This report describes Tail-Aware HiFloat4, our submission to the low-bit text-to-video generation quantization challenge. Our method adapts the public ViDiT-Q post-training quantization pipeline to Wan2.2 under the HiFloat4 numerical format. We quantize the main linear layers in both Wan2.2 transformer modules with W4A4 HiFloat4 fake quantization, keep numerically sensitive boundary modules in high precision, and introduce an activation-tail-aware percentile calibration module for channel-mask construction. Together with compact PTQ-state restoration, this design reduces the influence of rare calibration outliers while keeping the runtime HiFloat4 arithmetic and sampling pipeline unchanged.

Continuous Spatio-Temporal Video Super-Resolution (C-STVSR) aims to simultaneously enhance the spatial resolution and frame rate of videos by arbitrary scale factors, offering greater flexibility than fixed-scale methods that are constrained by predefined upsampling ratios. In recent years, methods based on Implicit Neural Representations (INR) have made significant progress in C-STVSR by learning continuous mappings from spatio-temporal coordinates to pixel values. However, these methods fundamentally rely on dense pixel-wise grid queries, causing computational cost to scale linearly with the number of interpolated frames and severely limiting inference efficiency. We propose GS-STVSR, an ultra-efficient C-STVSR framework based on 2D Gaussian Splatting (2D-GS) that drives the spatiotemporal evolution of Gaussian kernels through continuous motion modeling, bypassing dense grid queries entirely. We exploit the strong temporal stability of covariance parameters for lightweight intermediate fitting, design an optical flow-guided motion module to derive Gaussian position and color at arbitrary time steps, introduce a Covariance resampling alignment module to prevent covariance drift, and propose an adaptive offset window for large-scale motion. Extensive experiments on Vid4, GoPro, and Adobe240 show that GS-STVSR achieves state-of-the-art quality across all benchmarks. Moreover, its inference time remains nearly constant at conventional temporal scales (X2–X8) and delivers over X3 speedup at extreme scales X32, demonstrating strong practical applicability.

Video temporal localization (VTL) aims to accurately identify the temporal boundary of a specific segment corresponding to a sentence query from an untrimmed video. Recently, DETR-based methods have emerged as effective solutions to VTL. The primary idea of these methods is to fuse multi-modal features and refine moment queries to obtain precise locations. Despite their success, these methods adopt fixed trainable moment queries, which implicitly model the statistical patterns of groundtruth segments in the training set and thus limit the model's generalization ability to out-of-distribution scenarios. To tackle this problem, we propose a Dynamic moment Query Recalibrated DEtection TRansformer (DQR-DETR) that adaptively generates and recalibrates multi-scale moment queries on the target segments for out-of-distribution VTL. Our method contains three key parts: (1) progressive semantic regularization that performs primitive modality consensus loss and refined semantic calibration loss at two phases for better global and fine-grained cross-modal associations; (2) hierarchical query decoupling that first dynamically extracts multi-scale and context-aware event queries from the joint visual-language representations to cover a wide range of potential events in video, and then select the sentence-aware queries with semantic guidance; (3) moment query recalibrating that performs coarse-to-fine modulation on multi-scale queries by first activating relevant queries with video highlight and sentence semantic guidance, and then refining them to focus on the target segments. Extensive experiments on QVHighlights, Charades-STA and TACoS datasets show that our DQR-DETR achieves state-of-the-art performance, especially in out-of-distribution scenarios where our method substantially outperforms existing approaches by large margins.

