannot handle large motions effectively. In this paper, inspired by the traditional matching-optimization methods where matching is introduced to handle large displacements before energy-based optimizations, we introduce a simple but effective global matching step before the direct regression and develop a learning-based matching-optimization framework, namely GMFlowNet. In GMFlowNet, global matching is efficiently calculated by applying argmax on 4D cost volumes. Additionally, to improve the matching quality, we propose patch-based overlapping attention to extract large context features. Extensive experiments demonstrate that GM-FlowNet outperforms RAFT, the most popular optimization-only method, by a large margin and achieves state-of-the-art performance on standard benchmarks. Thanks to the matching and overlapping attention, GMFlowNet obtains major improvements on the predictions for textureless regions and large motions. Our code is made publicly available at https://github.com/xiaofeng94/GMFlowNet.

Abstract 3D facial animation synthesis from audio has been a focus in recent years. However, most existing literature works are designed to map audio and visual content, providing limited knowledge regarding the relationship between emotion in audio and expressive facial animation. This work generates audio‐matching facial animations with the specified emotion label. In such a task, we argue that separating the content from audio is indispensable—the proposed model must learn to generate facial content from audio content while expressions from the specified emotion. We achieve it by an adaptive instance normalization module that isolates the content in the audio and combines the emotion embedding from the specified label. The joint content‐emotion embedding is then used to generate 3D facial vertices and texture maps. We compare our method with state‐of‐the‐art baselines, including the facial segmentation‐based and voice conversion‐based disentanglement approaches. We also conduct a user study to evaluate the performance of emotion conditioning. The results indicate that our proposed method outperforms the baselines in animation quality and expression categorization accuracy.

A common assumption in multimodal learning is the completeness of training data, i.e., full modalities are available in all training examples. Although there exists research endeavor in developing novel methods to tackle the incompleteness of testing data, e.g., modalities are partially missing in testing examples, few of them can handle incomplete training modalities. The problem becomes even more challenging if considering the case of severely missing, e.g., ninety percent of training examples may have incomplete modalities. For the first time in the literature, this paper formally studies multimodal learning with missing modality in terms of flexibility (missing modalities in training, testing, or both) and efficiency (most training data have incomplete modality). Technically, we propose a new method named SMIL that leverages Bayesian meta-learning in uniformly achieving both objectives. To validate our idea, we conduct a series of experiments on three popular benchmarks: MM-IMDb, CMU-MOSI, and avMNIST. The results prove the state-of-the-art performance of SMIL over existing methods and generative baselines including autoencoders and generative adversarial networks.

We introduce a novel representation learning method to disentangle pose-dependent as well as view-dependent factors from 2D human poses. The method trains a network using cross-view mutual information maximization (CV-MIM) which maximizes mutual information of the same pose performed from different viewpoints in a contrastive learning manner. We further propose two regularization terms to ensure disentanglement and smoothness of the learned representations. The resulting pose representations can be used for cross-view action recognition.To evaluate the power of the learned representations, in addition to the conventional fully-supervised action recognition settings, we introduce a novel task called single-shot cross-view action recognition. This task trains models with actions from only one single viewpoint while models are evaluated on poses captured from all possible viewpoints. We evaluate the learned representations on standard benchmarks for action recognition, and show that (i) CV-MIM performs competitively compared with the state-of-the-art models in the fully-supervised scenarios; (ii) CV-MIM outperforms other competing methods by a large margin in the single-shot cross-view setting; (iii) and the learned representations can significantly boost the performance when reducing the amount of supervised training data. Our code is made publicly available at https://github.com/google-research/google-research/tree/master/poem.

Attention-based models, exemplified by the Transformer, can effectively model long range dependency, but suffer from the quadratic complexity of self-attention operation, making them difficult to be adopted for high-resolution image generation based on Generative Adversarial Networks (GANs). In this paper, we introduce two key ingredients to Transformer to address this challenge. First, in low-resolution stages of the generative process, standard global self-attention is replaced with the proposed multi-axis blocked self-attention which allows efficient mixing of local and global attention. Second, in high-resolution stages, we drop self-attention while only keeping multi-layer perceptrons reminiscent of the implicit neural function. To further improve the performance, we introduce an additional self-modulation component based on cross-attention. The resulting model, denoted as HiT, has a nearly linear computational complexity with respect to the image size and thus directly scales to synthesizing high definition images. We show in the experiments that the proposed HiT achieves state-of-the-art FID scores of 30.83 and 2.95 on unconditional ImageNet $128 \times 128$ and FFHQ $256 \times 256$, respectively, with a reasonable throughput. We believe the proposed HiT is an important milestone for generators in GANs which are completely free of convolutions. Our code is made publicly available at https://github.com/google-research/hit-gan.

