Human beings are endowed with a complementary learning system, which bridges the slow learning of general world dynamics with fast storage of episodic memory from a new experience. Previous video generation models, however, primarily focus on slow learning by pre-training on vast amounts of data, overlooking the fast learning phase crucial for episodic memory storage. This oversight leads to inconsistencies across temporally distant frames when generating longer videos, as these frames fall beyond the model's context window. To this end, we introduce SlowFast-VGen, a novel dual-speed learning system for action-driven long video generation. Our approach incorporates a masked conditional video diffusion model for the slow learning of world dynamics, alongside an inference-time fast learning strategy based on a temporal LoRA module. Specifically, the fast learning process updates its temporal LoRA parameters based on local inputs and outputs, thereby efficiently storing episodic memory in its parameters. We further propose a slow-fast learning loop algorithm that seamlessly integrates the inner fast learning loop into the outer slow learning loop, enabling the recall of prior multi-episode experiences for context-aware skill learning. To facilitate the slow learning of an approximate world model, we collect a large-scale dataset of 200k videos with language action annotations, covering a wide range of scenarios. Extensive experiments show that SlowFast-VGen outperforms baselines across various metrics for action-driven video generation, achieving an FVD score of 514 compared to 782, and maintaining consistency in longer videos, with an average of 0.37 scene cuts versus 0.89. The slow-fast learning loop algorithm significantly enhances performances on long-horizon planning tasks as well.

Despite the impressive performance obtained by recent single-image hand modeling techniques, they lack the capability to capture sufficient details of the 3D hand mesh. This deficiency greatly limits their applications when high-fidelity hand modeling is required, e.g., personalized hand modeling. To address this problem, we design a frequency split network to generate 3D hand meshes using different frequency bands in a coarse-to-fine manner. To capture high-frequency personalized details, we transform the 3D mesh into the frequency domain, and proposed a novel frequency decomposition loss to supervise each frequency component. By leveraging such a coarse-to-fine scheme, hand details that correspond to the higher frequency domain can be preserved. In addition, the proposed network is scalable, and can stop the inference at any resolution level to accommodate different hardware with varying computational powers. To feed the scalable frequency network with frequency split image features, we proposed an image-graph ring feature mapping strategy. To train our network with per-vertex supervision, we use a bidirectional registration strategy to generate a topology-fixed ground-truth. To quantitatively evaluate the performance of our method in terms of recovering personalized shape details, we introduce a new evaluation metric named Mean-frequency Signal-to-Noise Ratio (MSNR) to measure the mean signal-to-noise ratio of mesh signal on each frequency component. Extensive experiments demonstrate that our approach generates fine-grained details for high-fidelity 3D hand reconstruction, and our evaluation metric is more effective than traditional metrics for measuring mesh details.

Generative AI has made significant strides in computer vision, particularly in text-driven image/video synthesis (T2I/T2V). Despite the notable advancements, it remains challenging in human-centric content synthesis such as realistic dance generation. Current methodologies, primarily tailored for human motion transfer, encounter difficulties when confronted with real-world dance scenarios (e.g., social media dance), which require to generalize across a wide spectrum of poses and intricate human details. In this paper, we depart from the traditional paradigm of human motion transfer and emphasize two additional critical attributes for the synthesis of human dance content in social media contexts: (i) Generalizability: the model should be able to generalize beyond generic human viewpoints as well as unseen human subjects, backgrounds, and poses; (ii) Compositionality: it should allow for the seamless composition of seen/unseen subjects, backgrounds, and poses from different sources. To address these challenges, we introduce Disco, which includes a novel model architecture with disentangled control to improve the compositionality of dance synthesis, and an effective human attribute pre-training for better generalizability to unseen humans. Extensive qualitative and quantitative results demonstrate that DISCO can generate high-quality human dance images and videos with diverse appearances and flexible motions. Code is available at https://disco-dance.github.io/.

Image diffusion distillation achieves high-fidelity generation with very few sampling steps. However, directly applying these techniques to video models results in unsatisfied frame quality. This issue arises from the limited frame appearance quality in public video datasets, affecting the performance of both teacher and student video diffusion models. Our study aims to improve video diffusion distillation and meanwhile enabling the student model to improve frame appearance using the abundant high-quality image data. To this end, we propose motion consistency models (MCM), a single-stage video diffusion distillation method that disentangles motion and appearance learning. Specifically, MCM involves a video consistency model that distills motion from the video teacher model, and an image discriminator that boosts frame appearance to match high-quality image data. However, directly combining these components leads to two significant challenges: a conflict in frame learning objectives, where video distillation learns from low-quality video frames while the image discriminator targets high-quality images, and training-inference discrepancies due to the differing quality of video samples used during training and inference. To address these challenges, we introduce disentangled motion distillation and mixed trajectory distillation. The former applies the distillation objective solely to the motion representation, while the latter mitigates training-inference discrepancies by mixing distillation trajectories from both the low- and high-quality video domains. Extensive experiments show that our MCM achieves state-of-the-art video diffusion distillation performance. Additionally, our method can enhance frame quality in video diffusion models, producing frames with high aesthetic value or specific styles.