Current image de-raining methods primarily learn from a limited dataset, leading to inadequate performance in varied real-world rainy conditions. To tackle this, we introduce a new framework that enables networks to progressively expand their de-raining knowledge base by tapping into a growing pool of datasets, significantly boosting their adaptability. Drawing inspiration from the human brain’s ability to continually absorb and generalize from ongoing experiences, our approach borrows the mechanism of the complementary learning system. Specifically, we first deploy generative adversarial networks (GANs) to capture and retain the unique features of new data, mirroring the hippocampus’s role in learning and memory. Then, the de-raining network is trained with both existing and GAN-synthesized data, mimicking the process of hippocampal replay and interleaved learning. Furthermore, we employ knowledge distillation with the replayed data to replicate the synergy between the neocortex’s activity patterns triggered by hippocampal replays and the preexisting neocortical knowledge. This comprehensive framework empowers the de-raining network to accumulate knowledge from various datasets, continually enhancing its performance on previously unseen rainy scenes. Our testing on three benchmark de-raining networks confirms the framework’s effectiveness. It not only facilitates continual knowledge accumulation across six datasets but also surpasses state-of-the-art methods in generalizing to new real-world scenarios. Our code is available at https://github.com/wangkunyu241/CLGID

Spiking Neural Networks (SNNs) are well-regarded for their biological plausibility and energy efficiency in processing sequential data. However, dominant SNN architectures typically rely on first-order Ordinary Differential Equations (ODEs) to govern neuronal state transitions. This first-order assumption imposes a "memoryless" bottleneck, limiting the model's capacity to capture the complex, long-range dependencies inherent in long-sequence tasks. In this work, we propose LongSpike, a novel SNN framework that integrates fractional-order State-Space Modeling, or f-SSM, from control theory into the spiking domain. By extending traditional integer-order SSMs to the fractional-calculus regime, LongSpike enables the hierarchical integration of neuronal dynamics with long-memory kernels. To mitigate the computational overhead and parallelization challenges typically associated with fractional operators, we leverage a state-space formulation that supports efficient, parallel training. Empirical evaluations on challenging benchmarks, including Long Range Arena (LRA), large-scale WikiText-103, and Speech Commands, demonstrate that LongSpike outperforms state-of-the-art SNNs in accuracy while preserving sparse synaptic computation. The code is available at https://github.com/xinruihe389-commits/LongSpike.

Uncrewed aerial vehicles (UAV) are increasingly deployed to assist humans in diverse tasks, where understanding human intentions is critical to effective collaboration. Referring expression comprehension (REC) links language to visual targets, allowing UAV to recognize human-intended targets of interest, thereby supporting subsequent actions. However, existing REC research is almost exclusively confined to ground-based scenarios, leaving aerial scenarios largely unexplored. In this paper, we formally define UAV-based REC as a new research problem and highlight its unique challenges, including abundant background interference, small target size, and complex referring relations. To enable systematic study, we introduce SkyFind, a large-scale dataset with one million high-quality target-expression pairs, providing a solid foundation. In addition, we propose AerialREC, a baseline framework that reduces background interference in UAV imagery by searching for a potential target region before localization. We establish benchmark results on SkyFind using ten representative REC methods and validate the effectiveness of the AerialREC framework.

This paper provides a review of the NTIRE 2026 challenge on mobile real-world image super-resolution, highlighting the proposed solutions and the resulting outcomes. The challenge aims to recover high-resolution (HR) images from low-resolution (LR) counterparts generated through unknown degradations with a x4 scaling factor while ensuring the models remain executable on mobile devices. The objective is to develop effective and efficient network designs or solutions that achieve state-of-the-art real-world image super-resolution performance. The track of the challenge evaluates performance using a weighted combination of image quality assessment (IQA) score and speedup ratios. The competition attracted 108 registrants, with 16 teams achieving a valid score in the final ranking. This collaborative effort advances the performance of mobile real-world image super-resolution while offering an in-depth overview of the latest trends in the field.

