To improve Multimodal Large Language Models' (MLLMs) ability to process images and complex instructions, researchers predominantly curate large-scale visual instruction tuning datasets, which are either sourced from existing vision tasks or synthetically generated using LLMs and image descriptions. However, they often suffer from critical flaws, including misaligned instruction-image pairs and low-quality images. Such issues hinder training efficiency and limit performance improvements, as models waste resources on noisy or irrelevant data with minimal benefit to overall capability. To address this issue, we propose a \textbf{Vi}sual-Centric \textbf{S}election approach via \textbf{A}gents Collaboration (ViSA), which centers on image quality assessment and image-instruction relevance evaluation. Specifically, our approach consists of 1) an image information quantification method via visual agents collaboration to select images with rich visual information, and 2) a visual-centric instruction quality assessment method to select high-quality instruction data related to high-quality images. Finally, we reorganize 80K instruction data from large open-source datasets. Extensive experiments demonstrate that ViSA outperforms or is comparable to current state-of-the-art models on seven benchmarks, using only 2.5\% of the original data, highlighting the efficiency of our data selection approach. Moreover, we conduct ablation studies to validate the effectiveness of each component of our method. The code is available at https://github.com/HITsz-TMG/ViSA.

Deep learning models are usually black boxes when deployed on machine learning platforms. Prior works have shown that the attributes (e.g., the number of convolutional layers) of a target black-box model can be exposed through a sequence of queries. There is a crucial limitation: these works assume the training dataset of the target model is known beforehand and leverage this dataset for model attribute attack. However, it is difficult to access the training dataset of the target black-box model in reality. Therefore, whether the attributes of a target black-box model could be still revealed in this case is doubtful. In this paper, we investigate a new problem of black-box reverse engineering, without requiring the availability of the target model's training dataset. We put forward a general and principled framework DREAM, by casting this problem as out-of-distribution (OOD) generalization. In this way, we can learn a domain-agnostic meta-model to infer the attributes of the target black-box model with unknown training data. This makes our method one of the kinds that can gracefully apply to an arbitrary domain for model attribute reverse engineering with strong generalization ability. Extensive experimental results demonstrate the superiority of our proposed method over the baselines.

Recent advancements in generative models have significantly enhanced talking face video generation, yet singing video generation remains underexplored. The differences between human talking and singing limit the performance of existing talking face video generation models when applied to singing. The fundamental differences between talking and singing-specifically in audio characteristics and behavioral expressions-limit the effectiveness of existing models. We observe that the differences between singing and talking audios manifest in terms of frequency and amplitude. To address this, we have designed a multi-scale spectral module to help the model learn singing patterns in the spectral domain. Additionally, we develop a spectral-filtering module that aids the model in learning the human behaviors associated with singing audio. These two modules are integrated into the diffusion model to enhance singing video generation performance, resulting in our proposed model, SINGER. Furthermore, the lack of high-quality real-world singing face videos has hindered the development of the singing video generation community. To address this gap, we have collected an in-the-wild audio-visual singing dataset to facilitate research in this area. Our experiments demonstrate that SINGER is capable of generating vivid singing videos and outperforms state-of-the-art methods in both objective and subjective evaluations.

World models integrate raw data from various modalities, such as images and language to simulate comprehensive interactions in the world, thereby displaying crucial roles in fields like mixed reality and robotics. Yet, applying the world model for accurate video prediction is quite challenging due to the complex and dynamic intentions of the various scenes in practice. In this paper, inspired by the human rethinking process, we decompose the complex video prediction into four meta-tasks that enable the world model to handle this issue in a more fine-grained manner. Alongside these tasks, we introduce a new benchmark named Embodied Video Anticipation Benchmark (EVA-Bench) to provide a well-rounded evaluation. EVA-Bench focused on evaluating the video prediction ability of human and robot actions, presenting significant challenges for both the language model and the generation model. Targeting embodied video prediction, we propose the Embodied Video Anticipator (EVA), a unified framework aiming at video understanding and generation. EVA integrates a video generation model with a visual language model, effectively combining reasoning capabilities with high-quality generation. Moreover, to enhance the generalization of our framework, we tailor-designed a multi-stage pretraining paradigm that adaptatively ensembles LoRA to produce high-fidelity results. Extensive experiments on EVA-Bench highlight the potential of EVA to significantly improve performance in embodied scenes, paving the way for large-scale pre-trained models in real-world prediction tasks.

Abstract--Recent advancements in Multimodal Large Language Models (MLLMs) underscore the significance of scalable models and data to boost performance, yet this often incurs substantial computational costs. Although the Mixture of Experts (MoE) architecture has been employed to efficiently scale large language and image-text models, these efforts typically involve fewer experts and limited modalities. To address this, our work presents the pioneering attempt to develop a unified MLLM with the MoE architecture, named Uni-MoE that can handle a wide array of modalities. Specifically, it features modality-specific encoders with connectors for a unified multimodal representation. We also implement a sparse MoE architecture within the LLMs to enable efficient training and inference through modality-level data parallelism and expert-level model parallelism. To enhance the multi-expert collaboration and generalization, we present a progressive training strategy: 1) Cross-modality alignment using various connectors with different cross-modality data, 2) Training modality-specific experts with cross-modality instruction data to activate experts' preferences, and 3) Tuning the Uni-MoE framework utilizing Low-Rank Adaptation (LoRA) on mixed multimodal instruction data. We evaluate the instruction-tuned Uni-MoE on a comprehensive set of multimodal datasets. The extensive experimental results demonstrate Uni-MoE's principal advantage of significantly reducing performance bias in handling mixed multimodal datasets, alongside improved multi-expert collaboration and generalization. Additionally, there is a growing trend [6], [7], [8], [9] toward building a unified MLLM that could comprehend more modalities such as video, audio, and speech, moving beyond the traditional imagetext paradigm. To catch up with superior closed-source MLLMs like GPT-4V [10] and Gemini [11], the main efforts of open-source community contain enlarging model sizes [12], as seen with the expansion of vision foundation models to 6 billion parameters [12] and the integration with 70B Large Language models (LLMs) [13], [14], and enhancing instruction tuning with diverse multimodal datasets [3], [15], [16]. These developments underscore the increasing ability of MLLMs to process and reason across multiple modalities, showing the importance of both model scalability and the expansion of multimodal instructional data.