Despite the impressive performance obtained by recent single-image hand modeling techniques, they lack the capability to capture sufficient details of the 3D hand mesh. This deficiency greatly limits their applications when high-fidelity hand modeling is required, e.g., personalized hand modeling. To address this problem, we design a frequency split network to generate 3D hand mesh using different frequency bands in a coarse-to-fine manner. To capture high-frequency personalized details, we transform the 3D mesh into the frequency domain, and propose a novel frequency decomposition loss to supervise each frequency component. By leveraging such a coarse-to-fine scheme, hand details that correspond to the higher frequency domain can be preserved. In addition, the proposed network is scalable, and can stop the inference at any resolution level to accommodate different hardware with varying computational powers. To quantitatively evaluate the performance of our method in terms of recovering personalized shape details, we introduce a new evaluation metric named Mean Signal-to-Noise Ratio (MSNR) to measure the signal-to-noise ratio of each mesh frequency component. Extensive experiments demonstrate that our approach generates fine-grained details for high-fidelity 3D hand reconstruction, and our evaluation metric is more effective for measuring mesh details compared with traditional metrics. The code is available at https://github.com/tyluann/FreqHand.

As advanced image manipulation techniques emerge, detecting the manipulation becomes increasingly important. Despite the success of recent learning-based approaches for image manipulation detection, they typically require expensive pixel-level annotations to train, while exhibiting degraded performance when testing on images that are differently manipulated compared with training images. To address these limitations, we propose weakly-supervised image manipulation detection, such that only binary image-level labels (authentic or tampered with) are required for training purpose. Such a weakly-supervised setting can leverage more training images and has the potential to adapt quickly to new manipulation techniques. To improve the generalization ability, we propose weakly-supervised self-consistency learning (WSCL) to leverage the weakly annotated images. Specifically, two consistency properties are learned: multi-source consistency (MSC) and inter-patch consistency (IPC). MSC exploits different content-agnostic information and enables cross-source learning via an online pseudo label generation and refinement process. IPC performs global pair-wise patch-patch relationship reasoning to discover a complete region of manipulation. Extensive experiments validate that our WSCL, even though is weakly supervised, exhibits competitive performance compared with fully-supervised counterpart under both in-distribution and out-of-distribution evaluations, as well as reasonable manipulation localization ability.

Deep learning models have a risk of utilizing spurious clues to make predictions, such as recognizing actions based on the background scene. This issue can severely degrade the open-set action recognition performance when the testing samples have different scene distributions from the training samples. To mitigate this problem, we propose a novel method, called Scene-debiasing Open-set Action Recognition (SOAR), which features an adversarial scene reconstruction module and an adaptive adversarial scene classification module. The former prevents the decoder from reconstructing the video background given video features, and thus helps reduce the background information in feature learning. The latter aims to confuse scene type classification given video features, with a specific emphasis on the action foreground, and helps to learn scene-invariant information. In addition, we design an experiment to quantify the scene bias. The results indicate that the current open-set action recognizers are biased toward the scene, and our proposed SOAR method better mitigates such bias. Furthermore, our extensive experiments demonstrate that our method outperforms state-of-the-art methods, and the ablation studies confirm the effectiveness of our proposed modules.

Language-guided human motion synthesis has been a challenging task due to the inherent complexity and diversity of human behaviors. Previous methods face limitations in generalization to novel actions, often resulting in unrealistic or incoherent motion sequences. In this paper, we propose ATOM (ATomic mOtion Modeling) to mitigate this problem, by decomposing actions into atomic actions, and employing a curriculum learning strategy to learn atomic action composition. First, we disentangle complex human motions into a set of atomic actions during learning, and then assemble novel actions using the learned atomic actions, which offers better adaptability to new actions. Moreover, we introduce a curriculum learning training strategy that leverages masked motion modeling with a gradual increase in the mask ratio, and thus facilitates atomic action assembly. This approach mitigates the overfitting problem commonly encountered in previous methods while enforcing the model to learn better motion representations. We demonstrate the effectiveness of ATOM through extensive experiments, including text-to-motion and action-to-motion synthesis tasks. We further illustrate its superiority in synthesizing plausible and coherent text-guided human motion sequences.

