Implicit neural representations (INRs) model videos as mappings from temporal indices to frames, providing an effective paradigm for video representation and compression. However, existing methods still struggle to effectively align learned representations with the spatial, temporal, and frequency characteristics of target videos. To address this issue, we propose STF-NeRV, a unified Spatial–Temporal-Frequency (STF) Alignment framework. Specifically, it first employs a Multi-Scale Spatial Alignment (MSSA) mechanism to construct spatially consistent latent representations through scale-aware spatial modulation. It then introduces a Temporal Identity Alignment (TIA) mechanism that progressively evolves temporal conditioning from shared temporal structures to timestamp-specific identities, thereby providing conditioning information with increasingly finer granularity for different decoding stages. Finally, an Adaptive Frequency Alignment (AFA) mechanism performs gated sinusoidal refinement to selectively enhance high-frequency details while maintaining structural stability. Experiments demonstrate that STF-NeRV achieves superior reconstruction quality and competitive rate–distortion performance over state-of-the-art implicit video representation methods.
This work investigates the capacity of a pure transformer architecture to perform interactive next-frame prediction in visual sequence modelling. Unlike conventional convolutional or recurrent approaches, the proposed model relies exclusively on self-attention mechanisms, augmented with lightweight conditioning pathways to accommodate user input during inference. To support evaluation, a synthetic benchmarking dataset was generated, capturing gravity-driven motion with controllable interventions. Quantitative comparisons against recurrent and convolutional baselines demonstrate that the transformer achieves superior reconstruction accuracy and perceptual fidelity across multiple spatio-temporal metrics, albeit with increased computational cost. Qualitative assessments further highlight the model’s ability to generalise across regimes and maintain temporal consistency under interactive conditions, though limitations remain in reproducing some dynamic behaviours.
In recent years, the large-scale capture and sharing of image data, together with advances in artificial intelligence, have raised growing concerns about personal privacy and security. To mitigate these risks, numerous face anonymization techniques have been proposed. However, existing methods often suffer from limited image resolution and diversity, as well as inadequate preservation of facial expressions and other non-identity attributes. This paper proposes a face anonymization framework based on a StyleGAN encoder and a skin color matching generative adversarial network (SCMGAN). The anonymization process is decomposed into three stages through a multi-stage training strategy, with each stage focusing on a specific aspect of anonymization. This design enables effective identity suppression while maintaining high face detection rates, visual quality, output diversity, and facial expression fidelity. Extensive experiments show that the proposed method achieves a favorable balance among anonymization effectiveness, detectability, image quality, diversity, and expression preservation compared with recent anonymization approaches.
In recent times, deepfakes have attained super-realistic appearance through use of advanced AI. Although, majority of deepfake research is based on individual audio/video modes for deepfake detection, it has been proven that a multimodal way of handling deepfakes provides a robust and promising solution towards detection. In this work, we present a deepfake detection framework combining visual and auditory modalities for identifying deepfakes. We introduce a convolutional bidirectional long short-term memory model for spatial and temporal information extraction based on video sequences. We also collect spatial information from audio spectrogramic components. Finally, we combine the above in both spatial and temporal domains using a fully connected neural network, to identify deepfakes. Our experimental findings on standard datasets prove that the proposed model generates SOTA detection results, with ablation study to test its different variants operated on diverse modes.
Three-dimensional Gaussian Splatting (3DGS) provides efficient photorealistic rendering, but image-only Structure-from-Motion initialization does not provide metric scale and can be unreliable over long trajectories or in texture-poor regions. We present a systems-level LiDAR–inertial–visual 3DGS pipeline for static, spatially extended scenes. The pipeline combines global multi-sensor poses, occlusion-aware colorized LiDAR initialization, and a hybrid objective comprising photometric, depth, color-gradient, and depth-gradient terms. Unlike approaches centered on new Gaussian primitives or planar/normal regularizers, our contribution focuses on integrating metric SLAM outputs with standard 3DGS and compensating for the photometric mismatch introduced by real LiDAR initialization. Experiments on outdoor and indoor scenes show that LiDAR initialization alone can reduce rendering scores, whereas colorization and hybrid supervision recover this loss and yield modest improvements. The method is intended for applications where metric scale and a deployable multi-sensor reconstruction pipeline are important, rather than as a general replacement for image-only 3DGS.
