Recent large-scale video datasets have facilitated the generation of diverse open-domain videos of Video Diffusion Models (VDMs). Nonetheless, the efficacy of VDMs in assimilating complex knowledge from these datasets remains constrained by their inherent scale, leading to suboptimal comprehension and synthesis of numerous actions. In this paper, we introduce EchoReel, a novel approach to augment the capability of VDMs in generating intricate actions by emulating motions from pre-existing videos, which are readily accessible from databases or online repositories. EchoReel seamlessly integrates with existing VDMs, enhancing their ability to produce realistic motions without compromising their fundamental capabilities. Specifically, the Action Prism (AP), is introduced to distill motion information from reference videos, which requires training on only a small dataset. Leveraging the knowledge from pre-trained VDMs, EchoReel incorporates new action features into VDMs through the additional layers, eliminating the need for any further fine-tuning of untrained actions. Extensive experiments demonstrate that EchoReel is not merely replicating the whole content from references, and it significantly improves the generation of realistic actions, even in situations where existing VDMs might directly fail.
Instruction tuning represents a prevalent strategy employed by Multimodal Large Language Models (MLLMs) to align with human instructions and adapt to new tasks. Nevertheless, MLLMs encounter the challenge of adapting to users' evolving knowledge and demands. Therefore, how to retain existing skills while acquiring new knowledge needs to be investigated. In this paper, we present a comprehensive benchmark, namely Continual Instruction tuNing (CoIN), to assess existing MLLMs in the sequential instruction tuning paradigm. CoIN comprises 10 commonly used datasets spanning 8 task categories, ensuring a diverse range of instructions and tasks. Besides, the trained model is evaluated from two aspects: Instruction Following and General Knowledge, which assess the alignment with human intention and knowledge preserved for reasoning, respectively. Experiments on CoIN demonstrate that current powerful MLLMs still suffer catastrophic forgetting, and the failure in intention alignment assumes the main responsibility, instead of the knowledge forgetting. To this end, we introduce MoELoRA to MLLMs which is effective to retain the previous instruction alignment. Experimental results consistently illustrate the forgetting decreased from this method on CoIN.
The video composition task aims to integrate specified foregrounds and backgrounds from different videos into a harmonious composite. Current approaches, predominantly trained on videos with adjusted foreground color and lighting, struggle to address deep semantic disparities beyond superficial adjustments, such as domain gaps. Therefore, we propose a training-free pipeline employing a pre-trained diffusion model imbued with semantic prior knowledge, which can process composite videos with broader semantic disparities. Specifically, we process the video frames in a cascading manner and handle each frame in two processes with the diffusion model. In the inversion process, we propose Balanced Partial Inversion to obtain generation initial points that balance reversibility and modifiability. Then, in the generation process, we further propose Inter-Frame Augmented attention to augment foreground continuity across frames. Experimental results reveal that our pipeline successfully ensures the visual harmony and inter-frame coherence of the outputs, demonstrating efficacy in managing broader semantic disparities.
Unbiased Scene Graph Generation (USGG) aims to address biased predictions in SGG. To that end, data transfer methods are designed to convert coarse-grained predicates into fine-grained ones, mitigating imbalanced distribution. However, them overlook contextual relevance between transferred labels and subject-object pairs, such as unsuitability of 'eating' for 'woman-table'. Furthermore, they typically involve a two-stage process with significant computational costs, starting with pre-training a model for data transfer, followed by training from scratch using transferred labels. Thus, we introduce a plug-and-play method named CITrans, which iteratively trains SGG models with progressively enhanced data. First, we introduce Context-Restricted Transfer (CRT), which imposes subject-object constraints within predicates' semantic space to achieve fine-grained data transfer. Subsequently, Efficient Iterative Learning (EIL) iteratively trains models and progressively generates enhanced labels which are consistent with model's learning state, thereby accelerating the training process. Finally, extensive experiments show that CITrans achieves state-of-the-art and results with high efficiency.
The goal of Universal Cross-Domain Retrieval (UCDR) is to achieve robust performance in generalized test scenarios, wherein data may belong to strictly unknown domains and categories during training. Recently, pre-trained models with prompt tuning have shown strong generalization capabilities and attained noteworthy achievements in various downstream tasks, such as few-shot learning and video-text retrieval. However, applying them directly to UCDR may not sufficiently to handle both domain shift (i.e., adapting to unfamiliar domains) and semantic shift (i.e., transferring to unknown categories). To this end, we propose Prompting-to-Simulate (ProS), the first method to apply prompt tuning for UCDR. ProS employs a two-step process to simulate Content-aware Dynamic Prompts (CaDP) which can impact models to produce generalized features for UCDR. Concretely, in Prompt Units Learning stage, we introduce two Prompt Units to individually capture domain and semantic knowledge in a mask-and-align way. Then, in Context-aware Simulator Learning stage, we train a Content-aware Prompt Simulator under a simulated test scenarios to produce the corresponding CaDP. Extensive experiments conducted on three benchmark datasets show that our method achieves new state-of-the-art performance without bringing excessive parameters. Our method is publicly available at https://anonymous.4open.science/r/ProS
Recently Text-to-Video (T2V) synthesis has undergone a breakthrough by training transformers or diffusion models on large-scale datasets. Nevertheless, inferring such large models incurs huge costs.Previous inference acceleration works either require costly retraining or are model-specific.To address this issue, instead of retraining we explore the inference process of two mainstream T2V models using transformers and diffusion models.The exploration reveals the redundancy in temporal attention modules of both models, which are commonly utilized to establish temporal relations among frames.Consequently, we propose a training-free and generalized pruning strategy called F3-Pruning to prune redundant temporal attention weights.Specifically, when aggregate temporal attention values are ranked below a certain ratio, corresponding weights will be pruned.Extensive experiments on three datasets using a classic transformer-based model CogVideo and a typical diffusion-based model Tune-A-Video verify the effectiveness of F3-Pruning in inference acceleration, quality assurance and broad applicability.