Spatial understanding of the physical world from 2D visual inputs hinges on two complementary forms of geometric knowledge: holistic 3D structural perception and fine-grained metric scale estimation. Existing multimodal large language models (MLLMs) typically address only one facet, ingesting either depth maps or point clouds as additional model inputs, which incurs substantial computational overhead and inherits the generalization limitations of upstream prediction models. We propose GAMSI, a dual-pathway Geometry-Aware MLLM for Spatial Intelligence that takes only RGB images as input while internalizing both forms of geometric prior within a unified autoregressive backbone. Specifically, we introduce Metric-Structure Decoupled Queries (MSDQ) which employ two groups of learnable queries to respectively extract dense metric signals and sparse structural cues from the shared visual context, with a task-decoupled attention mask further preventing the two pathways from contaminating each other. Building on this, an Expert-Guided Visual Grounding (EVG) module projects the aggregated cues back to frame-level visual features and aligns them with vision foundation models, which serve purely as training-time supervision, rather than as model inputs. We further build a multi-task spatial instruction-tuning dataset (MTS) comprising 152,776 samples spanning 13 task types and three visual modalities, consolidated from six public datasets. Trained with a two-stage curriculum, GAMSI achieves state-of-the-art performance on seven spatial intelligence benchmarks.

Natural language video localization (NLVL), a crucial task in video understanding, localizes target moments in videos corresponding to a language description. The point-supervised paradigm, which requires only a single annotated frame instead of complete temporal boundaries, offers a favorable balance between performance and annotation cost. However, the absence of full annotations impedes effective video-language alignment, thereby reducing prediction accuracy. To overcome this limitation, we propose COTEL (COllaborative Temporal consistEncy Learning), a framework that enhances alignment by leveraging the synergy between saliency detection and moment localization. COTEL integrates frame-level and segment-level consistency learning branches, which are trained to reinforce each other via a cross-consistency guidance scheme. Furthermore, we introduce a hierarchical contrastive alignment loss that combines intra-video positive alignment with inter-video negative mining to achieve comprehensive video-text matching. Extensive experiments on three widely used benchmarks demonstrate that our method performs favorably against state-of-the-art approaches. All source code will be released.

Diffusion models have emerged as state-of-the-art generative models, capable of producing high-quality synthetic outputs. These generated contents serve two key purposes in practice: as end-user facing products (“Generated Content as Product”, GCAP) and for data augmentation in machine learning pipelines (“Generated Content as Data”, GCAD). The expedited deployment of diffusion models has heightened awareness of their potential vulnerabilities, among which backdoor attacks are identified as a prominent concern. However, existing backdoor attack methods primarily focus on the conventional GCAP scenario and are constrained by their reliance on perceptible, pixel-based triggers. The lack of exploration on imperceptible attacks and the neglect of GCAD scenario may exacerbate security risks and broaden the reach of adversarial impacts. To unveil hidden attack surfaces, we address the noted challenges and propose Evil Diffusion (ELF), a novel framework for launching stealthy backdoor attacks on diffusion models, including two algorithms tailored for the GCAP and GCAD scenarios, respectively. First, we introduce ELF-P, a two-stage training algorithm that autonomously learns a frequency-based stealthy trigger for the GCAP scenario. ELF-P employs perceptual and defense-resistant constraints in the frequency domain to craft a human-imperceptible trigger that is resilient to common steganography defenses, enabling highly covert backdooring of diffusion models. Second, we present ELF-D, an algorithm designed for the GCAD scenario that poisons a conditional diffusion model to compromise specific downstream tasks. ELF-D introduces perturbations into task-critical frequency components of generated images, allowing compromised diffusion models to subtly undermine the performance of downstream models trained on their outputs. Extensive experiments validate the effectiveness of the proposed ELF framework in both scenarios, unveiling the hidden vulnerabilities in diffusion models and emphasize the broader, potentially harmful implications of these attacks.

