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.
Generative steganography can directly synthesize images that carry secret messages, thereby avoiding explicit cover modifications. This property makes detecting generative steganography more challenging under semantic shift. To address this problem, we propose a semantic-invariant steganalysis framework via trace disentanglement and prototype-guided meta-learning. Specifically, a dual-view disentanglement pipeline is designed to learn steganalysis related trace features by separating steganographic traces from semantic content through self- and cross-reconstruction. Furthermore, multiple prototypes are introduced as cover reference points in the trace feature space, which help describe the mixed cover distribution and provide complementary evidence for cover and stego discrimination. In addition, semantic domain meta-learning is used to simulate semantic shifts with asymmetric prototype constraint during training. Extensive experiments are conducted on datasets built from several generative steganography methods. Experiments show that SITraNet achieves the highest average ACCs of 91.11% for cross-semantic evaluation within the same steganography scheme and 90.76% with unseen steganography schemes.
With the rapid growth of Internet and multimedia content, image steganography has become important for secure information transmission. However, existing methods still struggle to balance visual quality, embedding efficiency, security, and robustness. To address this issue, we propose MSFD-Stega, a training-free generative image steganography framework based on Stable Diffusion. The proposed method embeds secret information into the latent space through a multi-scale frequency domain injection strategy, selectively modifying frequency subbands to preserve visual fidelity and semantic consistency. By leveraging DDIM deterministic inversion and fast sampling, MSFD-Stega enables efficient and reversible information embedding and extraction without additional model training. Experiments on multiple datasets show that the proposed method achieves a favorable trade-off among image fidelity, extraction accuracy, embedding capacity, and efficiency, while maintaining competitive robustness and relatively low detectability under the evaluated settings.
Open Set Recognition (OSR) requires a model to classify known classes while simultaneously identifying unknown samples not encountered during training. Existing generative OSR approaches synthesize pseudo open set samples to help establish decision boundaries; however, the highly variable quality of these generated samples often introduces identity ambiguity, leading to classifier prediction bias and error accumulation. To overcome this challenge, we incorporate meta learning into OSR, formulating the dynamic sample selection process as an adaptive meta learning process that effectively blocks error accumulation. Specifically, we propose a Meta Open Set Filtering (MOF) framework, which uses a meta network that integrates the initial, current and historical trends of predictions to dynamically filter out high quality pseudo open set candidates. Through a bi-level optimization strategy guided by a small meta dataset of authentic open set samples, MOF autonomously calibrates pseudo labels and extends the closed set classifier to directly output open set probabilities. Extensive experiments on benchmark datasets demonstrate that MOF achieves superior performance in unknown class detection while maintaining high accuracy on known classes. The code is available at https://github.com/wangsfan/MOF.
Two-dimensional maps of a video collection, in which nearby points correspond to related content, would make large archives browsable and searchable. Applying PCA, t-SNE, or UMAP to raw clips does not produce such maps: distances in pixel space are dominated by appearance, and the resulting layouts appear to be organized by recording conditions rather than by action. We examine why this happens and study a training-free alternative in which a frozen pre-trained network first maps each clip to a compact descriptor and a conventional embedding method then operates in that space. No component is trained or fine-tuned on the target data. Across sixteen encoders, spanning image and video architectures of both convolutional and attention-based design, evaluated on UCF101 and HMDB51, we separate three properties that a single score conflates: the quality of the descriptor, the fidelity of the reduction, and the class structure a reader sees in the map. The separation is informative. The encoder governs what the map shows, while the embedding method governs how faithfully it reproduces the space it came from. Video encoders lead image encoders in descriptor quality and in the class structure of the maps, a compact spatiotemporal network offers the best trade-off between quality and computation, and the choice of which token to read from a Transformer can matter as much as the choice of architecture.