Story Visualization aims to generate images aligned with story prompts, reflecting the coherence of storybooks through visual consistency among characters and scenes.Whereas current approaches exclusively concentrate on characters and neglect the visual consistency among contextually correlated scenes, resulting in independent character images without inter-image coherence.To tackle this issue, we propose a new presentation form for Story Visualization called Storyboard, inspired by film-making, as illustrated in Fig.1.Specifically, a Storyboard unfolds a story into visual representations scene by scene. Within each scene in Storyboard, characters engage in activities at the same location, necessitating both visually consistent scenes and characters.For Storyboard, we design a general framework coined as Make-A-Storyboard that applies disentangled control over the consistency of contextual correlated characters and scenes and then merge them to form harmonized images.Extensive experiments demonstrate 1) Effectiveness.the effectiveness of the method in story alignment, character consistency, and scene correlation; 2) Generalization. Our method could be seamlessly integrated into mainstream Image Customization methods, empowering them with the capability of story visualization.
Catastrophic Forgetting (CF) is a prominent issue in continual learning. Parameter isolation addresses this challenge by masking a sub-network for each task to mitigate interference with old tasks. However, these sub-networks are constructed relying on weight magnitude, which does not necessarily correspond to the importance of weights, resulting in maintaining unimportant weights and constructing redundant sub-networks. To overcome this limitation, inspired by information bottleneck, which removes redundancy between adjacent network layers, we propose \textbf{\underline{I}nformation \underline{B}ottleneck \underline{M}asked sub-network (IBM)} to eliminate redundancy within sub-networks. Specifically, IBM accumulates valuable information into essential weights to construct redundancy-free sub-networks, not only effectively mitigating CF by freezing the sub-networks but also facilitating new tasks training through the transfer of valuable knowledge. Additionally, IBM decomposes hidden representations to automate the construction process and make it flexible. Extensive experiments demonstrate that IBM consistently outperforms state-of-the-art methods. Notably, IBM surpasses the state-of-the-art parameter isolation method with a 70\% reduction in the number of parameters within sub-networks and an 80\% decrease in training time.
Zero-shot Text-to-Video synthesis generates videos based on prompts without any videos. Without motion information from videos, motion priors implied in prompts are vital guidance. For example, the prompt "airplane landing on the runway" indicates motion priors that the "airplane" moves downwards while the "runway" stays static. Whereas the motion priors are not fully exploited in previous approaches, thus leading to two nontrivial issues: 1) the motion variation pattern remains unaltered and prompt-agnostic for disregarding motion priors; 2) the motion control of different objects is inaccurate and entangled without considering the independent motion priors of different objects. To tackle the two issues, we propose a prompt-adaptive and disentangled motion control strategy coined as MotionZero, which derives motion priors from prompts of different objects by Large-Language-Models and accordingly applies motion control of different objects to corresponding regions in disentanglement. Furthermore, to facilitate videos with varying degrees of motion amplitude, we propose a Motion-Aware Attention scheme which adjusts attention among frames by motion amplitude. Extensive experiments demonstrate that our strategy could correctly control motion of different objects and support versatile applications including zero-shot video edit.
In weakly supervised video anomaly detection (WVAD), where only video-level labels indicating the presence or absence of abnormal events are available, the primary challenge arises from the inherent ambiguity in temporal annotations of abnormal occurrences. Inspired by the statistical insight that temporal features of abnormal events often exhibit outlier characteristics, we propose a novel method, BN-WVAD, which incorporates BatchNorm into WVAD. In the proposed BN-WVAD, we leverage the Divergence of Feature from Mean vector (DFM) of BatchNorm as a reliable abnormality criterion to discern potential abnormal snippets in abnormal videos. The proposed DFM criterion is also discriminative for anomaly recognition and more resilient to label noise, serving as the additional anomaly score to amend the prediction of the anomaly classifier that is susceptible to noisy labels. Moreover, a batch-level selection strategy is devised to filter more abnormal snippets in videos where more abnormal events occur. The proposed BN-WVAD model demonstrates state-of-the-art performance on UCF-Crime with an AUC of 87.24%, and XD-Violence, where AP reaches up to 84.93%. Our code implementation is accessible at https://github.com/cool-xuan/BN-WVAD.