Forward-looking sonar (FLS) plays an important role in underwater perception for autonomous marine robots. However, due to the inherent characteristics of sonar imaging, such as strong noise interference, weak semantic representation and complex background structures, FLS-based object detection remains a challenging problem. Moreover, relying on single-modal detection methods often leads to limited accuracy and poor robustness in complex underwater environments. To address these limitations, this paper proposes a novel sonar-camera fusion detection network named SCFusion, which exploits the complementary characteristics between visual semantic cues and sonar geometric information for robust underwater object detection. Specifically, SCFusion employs a camera-sonar complementary reasoning module to jointly capture the semantic cues from camera images and the geometric characteristics inherent in sonar measurements, enabling more effective multimodal feature representation. Moreover, a vision-guided context aggregation module is designed to combine highlevel semantic information from the camera modality to guide the contextual enhancement of sonar features. Extensive experiments on a public multimodal dataset demonstrate that the proposed SCFusion framework significantly outperforms existing single-modal and multimodal detection approaches in terms of detection accuracy. The results also highlight the effectiveness of sonar-vision fusion for improving reliable underwater object perception in challenging marine environments.

In recent years, autonomous docking technology for autonomous surface vessels (ASVs) has attracted increasing research interest, and boat berth detection plays a crucial role in this task. However, common berth detection methods for ASVs adopt specific marks, such as ArUco codes, to implement boat berth detection, which significantly limits their applicability in practical deployments, since most conventional wharfs do not contain such markers. To support tasks involving human-less return to dock and enhance the generalizability of berth detection in real-world wharf scenarios, this article proposes a novel markless berth detection framework based on high-level scene semantics, without relying on artificial markers. Our proposed detection framework, based on the fusion of camera and light detection and ranging (LiDAR), comprises two important components. First, a LiDAR-camera fusion point cloud semantic segmentation model is constructed to generate 3-D semantic point clouds. Second, a berth extraction algorithm is designed to process point cloud segmentation results and detect free berths for ASVs. In addition, to improve the performance of the point cloud semantic segmentation model for ASV scenes, we developed a LiDAR-assisted view transformer module and a LiDAR-camera fusion attention module. Moreover, in our collected point cloud segmentation data set for ASV wharf scenes, our proposed method outperforms other point cloud segmentation baselines. Moreover, we conducted real-world prototype experiments to validate the feasibility and robustness of our markless boat docking detection method. The experimental results demonstrate that our proposed berth detection method has a high potential to support applications involving the autonomous docking of ASVs.

Anticipating diverse future states is a central challenge in video world modeling. A key limitation lies in the computational cost of generating multiple plausible futures with existing world models. Recent work demonstrates that predicting the future in the latent space of a vision foundation model (VFM), rather than in raw pixel space, greatly improves efficiency. Despite this progress, efficient VFM-based world models are still predominantly discriminative, producing predictions that implicitly average over many possible futures. To explicitly and efficiently model diverse plausible futures, we introduce DeltaWorld, the first VFM-based world model which shifts from deterministic prediction to the ability to generate multiple plausible futures in a single forward pass. At the core of DeltaWorld is DeltaTok, a tokenizer that encodes feature differences between consecutive frames into a single compact “delta” token, effectively reducing redundancy among temporally adjacent feature maps. By representing futures as delta tokens, DeltaWorld efficiently generates multiple diverse predictions in parallel. Experiments on dense forecasting tasks demonstrate that DeltaWorld is capable of predicting futures that more closely align with real-world outcomes, while being orders of magnitude more efficient than existing generative world models. Code will be made publicly available.