Weakly-supervised temporal action localization (W-TAL) aims to classify and localize all action instances in untrimmed videos under only video-level supervision. Without frame-level annotations, it is challenging for W-TAL methods to clearly distinguish actions and background, which severely degrades the action boundary localization and action proposal scoring. In this paper, we present an adaptive two-stream consensus network (A-TSCN) to address this problem. Our A-TSCN features an iterative refinement training scheme: a frame-level pseudo ground truth is generated and iteratively updated from a late-fusion activation sequence, and used to provide frame-level supervision for improved model training. Besides, we introduce an adaptive attention normalization loss, which adaptively selects action and background snippets according to video attention distribution. By differentiating the attention values of the selected action snippets and background snippets, it forces the predicted attention to act as a binary selection and promotes the precise localization of action boundaries. Furthermore, we propose a video-level and a snippet-level uncertainty estimator, and they can mitigate the adverse effect caused by learning from noisy pseudo ground truth. Experiments conducted on the THUMOS14, ActivityNet v1.2, ActivityNet v1.3, and HACS datasets show that our A-TSCN outperforms current state-of-the-art methods, and even achieves comparable performance with several fully-supervised methods.

Weakly-supervised Temporal Action Localization (W-TAL) aims at simultaneously classifying and locating all action instances with only video-level supervision. However, current W-TAL methods have two limitations. First, they ignore the difference in video representations between an action instance and its surrounding background when generating and scoring action proposals. Second, the unique characteristics of the RGB frames and optical flow are largely ignored when fusing these two modalities. To address these problems, an Action Coherence Network (ACN) is proposed in this paper. Its core is a new coherence loss which exploits both classification predictions and video content representations to supervise action boundary regression and thus leads to more accurate action localization results. Besides, the proposed ACN explicitly takes into account the specific characteristics of RGB frames and optical flow by training two separate sub-networks, each of which is able to generate modality-specific action proposals independently. Finally, to take advantage of the complementary action proposals generated by two streams, a novel fusion module is introduced to reconcile them and obtain the final action localization results. Experiments on the THUMOS14 and ActivityNet datasets show that our ACN outperforms the state-of-the-art W-TAL methods, and is even comparable to some recent fully-supervised methods. Particularly, ACN achieves a mean average precision of 26.4% on the THUMOS14 dataset under the IoU threshold 0.5.

The lack of automatic tools to identify giant panda makes it hard to keep track of and manage giant pandas in wildlife conservation missions. In this paper, we introduce a new Giant Panda Identification (GPID) task, which aims to identify each individual panda based on an image. Though related to the human re-identification and animal classification problem, GPID is extraordinarily challenging due to subtle visual differences between pandas and cluttered global information. In this paper, we propose a new benchmark dataset iPanda-50 for GPID. The iPanda-50 consists of 6, 874 images from 50 giant panda individuals, and is collected from panda streaming videos. We also introduce a new Feature-Fusion Network with Patch Detector (FFN-PD) for GPID. The proposed FFN-PD exploits the patch detector to detect discriminative local patches without using any part annotations or extra location sub-networks, and builds a hierarchical representation by fusing both global and local features to enhance the inter-layer patch feature interactions. Specifically, an attentional cross-channel pooling is embedded in the proposed FFN-PD to improve the identify-specific patch detectors. Experiments performed on the iPanda-50 datasets demonstrate the proposed FFN-PD significantly outperforms competing methods. Besides, experiments on other fine-grained recognition datasets (i.e., CUB-200-2011, Stanford Cars, and FGVC-Aircraft) demonstrate that the proposed FFN-PD outperforms existing state-of-the-art methods.

Most prominent temporal action localization methods are of the fully-supervised type, which rely heavily on frame-level labels, which could be prohibitively expensive to annotate. Thanks to recent developments on the Weakly-supervised Temporal Action Localization (W-TAL), this alternative paradigm requires only video-level labels in training, alleviating such annotation efforts. Specifically, we present Action Coherence Network (ACN) for W-TAL, which features a new coherence loss that better supervises action boundary learning and facilitate proposal regression. In addition, a purpose-built fusion module is proposed for localization inference based on features extracted by two streams of convolutional neural network. Overall, the proposed ACN achieves state-of-the-art W-TAL performance on two challenging datasets (THU-MOS14 and ActivityNet1.2, particularly ACN attains mAP of 24.2% on THUMOS14 under IoU threshold 0.5), which is approaching some recent fully-supervised TAL methods.