Autoregressive image generation aims to predict the next token based on previous ones. However, this process is challenged by the bidirectional dependencies inherent in conventional image tokenizations, which creates a fundamental misalignment with the unidirectional nature of autoregressive models. To resolve this, we introduce AliTok, a novel Aligned Tokenizer that alters the dependency structure of the token sequence. AliTok employs a bidirectional encoder constrained by a causal decoder, a design that compels the encoder to produce a token sequence with both semantic richness and forward-dependency. Furthermore, by incorporating prefix tokens and employing a two-stage tokenizer training process to enhance reconstruction performance, AliTok achieves high fidelity and predictability simultaneously. Building upon AliTok, a standard decoder-only autoregressive model with just 177M parameters achieves a gFID of 1.44 and an IS of 319.5 on ImageNet-256. Scaling to 662M, our model reaches a gFID of 1.28, surpassing the SOTA diffusion method with 10x faster sampling. On ImageNet-512, our 318M model also achieves a SOTA gFID of 1.39. Code and weights will be released.

Transformers have excelled in image restoration due to their advanced representational abilities. However, their reliance on a fixed local window for attention often undermines translation invariance and local relationship preservation. This limitation can reduce network stability, especially when dealing with positional changes in degradation scenarios. In this research, we present a new Bayesian Window Transformer, which innovates by employing a probability distribution for window shifts, overcoming the limitations of fixed window configurations in traditional transformers. This approach allows for more flexible coverage beyond a predetermined region. During the evaluation procedure, we further develop two approximate inference algorithms: Layer Expectation Propagation and Monte Carlo Average. These two algorithms calculate expectations derived from the introduced distribution to effectively approximate the marginalization results of the probabilistic variables. Hence, our Bayesian Window Transformer not only inherits the powerful representation ability but also maintains essential properties like translation invariance and local relationship preservation for image restoration. We also provide a theoretical guarantee, demonstrating that our method is aligned with the classic sliding window technique in terms of receptive field sizes and sliding behavior. Comprehensive experiments validate the exceptional effectiveness of our Bayesian Window Transformer across multiple image restoration tasks, including image deraining, denoising, and deblurring.

Arbitrary-scale super-resolution (ASSR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs with arbitrary upsampling factors using a single model, addressing the limitations of traditional SR methods constrained to fixed-scale factors (\textit{e.g.}, $\times$ 2). Recent advances leveraging implicit neural representation (INR) have achieved great progress by modeling coordinate-to-pixel mappings. However, the efficiency of these methods may suffer from repeated upsampling and decoding, while their reconstruction fidelity and quality are constrained by the intrinsic representational limitations of coordinate-based functions. To address these challenges, we propose a novel ContinuousSR framework with a Pixel-to-Gaussian paradigm, which explicitly reconstructs 2D continuous HR signals from LR images using Gaussian Splatting. This approach eliminates the need for time-consuming upsampling and decoding, enabling extremely fast ASSR. Once the Gaussian field is built in a single pass, ContinuousSR can perform arbitrary-scale rendering in just 1ms per scale. Our method introduces several key innovations. Through statistical analysis, we uncover the Deep Gaussian Prior (DGP) and propose DGP-Driven Covariance Weighting, which dynamically optimizes covariance via adaptive weighting. Additionally, we present Adaptive Position Drifting, which refines the positional distribution of the Gaussian space based on image content, further enhancing reconstruction quality. Extensive experiments on seven benchmarks demonstrate that our ContinuousSR delivers significant improvements in SR quality across all scales, with an impressive 19.5× speedup when continuously upsampling an image across forty scales.

In language reasoning, longer chains of thought consistently yield better performance, which naturally suggests that visual latent reasoning may likewise benefit from longer latent sequences. However, we discover a counterintuitive phenomenon: the performance of existing latent visual reasoning methods systematically degrades as the latent sequence grows longer. We reveal the root cause: Information Gain Collapse – autoregressive generation makes each step highly dependent on prior outputs, so subsequent tokens can barely introduce new information. We further identify that heavily pooled (≥ 128×) image embeddings used as supervision targets provide no more signal than meaningless placeholders. Motivated by these insights, we propose SCOLAR (Self-COnsistent LAtent Reasoning), which introduces a lightweight detransformer that leverages the LLM's full-sequence hidden states to generate auxiliary visual tokens in a single shot, with each token independently anchored to the original visual space. Combined with three-stage SFT and ALPO reinforcement learning, SCOLAR extends acceptable latent CoT length by over 30×, achieves state-of-the-art among open-source models on real-world reasoning benchmarks (+14.12