Side scan sonar (SSS) is crucial for seafloor exploration owing to its unique ability to provide both high long-range visibility and extensive area coverage. However, inherent challenges such as the scarcity of sonar images, low signal-to-noise ratios, and degraded image quality persist, limiting the practical applicability of sonar image detection. Existing sonar image detection networks lack sufficient utilization of underwater imaging mechanisms. To address these challenges, we propose SonarDeNet, a novel side scan sonar image detection framework that incorporates prior-driven acoustic knowledge into sonar data simulation and deep-learning based detection network. Specifically, we enhance the traditional ray tracing SSS image simulation method with acoustic imaging mechanisms to provide high-quality simulation SSS images. Furthermore, we introduce the acoustic shadow as key prior-driven prompt for our SSS image detector, and a soft knowledge association mechanism is designed to eliminate the cognitive bias in prompt features. The experiments demonstrate that SonarDeNet achieves state-of-the-art detection performance compared to other detection baseline methods, with improvements of 3.6% in AP50 and 3% in mAP, highlighting its significant potential in incorporating prior knowledge for SSS object detection.

Flow matching models have emerged as a powerful framework for realistic image generation by learning to reverse a corruption process that progressively adds Gaussian noise. However, because noise is injected in the latent domain, its impact on different frequency components is non-uniform. As a result, during inference, flow matching models tend to generate low-frequency components (global structure) in the early stages, while high-frequency components (fine details) emerge only later in the reverse process. Building on this insight, we propose Frequency-Aware Flow Matching (FreqFlow), a novel approach that explicitly incorporates frequency-aware conditioning into the flow matching framework via time-dependent adaptive weighting. We introduce a two-branch architecture: (1) a frequency branch that separately processes low- and high-frequency components to capture global structure and refine textures and edges, and (2) a spatial branch that synthesizes images in the latent domain, guided by the frequency branch's output. By explicitly integrating frequency information into the generation process, FreqFlow ensures that both large-scale coherence and fine-grained details are effectively modeled—low-frequency conditioning reinforces global structure, while high-frequency conditioning enhances texture fidelity and detail sharpness. On the class-conditional ImageNet-256 generation benchmark, our method achieves state-of-the-art performance with an FID of 1.38, surpassing the prior diffusion model DiT and flow matching model SiT by 0.79 and 0.58 FID, respectively. Code will be made available.

In recent years, video generation has seen significant advancements. However, challenges still persist in generating complex motions and interactions. To address these challenges, we introduce ReVision, a plug-and-play framework that explicitly integrates parameterized 3D model knowledge into a pretrained conditional video generation model, significantly enhancing its ability to generate high-quality videos with complex motion and interactions. Specifically, ReVision consists of three stages. First, a video diffusion model is used to generate a coarse video. Next, we extract a set of 2D and 3D features from the coarse video to construct a 3D object-centric representation, which is then refined by our proposed parameterized motion prior model to produce an accurate 3D motion sequence. Finally, this refined motion sequence is fed back into the same video diffusion model as additional conditioning, enabling the generation of motion-consistent videos, even in scenarios involving complex actions and interactions. We validate the effectiveness of our approach on Stable Video Diffusion, where ReVision significantly improves motion fidelity and coherence. Remarkably, with only 1.5B parameters, it even outperforms a state-of-the-art video generation model with over 13B parameters on complex video generation by a substantial margin. Our results suggest that, by incorporating 3D motion knowledge, even a relatively small video diffusion model can generate complex motions and interactions with greater realism and controllability, offering a promising solution for physically plausible video generation.

Forward-Looking Sonar (FLS) is a critical tool for underwater surveying, resource exploration, and deep-sea mapping. Recent CNN-Transformer hybrid frameworks have demonstrated strong performance on FLS segmentation by jointly capturing fine-grained local details and long-range dependencies. Nevertheless, their reliance on self-attention mechanisms incurs substantial computational overhead, limiting their efficiency and scalability. To address these limitations, we propose MambaSonar, an efficient CNN-Mamba model for FLS image segmentation. MambaSonar combines the local feature extraction strengths of convolutional neural networks (CNN) with Mamba's efficient global dependency modeling via its selective state-space mechanism. To better adapt Mamba to the characteristics of FLS images, we introduce an acoustic background suppression block, which reduces environmental noise and emphasizes meaningful target responses, and a Mamba-CNN fusion block, which bridges semantic gaps and effectively integrates multi-scale features. Extensive experiments on several public FLS datasets demonstrate that MambaSonar delivers superior segmentation accuracy while maintaining high computational efficiency. In addition, ablation studies also validate the effectiveness of the proposed model and its components, highlighting the potential of hybrid CNN-Mamba architectures for challenging underwater imaging tasks.