Cross-modal knowledge distillation deals with transferring knowledge from a model trained with superior modalities (Teacher) to another model trained with weak modalities (Student). Existing approaches require paired training examples exist in both modalities. However, accessing the data from superior modalities may not always be feasible. For example, in the case of 3D hand pose estimation, depth maps, point clouds, or stereo images usually capture better hand structures than RGB images, but most of them are expensive to be collected. In this paper, we propose a novel scheme to train the Student in a Target dataset where the Teacher is unavailable. Our key idea is to generalize the distilled cross-modal knowledge learned from a Source dataset, which contains paired examples from both modalities, to the Target dataset by modeling knowledge as priors on parameters of the Student. We name our method "Cross-Modal Knowledge Generalization" and demonstrate that our scheme results in competitive performance for 3D hand pose estimation on standard benchmark datasets.

We are concerned with a worst-case scenario in model generalization, in the sense that a model aims to perform well on many unseen domains while there is only one single domain available for training. We propose a new method named adversarial domain augmentation to solve this Out-of-Distribution (OOD) generalization problem. The key idea is to leverage adversarial training to create "fictitious" yet "challenging" populations, from which a model can learn to generalize with theoretical guarantees. To facilitate fast and desirable domain augmentation, we cast the model training in a meta-learning scheme and use a Wasserstein Auto-Encoder (WAE) to relax the widely used worst-case constraint. Detailed theoretical analysis is provided to testify our formulation, while extensive experiments on multiple benchmark datasets indicate its superior performance in tackling single domain generalization.

Adversarial data augmentation has shown promise for training robust deep neural networks against unforeseen data shifts or corruptions. However, it is difficult to define heuristics to generate effective fictitious target distributions containing "hard" adversarial perturbations that are largely different from the source distribution. In this paper, we propose a novel and effective regularization term for adversarial data augmentation. We theoretically derive it from the information bottleneck principle, which results in a maximum-entropy formulation. Intuitively, this regularization term encourages perturbing the underlying source distribution to enlarge predictive uncertainty of the current model, so that the generated "hard" adversarial perturbations can improve the model robustness during training. Experimental results on three standard benchmarks demonstrate that our method consistently outperforms the existing state of the art by a statistically significant margin.

In this paper, we study the problem of learning Graph Convolutional Networks (GCNs) for regression. Current architectures of GCNs are limited to the small receptive field of convolution filters and shared transformation matrix for each node. To address these limitations, we propose Semantic Graph Convolutional Networks (SemGCN), a novel neural network architecture that operates on regression tasks with graph-structured data. SemGCN learns to capture semantic information such as local and global node relationships, which is not explicitly represented in the graph. These semantic relationships can be learned through end-to-end training from the ground truth without additional supervision or hand-crafted rules. We further investigate applying SemGCN to 3D human pose regression. Our formulation is intuitive and sufficient since both 2D and 3D human poses can be represented as a structured graph encoding the relationships between joints in the skeleton of a human body. We carry out comprehensive studies to validate our method. The results prove that SemGCN outperforms state of the art while using 90% fewer parameters.

Graph kernels are kernel methods measuring graph similarity and serve as a standard tool for graph classification. However, the use of kernel methods for node classification, which is a related problem to graph representation learning, is still ill-posed and the state-of-the-art methods are heavily based on heuristics. Here, we present a novel theoretical kernel-based framework for node classification that can bridge the gap between these two representation learning problems on graphs. Our approach is motivated by graph kernel methodology but extended to learn the node representations capturing the structural information in a graph. We theoretically show that our formulation is as powerful as any positive semidefinite kernels. To efficiently learn the kernel, we propose a novel mechanism for node feature aggregation and a data-driven similarity metric employed during the training phase. More importantly, our framework is flexible and complementary to other graph-based deep learning models, e.g., Graph Convolutional Networks (GCNs). We empirically evaluate our approach on a number of standard node classification benchmarks, and demonstrate that our model sets the new state of the art.

Generating multi-view images from a single-view input is an important yet challenging problem. It has broad applications in vision, graphics, and robotics. Our study indicates that the widely-used generative adversarial network (GAN) may learn ?incomplete? representations due to the single-pathway framework: an encoder-decoder network followed by a discriminator network.We propose CR-GAN to address this problem. In addition to the single reconstruction path, we introduce a generation sideway to maintain the completeness of the learned embedding space. The two learning paths collaborate and compete in a parameter-sharing manner, yielding largely improved generality to ?unseen? dataset. More importantly, the two-pathway framework makes it possible to combine both labeled and unlabeled data for self-supervised learning, which further enriches the embedding space for realistic generations. We evaluate our approach on a wide range of datasets. The results prove that CR-GAN significantly outperforms state-of-the-art methods, especially when generating from ?unseen? inputs in wild conditions.

Hierarchical segmentation based object proposal methods have become an important step in modern object detection paradigm. However, standard single-way hierarchical methods are fundamentally flawed in that the errors in early steps cannot be corrected and accumulate. In this work, we propose a novel multi-branch hierarchical segmentation approach that alleviates such problems by learning multiple merging strategies in each step in a complementary manner, such that errors in one merging strategy could be corrected by the others. Our approach achieves the state-of-the-art performance for both object proposal and object detection tasks, comparing to previous object proposal methods.