Multimodal large language models (MLLMs) are increasingly considered as a foundation for embodied agents, yet it remains unclear whether they can reliably reason about the long-term physical consequences of actions from an egocentric viewpoint. We study this gap through a new task, Egocentric Scene Prediction with LOng-horizon REasoning: given an initial-scene image and a sequence of atomic action descriptions, a model is asked to predict the final scene after all actions are executed. To enable systematic evaluation, we introduce EXPLORE-Bench, a benchmark curated from real first-person videos spanning diverse scenarios. Each instance pairs long action sequences with structured final-scene annotations, including object categories, visual attributes, and inter-object relations, which supports fine-grained, quantitative assessment. Experiments on a range of proprietary and open-source MLLMs reveal a significant performance gap to humans, indicating that long-horizon egocentric reasoning remains a major challenge. We further analyze test-time scaling via stepwise reasoning and show that decomposing long action sequences can improve performance to some extent, while incurring non-trivial computational overhead. Overall, EXPLORE-Bench provides a principled testbed for measuring and advancing long-horizon reasoning for egocentric embodied perception.

Egocentric point tracking aims to localize points on object surfaces from a first-person perspective and serves as a critical step toward embodied intelligence. Recent methods rely on video input, tracking query points through feature matching across consecutive frames. However, these methods struggle in highly dynamic settings—a common challenge in first-person perspectives, where the head-mounted camera undergoes frequent and abrupt rotations, resulting in high angular velocities, motion blur, and large inter-frame displacements. In contrast, event cameras capture motion at microsecond temporal resolution, naturally avoiding blur and delivering low-latency, high-fidelity cues crucial for egocentric point tracking. Moreover, rapid egocentric motion disrupts local smoothness, breaking the assumption that spatially adjacent regions share similar motion. Event dynamics expose global motion trends, guiding coherent modeling and consistent feature flow. Therefore, this paper proposes a mamba-based tracking framework that constructs feature modeling paths aligned with the dominant motion trend extracted from events, and modulates feature propagation along these paths based on local motion intensity, enhancing stability by suppressing unreliable signals and emphasizing consistent cues. Additionally, a motion-adaptive suppression module enhances temporal robustness by adaptively suppressing correlation features based on motion intensity variations, mitigating the effects of intensity fluctuations and partial observability. To facilitate research in this domain, a multimodal dataset named DVS-EgoPoints with both events and videos for egocentric point tracking is collected. Experiments on the DVS-EgoPoints dataset and a simulation benchmark demonstrate superior performance over state-of-the-art methods, especially under challenging motion and occlusion conditions.

Spiking Neural Networks (SNNs) draw inspiration from biological neurons to enable brain-like computation, demonstrating effectiveness in processing temporal information with energy efficiency and biological realism. Most existing SNNs are based on neural dynamics such as the (leaky) integrate-and-fire (IF/LIF) models, which are described by \emph{first-order} ordinary differential equations (ODEs) with Markovian characteristics. This means the potential state at any time depends solely on its immediate past value, potentially limiting network expressiveness. Empirical studies of real neurons, however, reveal long-range correlations and fractal dendritic structures, suggesting non-Markovian behavior better modeled by \emph{fractional-order} ODEs. Motivated by this, we propose a \emph{fractional-order} spiking neural network (\emph{f}-SNN) framework that strictly generalizes integer-order SNNs and captures long-term dependencies in membrane potential and spike trains via fractional dynamics, enabling richer temporal patterns. We also release an open-source toolbox to support the \emph{f}-SNN framework, applicable to diverse architectures and real-world tasks. Experimentally, fractional adaptations of established SNNs into the \emph{f}-SNN framework achieve superior accuracy, comparable energy efficiency, and improved robustness to noise, underscoring the promise of \emph{f}-SNNs as an effective extension of traditional SNNs.