Forward-looking sonar (FLS) image segmentation enables underwater robotics autonomous exploration in marine engineering. However, the inherent characteristics of sonar systems such as weak semantic information, complex backgrounds, and high environmental noise, pose significant challenges and cause performance bottlenecks for traditional FLS image segmentation algorithms. Moreover, relying on single-modal features further exacerbates these limitations, restricting the segmentation accuracy and robustness. To address these limitations, this paper proposes FLSFusionNet which models the multi-modal relationships between sonar space and vision modality, capturing complementary features and enhancing segmentation performance through adaptive cross-modal interactions. Specifically, FLSFusionNet comprises four key modules: (1) sonar and camera backbones that extract distinct features from each modality; (2) a camera-sonar spatial transformer module designed to bridge the representation gap between sonar and visual features; (3) a cross-modal adaptive fusion module that maintain the semantic consistency between the dual features for effective integration; and (4) a fusion-based segmentation head that predicts the final segmentation results. Experiments conducted on the multi-modal benchmark dataset demonstrate that the proposed fusion approach significantly outperforms existing single-modal and multi-modal sonar image methods in segmentation accuracy, especially under complex underwater conditions. The results underscore the potential of sonar-vision fusion strategy to enhance the robustness and precision of FLS-based applications.

Side-scan sonar (SSS) imaging plays a critical role in underwater perception for autonomous underwater vehicles (AUVs). However, the spatial sparsity of targets and the limited computational resources remain challenging for real-time object detection. Existing methods typically adopt dense inference strategies, leading to substantial computational redundancy and limited deployment feasibility. In this work, we propose a lightweight and ultra-fast SSS object detection framework based on target presence awareness. The proposed framework follows a coarse-to-fine inference paradigm, in which a target presence analysis module is first employed to rapidly filter out target-absent image patches, and only target-positive patches are forwarded to an Object Forward Detection (OFD) module for fine-grained detection. The TPA module integrates spatial–frequency convolution to efficiently capture both local structural cues and global contextual information with minimal computational overhead. Furthermore, an AttnConv-enhanced detection module is introduced in the OFD stage to strengthen high-frequency target features and improve fine-grained detection performance. Extensive experiments on public SSS datasets demonstrate that the proposed method achieves an mAP of 74.63% on the AI4Shipwrecks dataset and 63.02% on the SSS-Mine dataset. Notably, the framework delivers an ultra-fast inference speed of 174.74 FPS on embedded hardware, representing a 5.2× speedup over conventional dense-processing detection methods.

This paper challenges the dominance of continuous pipelines in visual generation. We systematically investigate the performance gap between discrete and continuous methods. Contrary to the belief that discrete tokenizers are intrinsically inferior, we demonstrate that the disparity arises primarily from the total number of bits allocated in the latent space (i.e., the compression ratio). We show that scaling up the codebook size effectively bridges this gap, allowing discrete tokenizers to match or surpass their continuous counterparts. However, existing discrete generation methods struggle to capitalize on this insight, suffering from performance degradation or prohibitive training costs with scaled codebook. To address this, we propose masked it utoegressive modeling (), a scalable framework that supports arbitrary codebook sizes. By equipping an autoregressive transformer with a masked bit modeling head, BAR predicts discrete tokens through progressively generating their constituent bits. BAR achieves a new state-of-the-art gFID of on ImageNet-256, outperforming leading methods across both continuous and discrete paradigms, while significantly reducing sampling costs and converging faster than prior continuous approaches.