We aim at exploiting additional auxiliary labels from an independent (auxiliary) task to boost the primary task performance which we focus on, while preserving a single task inference cost of the primary task. While most existing auxiliary learning methods are optimization-based relying on loss weights/gradients manipulation, our method is architecture-based with a flexible asymmetric structure for the primary and auxiliary tasks, which produces different networks for training and inference. Specifically, starting from two single task networks/branches (each representing a task), we propose a novel method with evolving networks where only primary-to-auxiliary links exist as the cross-task connections after convergence. These connections can be removed during the primary task inference, resulting in a single-task inference cost. We achieve this by formulating a Neural Architecture Search (NAS) problem, where we initialize bi-directional connections in the search space and guide the NAS optimization converging to an architecture with only the single-side primary-to-auxiliary connections. Moreover, our method can be incorporated with optimization-based auxiliary learning approaches. In this paper, we tackle the practical issue of auxiliary learning, which involves improving the performance of a specific task (i.e., the primary task) while incorporating additional auxiliary labels from different tasks (i.e., the auxiliary tasks). We aim to efficiently leverage these auxiliary labels to enhance the primary task's performance while maintaining a comparable computational and parameter cost to a single-task network when evaluating the primary task.

In this paper, we propose a new setting for generating product descriptions from images, augmented by marketing keywords. It leverages the combined power of visual and textual information to create descriptions that are more tailored to the unique features of products. For this setting, previous methods utilize visual and textual encoders to encode the image and keywords and employ a language model-based decoder to generate the product description. However, the generated description is often inaccurate and generic since same-category products have similar copy-writings, and optimizing the overall framework on large-scale samples makes models concentrate on common words yet ignore the product features. To alleviate the issue, we present a simple and effective Multimodal In-Context Tuning approach, named ModICT, which introduces a similar product sample as the reference and utilizes the in-context learning capability of language models to produce the description. During training, we keep the visual encoder and language model frozen, focusing on optimizing the modules responsible for creating multimodal in-context references and dynamic prompts. This approach preserves the language generation prowess of large language models (LLMs), facilitating a substantial increase in description diversity. To assess the effectiveness of ModICT across various language model scales and types, we collect data from three distinct product categories within the E-commerce domain. Extensive experiments demonstrate that ModICT significantly improves the accuracy (by up to 3.3% on Rouge-L) and diversity (by up to 9.4% on D-5) of generated results compared to conventional methods. Our findings underscore the potential of ModICT as a valuable tool for enhancing automatic generation of product descriptions in a wide range of applications. Code is at: https://github.com/HITsz-TMG/Multimodal-In-Context-Tuning

Deep learning models are usually black boxes when deployed on machine learning platforms. Prior works have shown that the attributes ($e.g.$, the number of convolutional layers) of a target black-box neural network can be exposed through a sequence of queries. There is a crucial limitation: these works assume the dataset used for training the target model to be known beforehand and leverage this dataset for model attribute attack. However, it is difficult to access the training dataset of the target black-box model in reality. Therefore, whether the attributes of a target black-box model could be still revealed in this case is doubtful. In this paper, we investigate a new problem of Domain-agnostic Reverse Engineering the Attributes of a black-box target Model, called DREAM, without requiring the availability of the target model's training dataset, and put forward a general and principled framework by casting this problem as an out of distribution (OOD) generalization problem. In this way, we can learn a domain-agnostic model to inversely infer the attributes of a target black-box model with unknown training data. This makes our method one of the kinds that can gracefully apply to an arbitrary domain for model attribute reverse engineering with strong generalization ability. Extensive experimental studies are conducted and the results validate the superiority of our proposed method over the baselines.

In active visual tracking, it is notoriously difficult when distracting objects appear, as distractors often mislead the tracker by occluding the target or bringing a confusing appearance. To address this issue, we propose a mixed cooperative-competitive multi-agent game, where a target and multiple distractors form a collaborative team to play against a tracker and make it fail to follow. Through learning in our game, diverse distracting behaviors of the distractors naturally emerge, thereby exposing the tracker's weakness, which helps enhance the distraction-robustness of the tracker. For effective learning, we then present a bunch of practical methods, including a reward function for distractors, a cross-modal teacher-student learning strategy, and a recurrent attention mechanism for the tracker. The experimental results show that our tracker performs desired distraction-robust active visual tracking and can be well generalized to unseen environments. We also show that the multi-agent game can be used to adversarially test the robustness of trackers.

This paper proposes a new approach to multi-object tracking by semantic topic discovery. We dynamically cluster frame-by-frame detections and treat objects as topics, allowing the application of the Dirichlet Process Mixture Model (DPMM). The tracking problem is cast as a topic-discovery task where the video sequence is treated analogously to a document. This formulation addresses tracking issues such as object exclusivity constraints as well as cannot-link constraints which are integrated without the need for heuristic thresholds. The video is temporally segmented into epochs to model the dynamics of word (superpixel) co-occurrences and to model the temporal damping effect. In experiments on public data sets we demonstrate the effectiveness of the proposed algorithm.