Estimating the 3D motion of scene points from 2D observations, typically parameterized by optical flow and motion in depth, is a fundamental problem in computer vision. Existing learning-based methods usually rely on supervised regression from densely labeled data, but their dependence on annotations and limited use of geometric constraints restricts generalization, motivating unsupervised solutions. Unsupervised 3D motion estimation is challenging because motion along the viewing direction is unobservable, and optical flow and motion in depth are geometrically coupled, making their separation ambiguous. Event cameras capture per-pixel brightness changes asynchronously with microsecond latency, providing high temporal resolution and motion continuity. Projecting event streams along different axes reveals spatiotemporal expansion and contraction patterns that encode depth variation and geometric structure, offering rich cues for unsupervised estimation. Leveraging these properties, we propose an unsupervised event-based 3D motion estimation framework that jointly models optical flow and motion in depth. We first derive an analytical relationship to infer initial motion in depth from estimated flow and further refine it using a directional expansion modulation module that captures horizontal and vertical expansion–contraction patterns in event projections. Finally, motion in depth is incorporated into optical flow warping under a contrast maximization objective. Experiments on the CarlaEvent3D dataset show that our method achieves competitive accuracy and strong generalization, advancing unsupervised 3D motion estimation in the event domain.

Few-shot class-incremental learning (FSCIL) aims to recognize novel classes from only a few labeled samples while retaining previously learned knowledge. Although recent FSCIL methods have achieved substantial progress on visual benchmarks, they remain limited in tactile sensing, where the same material may produce markedly different observations under different acquisition contexts, such as sensing devices, contact states, scanning trajectories, and interaction conditions. In tactile FSCIL, the challenges of few-shot learning and class-incremental learning are further amplified by acquisition context: the limited support samples may not only be scarce, but also carry context-induced biases. Once the resulting biased prototypes are inserted into the classifier, they may affect the decision boundaries in subsequent sessions. To address this problem, we propose Context-Probing Few-Shot Class-Incremental Learning (CoP-FSCIL), a context-aware framework for tactile FSCIL. CoP-FSCIL first employs Context-Probing Intervention (CPI) to diagnose local context-sensitive variations in tactile representations. It then introduces a Probe-Conditioned Quotient Adapter (PCQA) to suppress context-sensitive components identified by the probes. Finally, Probe-Stability Prototype Calibration (PSPC) estimates support sample reliability from probe-induced embedding fluctuations and calibrates stochastic prototypes accordingly. Experiments on HapTex and LMT108 show that CoP-FSCIL consistently outperforms representative FSCIL baselines, and extended experiments on audio FSCIL further demonstrate the generality of the proposed context probing mechanism. The source code is currently being prepared and will be released soon.

Monocular 4D human-object interaction (HOI) reconstruction - recovering a moving human and a manipulated object from a single RGB video - remains challenging due to depth ambiguity and frequent occlusions. Existing methods often rely on multi-stage pipelines or iterative optimization, leading to high inference latency, failing to meet real-time requirements, and susceptibility to error accumulation. To address these limitations, we propose THO, an end-to-end Spatial-Temporal Transformer that predicts human motion and coordinated object motion in a forward fashion from the given video and 3D template. THO achieves this by leveraging spatial-temporal HOI tuple priors. Spatial priors exploit contact-region proximity to infer occluded object features from human cues, while temporal priors capture cross-frame kinematic correlations to refine object representations and enforce physical coherence. Extensive experiments demonstrate that THO operates at an inference speed of 31.5 FPS on a single RTX 4090 GPU, achieving a >600x speedup over prior optimization-based methods while simultaneously improving reconstruction accuracy and temporal consistency. The project page is available at: https://nianheng.github.io/THO-project/