Bridging different modalities lies at the heart of cross-modality generation. While conventional approaches treat the text modality as a conditioning signal that gradually guides the denoising process from Gaussian noise to the target image modality, we explore a much simpler paradigm-directly evolving between text and image modalities through flow matching. This requires projecting both modalities into a shared latent space, which poses a significant challenge due to their inherently different representations: text is highly semantic and encoded as 1D tokens, whereas images are spatially redundant and represented as 2D latent embeddings. To address this, we introduce FlowTok, a minimal framework that seamlessly flows across text and images by encoding images into a compact 1D token representation. Compared to prior methods, this design reduces the latent space size by 3.3x at an image resolution of 256, eliminating the need for complex conditioning mechanisms or noise scheduling. Moreover, FlowTok naturally extends to image-to-text generation under the same formulation. With its streamlined architecture centered around compact 1D tokens, FlowTok is highly memory-efficient, requires significantly fewer training resources, and achieves much faster sampling speeds-all while delivering performance comparable to state-of-the-art models. Code is available at https://github.com/TACJu/FlowTok.

In this work, we present CoCal, an interpretable and consistent object parsing framework based on dictionary-based mask transformer. Designed around Contrastive Components and Logical Constraints, CoCal rethinks existing cluster-based mask transformer architectures used in segmentation; Specifically, CoCal utilizes a set of dictionary components, with each component being explicitly linked to a specific semantic class. To advance this concept, CoCal introduces a hierarchical formulation of dictionary components that aligns with the semantic hierarchy. This is achieved through the integration of both within-level contrastive components and cross-level logical constraints. Concretely, CoCal employs a component-wise contrastive algorithm at each semantic level, enabling the contrasting of dictionary components within the same class against those from different classes. Furthermore, CoCal addresses logical concerns by ensuring that the dictionary component representing a particular part is closer to its corresponding object component than to those of other objects through a cross-level contrastive learning objective. To further enhance our logical relation modeling, we implement a post-processing function inspired by the principle that a pixel assigned to a part should also be assigned to its corresponding object. With these innovations, CoCal establishes a new state-of-the-art performance on both PartImageNet and Pascal-Part-108, outperforming previous methods by a significant margin of 2.08 CoCal exhibits notable enhancements in object-level metrics across these benchmarks, highlighting its capacity to not only refine parsing at a finer level but also elevate the overall quality of object segmentation.

This paper presents Randomized AutoRegressive modeling (RAR) for visual generation, which sets a new state-of-the-art performance on the image generation task while maintaining full compatibility with language modeling frameworks. The proposed RAR is simple: during a standard autoregressive training process with a next-token prediction objective, the input sequence-typically ordered in raster form-is randomly permuted into different factorization orders with a probability r, where r starts at 1 and linearly decays to 0 over the course of training. This annealing training strategy enables the model to learn to maximize the expected likelihood over all factorization orders and thus effectively improve the model's capability of modeling bidirectional contexts. Importantly, RAR preserves the integrity of the autoregressive modeling framework, ensuring full compatibility with language modeling while significantly improving performance in image generation. On the ImageNet-256 benchmark, RAR achieves an FID score of 1.48, not only surpassing prior state-of-the-art autoregressive image generators but also outperforming leading diffusion-based and masked transformer-based methods. Code and models will be made available at https://github.com/bytedance/1d-tokenizer

Image tokenizers form the foundation of modern text-to-image generative models but are notoriously difficult to train. Furthermore, most existing text-to-image models rely on large-scale, high-quality private datasets, making them challenging to replicate. In this work, we introduce Text-Aware Transformer-based 1-Dimensional Tokenizer (TA-TiTok), an efficient and powerful image tokenizer that can utilize either discrete or continuous 1-dimensional tokens. TA-TiTok uniquely integrates textual information during the tokenizer decoding stage (i.e., de-tokenization), accelerating convergence and enhancing performance. TA-TiTok also benefits from a simplified, yet effective, one-stage training process, eliminating the need for the complex two-stage distillation used in previous 1-dimensional tokenizers. This design allows for seamless scalability to large datasets. Building on this, we introduce a family of text-to-image Masked Generative Models (MaskGen), trained exclusively on open data while achieving comparable performance to models trained on private data. We aim to release both the efficient, strong TA-TiTok tokenizers and the open-data, open-weight MaskGen models to promote broader access and democratize the field of text-to-image masked generative models.

