Infrared image super-resolution (IISR) is inherently more challenging than its visible-light counterpart due to the presence of structured, infrared-specific degradations, including thermal noise, fixed-pattern noise, and spatially non-uniform blur. These degradations deviate substantially from commonly assumed degradation models and often lead existing methods to produce unstable structures, amplified noise, or hallucinated textures. In this work, we present DIRSR, a degradation-aware framework for infrared image super-resolution that explicitly incorporates infrared degradation perception into the restoration process. The proposed approach characterizes instance-level infrared degradation conditions and leverages degradation-aware semantic guidance to regulate the trade-off between perceptual quality and structural fidelity. In addition, infrared imaging priors are embedded into a physics-informed, region-aware optimization objective to enhance foreground structural consistency while suppressing spurious background details. Extensive experiments on multiple public infrared benchmarks demonstrate that the proposed framework consistently outperforms state-of-the-art methods in both quantitative performance and visual quality, and further improves downstream infrared detection and segmentation tasks.
In real scenarios, multi-view data are often continuously acquired in the form of view streams, giving rise to the task of Streaming View Clustering (SVC). However, most existing methods rely on view-specific knowledge to guide the learning of the current view, while neglecting the shared, view-invariant knowledge structure across views. As a result, the model progressively forgets the commonalities between current and historical views during training, i.e., catastrophic forgetting, which leads to severe fluctuations in clustering performance under different input sequences. To address this issue, we propose a Dynamic kNowledge-Aware (DNA) framework for SVC, which is designed to mitigate the negative impact of catastrophic forgetting effectively. The proposed DNA framework consists of two main stages: the initial learning stage and the continual learning stage. Specifically, during the initial learning stage, we employ knowledge-aware learning to extract prototype knowledge from the first acquired view and construct an initial knowledge base. During the continual learning stage, we leverage knowledge aggregation learning to maximize the consistency between the prototype knowledge of the current view and the historical knowledge, and then fuse them into a new prototype knowledge and update the knowledge base, thereby ensuring the completeness of the learned knowledge. Finally, through structural guidance learning, the updated knowledge base is used to steer the sample distribution of the current view, thereby ensuring distributional consistency across different views. Extensive experimental results demonstrate that DNA consistently and significantly outperforms state-of-the-art methods. The code is available at https://github.com/hl-yuan/DNA.
The effectiveness of vision token pruning critically depends on the criterion used to define token redundancy. Existing approaches, which primarily rely on global saliency scores or similarity heuristics, often fail to account for a token’s essential contribution to the model’s internal representations. This oversight results in retaining visually prominent but redundant tokens or merging semantically distinct features, severely degrading performance under tight compression budgets. To address this, we propose a novel token pruning strategy aligned with representation complexity. We first introduce Local Representation Complexity (LRC), a token-level measure that estimates the intrinsic dimension of each token’s embedding. Tokens exhibiting high LRC are considered locally unpredictable and are therefore prioritized for preservation. To ensure selection diversity and mitigate over-concentration in the selected set, we frame the pruning process as a budget allocation problem and introduce a Greedy Selection algorithm with Penalized Diversity to actively suppress redundancy. We evaluate the proposed method across five vision-language model backbones from three model families, including LLaVA-1.5/LLaVA-NeXT, Qwen2.5-VL/Qwen3-VL, and InternVL3, covering a wide range of capacities and tasks. The results demonstrate competitive performance across different backbones, with particularly notable gains at aggressive compression ratios.
Sketch colorization in open-world scenarios presents unique challenges due to diverse, unpaired datasets and the variety in user-generated sketch styles. Traditional exemplar-based methods, which typically rely on dense feature matching and style normalization, struggle to generalize effectively in these conditions, as they require structured paired data and exhibit limited flexibility. We introduce Progressive Style Transfer (PROST), a completely unsupervised framework that addresses these limitations by progressively injecting style features into sketches without rigid content-style alignment. PROST comprises two key modules: Progressive Style Injection (PSI), which transfers style features adaptively across scales, and Progressive Style Enhancement (PSE), which refines style coherence through multi-layer style-adaptive attention. Together, these modules enable flexible style transfer that preserves content fidelity while achieving high stylistic coherence. Our approach also integrates a Dual Cycle-Consistency (DCC) framework to ensure both style and content consistency in generated outputs. Extensive evaluations demonstrate that PROST effectively generalizes to unseen styles, achieving robust performance across varied datasets and setting a new benchmark for open-world sketch colorization. The source code is publicly available at https://github.com/Vincotto/ProST.
Neural network-based in-loop filtering (ILF) technologies have achieved remarkable coding gains, emerging as a promising candidate for next-generation video coding standards. However, using deep neural networks (DNNs) brings significant computational and time complexity or high demands for dedicated hardware, making it challenging for general use. To address this limitation, we study a practical ILF solution by adopting look-up tables (LUTs). After training a DNN for ILF with a restricted reference range, all possible inputs are traversed and the output values of the DNN are stored into LUTs. During the coding process, the filtering process is performed by retrieving the filtered pixel through locating the input pixels and interpolating between the stored values, instead of relying on heavy DNN inference. We propose a universal LUT-based ILF framework, termed LUT-ILF++. First, we introduce the cooperation of multiple kinds of filtering LUTs and propose a series of customized indexing mechanisms to enable better filtering reference perception with limited storage consumption. Second, we propose the cross-component indexing mechanism to enable the filtering of different color components jointly. Third, in order to make our solution practical for coding uses, we propose the LUT compaction scheme to enable the LUT pruning, achieving a lower storage cost of the entire solution. The proposed framework is implemented in the Neural Network-based Video Coding reference software. Experimental results show that the proposed framework achieves on average 0.82%/2.97%/1.63% and 0.85%/4.11%/2.06% bitrate reduction for common test sequences, under the all-intra (AI) and random-access (RA) configurations, respectively. Compared to DNN inference-based solutions, our proposed solution has much lower time complexity and modest storage cost. Code is publicly available at https://github.com/milkbrotherli/LUT-family.