Recently, unmanned surface vessels (USVs) have played an increasingly important role in autonomous exploration, and boat detection is an important task for USVs. While most existing boat detection methods focus on 2-dimensional (2D) detection, 3-dimensional (3D) detection which provides valuable spatial direction for moving target estimation has not been studied in the boat detection field. However, 3D boat detection on water surfaces faces challenging problems, such as small sizes of detected targets and diverse moving directions. Considering that traditional LiDAR-based 3D boat detection methods require high hardware costs, we fuse millimeter wave (MMW) radar and high semantic camera to achieve low-cost and high-quality 3D boat detection. We propose a novel radar-camera fusion boat 3D detection model named RCBDet. The proposed RCBDet employs a new dual radar encoder and first introduces Doppler speed information from MMW radar into neural network to overcome sparse radar points. A new radar-camera attention module is designed to effectively combine camera features, radar spatial features and radar velocity features, encapsulating not only shape and semantic attributes but also spatial orientation information. In our collected boat 3D detection dataset, RCBDet achieves state-of-the-art accuracy compared to other single-modality baselines and radar-camera fusion baselines. Moreover, we conducted comprehensive ablation experiments to validate the efficacy of the designed modules. The experimental results demonstrated that the proposed radar-camera fusion model effectively fuses MMW radar features and camera features.

Forward-looking sonar (FLS) imaging systems can provide long-range underwater environments in ultralow visibility and contrast, holding a critical role in the recognition and segmentation of underwater objects. However, sonar image segmentation presents significant challenges due to the main factors: 1) the reverberation strength increased by the coherent scattering phenomenon, which hampers the segmentor's ability to capture object features accurately; and 2) insufficient multifeature utilization of weakly annotated samples and a lack of prior information on underwater object scales. To overcome this challenge, we proposed a novel FLS image segmentation method based on the reverberation suppression and multilayer context aware. Detailed statistical analysis of reverberation distribution using nonlinear least squares fitting, we design a self-supervised background suppression auxiliary branch (BSAB) to reduce inherent gamma noise and preserve pure features. To further leverage the multiscale context features extracted from different transformer encoding layers, we propose a multilayer aggregation transformer (MLAT) as an encoder to acquire more comprehensive target features. In addition, a global-local feature complementary fusion mechanism was designed for the interaction of these independent features, to achieve an optimal comprehensive state at the spatial and channel level. The comparative experiments demonstrate that FISNet achieves superior performance with state-of-the-art segmentation networks.

Autoregressive (AR) modeling, known for its next-token prediction paradigm, underpins state-of-the-art language and visual generative models. Traditionally, a “token” is treated as the smallest prediction unit, often a discrete symbol in language or a quantized patch in vision. However, the optimal token definition for 2D image structures remains an open question. Moreover, AR models suffer from exposure bias, where teacher forcing during training leads to error accumulation at inference. In this paper, we propose xAR, a generalized AR framework that extends the notion of a token to an entity X, which can represent an individual patch token, a cell (a k× k grouping of neighboring patches), a subsample (a non-local grouping of distant patches), a scale (coarse-to-fine resolution), or even a whole image. Additionally, we reformulate discrete token classification as continuous entity regression, leveraging flow-matching methods at each AR step. This approach conditions training on noisy entities instead of ground truth tokens, leading to Noisy Context Learning, which effectively alleviates exposure bias. As a result, xAR offers two key advantages: (1) it enables flexible prediction units that capture different contextual granularity and spatial structures, and (2) it mitigates exposure bias by avoiding reliance on teacher forcing. On ImageNet-256 generation benchmark, our base model, xAR-B (172M), outperforms DiT-XL/SiT-XL (675M) while achieving 20× faster inference. Meanwhile, xAR-H sets a new state-of-the-art with an FID of 1.24, running 2.2× faster than the previous best-performing model without relying on vision foundation modules (, DINOv2) or advanced guidance interval sampling.

Underwater object detection plays a crucial role in aquaculture and marine environmental protection. Compared to conventional images, underwater images often suffer from challenges such as low brightness, color distortions, blurred details, and noise. In recent years, frequency-domain techniques have demonstrated significant potential in underwater image processing. Notably, distinct patterns in blur distribution be observed across the low-frequency and high-frequency components of underwater images from various datasets. In response to these challenges, we propose a novel spatial-frequency domain modulation underwater object detection network, termed SFUDNet. Unlike existing spatial-domain underwater object detection methods, SFUDNet introduces an innovative spatial-frequency decoupling structure with a mixture of expert mechanism, which is implemented through the proposed frequency modulation block (FMB) and spatial frequency integration (SFI) module. The FMB employs a mixture-of-expert approach to dynamically learn diverse frequency features across different granularities and scales in a sample-adaptive manner, subsequently performing element-wise local feature modulation. Meanwhile, the SFI module effectively integrates frequency domain features with spatial-domain features, enabling a more comprehensive representation of underwater scenes. Extensive experiments on publicly available underwater datasets demonstrate that SFUDNet achieves state-of-the-art performance, outperforming existing underwater object detection baselines in both detection accuracy and robustness.

The marine object detection is an essential component for maritime autonomous surface ship (MASS) systems, enabling them to perceive obstacle objects and avoid them. Existing detection algorithms primarily rely on visual sensors and face serious limitations in complex maritime environments. In this article, we propose a novel joint object detection method based on the fusion of marine radar and camera data for robust maritime object detection. By employing a multistage hybrid-level fusion architecture, the proposed method efficiently fuses marine radar data with camera images, thereby enhancing the accuracy and robustness of object detection in complex maritime contexts. In addition, inspired by vision transformer detectors, a new transformer-based fusion module (TFM) is proposed to mitigate semantic disparity between the two modalities and enhance radar information guidance through extracted object queries. Extensive experiments on public maritime perception datasets demonstrate the significant advantages of the proposed method in various maritime environments. The framework also exhibits strong generalization capability, making it applicable to diverse maritime settings, including challenging conditions like low visibility and high-background noise. In public maritime perception datasets, the proposed method achieves state-of-the-art accuracy compared with other single-modality baselines and radar-camera fusion baselines, outperforming the excellent visual object detection model YOLOv8-x in terms of mAP0.35 increased by 13.72% in common scenarios and 24.85% in complex scenarios.

Autoregressive (AR) modeling has achieved remarkable success in natural language processing by enabling models to generate text with coherence and contextual understanding through next token prediction. Recently, in image generation, VAR proposes scale-wise autoregressive modeling, which extends the next token prediction to the next scale prediction, preserving the 2D structure of images. However, VAR encounters two primary challenges: (1) its complex and rigid scale design limits generalization in next scale prediction, and (2) the generator’s dependence on a discrete tokenizer with the same complex scale structure restricts modularity and flexibility in updating the tokenizer. To address these limitations, we introduce FlowAR, a general next scale prediction method featuring a streamlined scale design, where each subsequent scale is simply double the previous one. This eliminates the need for VAR’s intricate multi-scale residual tokenizer and enables the use of any off-the-shelf Variational AutoEncoder (VAE). Our simplified design enhances generalization in next scale prediction and facilitates the integration of Flow Matching for high-quality image synthesis. We validate the effectiveness of FlowAR on the challenging ImageNet-256 benchmark, demonstrating superior generation performance compared to previous methods. Codes is available at \href{https://github.com/OliverRensu/FlowAR}{https://github.com/OliverRensu/FlowAR}.