Rethinking Discrete Tokens: Treating Them as Conditions for Continuous Autoregressive Image Synthesis

Recent advances in large language models (LLMs) have spurred interests in encoding images as discrete tokens and leveraging autoregressive (AR) frameworks for visual generation. However, the quantization process in AR-based visual generation models inherently introduces information loss that degrades image fidelity. To mitigate this limitation, recent studies have explored to autoregressively predict continuous tokens. Unlike discrete tokens that reside in a structured and bounded space, continuous representations exist in an unbounded, high-dimensional space, making density estimation more challenging and increasing the risk of generating out-of-distribution artifacts. Based on the above findings, this work introduces DisCon (Discrete-Conditioned Continuous Autoregressive Model), a novel framework that reinterprets discrete tokens as conditional signals rather than generation targets. By modeling the conditional probability of continuous representations conditioned on discrete tokens, DisCon circumvents the optimization challenges of continuous token modeling while avoiding the information loss caused by quantization. DisCon achieves a gFID score of 1.38 on ImageNet 256$\times$256 generation, outperforming state-of-the-art autoregressive approaches by a clear margin.

Aerial Vision-and-Language Navigation with Grid-based View Selection and Map Construction

Aerial Vision-and-Language Navigation (Aerial VLN) aims to obtain an unmanned aerial vehicle agent to navigate aerial 3D environments following human instruction. Compared to ground-based VLN, aerial VLN requires the agent to decide the next action in both horizontal and vertical directions based on the first-person view observations. Previous methods struggle to perform well due to the longer navigation path, more complicated 3D scenes, and the neglect of the interplay between vertical and horizontal actions. In this paper, we propose a novel grid-based view selection framework that formulates aerial VLN action prediction as a grid-based view selection task, incorporating vertical action prediction in a manner that accounts for the coupling with horizontal actions, thereby enabling effective altitude adjustments. We further introduce a grid-based bird's eye view map for aerial space to fuse the visual information in the navigation history, provide contextual scene information, and mitigate the impact of obstacles. Finally, a cross-modal transformer is adopted to explicitly align the long navigation history with the instruction. We demonstrate the superiority of our method in extensive experiments.

OverLoCK: An Overview-first-Look-Closely-next ConvNet with Context-Mixing Dynamic Kernels

In the human vision system, top-down attention plays a crucial role in perception, wherein the brain initially performs an overall but rough scene analysis to extract salient cues (i.e., overview first), followed by a finer-grained examination to make more accurate judgments (i.e., look closely next). However, recent efforts in ConvNet designs primarily focused on increasing kernel size to obtain a larger receptive field without considering this crucial biomimetic mechanism to further improve performance. To this end, we propose a novel pure ConvNet vision backbone, termed OverLoCK, which is carefully devised from both the architecture and mixer perspectives. Specifically, we introduce a biomimetic Deep-stage Decomposition Strategy (DDS) that fuses semantically meaningful context representations into middle and deep layers by providing dynamic top-down context guidance at both feature and kernel weight levels. To fully unleash the power of top-down context guidance, we further propose a novel \textbf{Cont}ext-\textbf{Mix}ing Dynamic Convolution (ContMix) that effectively models long-range dependencies while preserving inherent local inductive biases even when the input resolution increases. These properties are absent in previous convolutions. With the support from both DDS and ContMix, our OverLoCK exhibits notable performance improvement over existing methods. For instance, OverLoCK-T achieves a Top-1 accuracy of 84.2\%, significantly surpassing ConvNeXt-B while only using around one-third of the FLOPs/parameters. On object detection with Cascade Mask R-CNN, our OverLoCK-S surpasses MogaNet-B by a significant 1\% in AP$^b$. On semantic segmentation with UperNet, our OverLoCK-T remarkably improves UniRepLKNet-T by 1.7\% in mIoU. Code is publicly available at https://github.com/LMMMEng/OverLoCK.

SuperNeRF-GAN: A Universal 3D-Consistent Super-Resolution Framework for Efficient and Enhanced 3D-Aware Image Synthesis

Neural volume rendering techniques, such as NeRF, have revolutionized 3D-aware image synthesis by enabling the generation of images of a single scene or object from various camera poses. However, the high computational cost of NeRF presents challenges for synthesizing high-resolution (HR) images. Most existing methods address this issue by leveraging 2D super-resolution, which compromise 3D-consistency. Other methods propose radiance manifolds or two-stage generation to achieve 3D-consistent HR synthesis, yet they are limited to specific synthesis tasks, reducing their universality. To tackle these challenges, we propose SuperNeRF-GAN, a universal framework for 3D-consistent super-resolution. A key highlight of SuperNeRF-GAN is its seamless integration with NeRF-based 3D-aware image synthesis methods and it can simultaneously enhance the resolution of generated images while preserving 3D-consistency and reducing computational cost. Specifically, given a pre-trained generator capable of producing a NeRF representation such as tri-plane, we first perform volume rendering to obtain a low-resolution image with corresponding depth and normal map. Then, we employ a NeRF Super-Resolution module which learns a network to obtain a high-resolution NeRF. Next, we propose a novel Depth-Guided Rendering process which contains three simple yet effective steps, including the construction of a boundary-correct multi-depth map through depth aggregation, a normal-guided depth super-resolution and a depth-guided NeRF rendering. Experimental results demonstrate the superior efficiency, 3D-consistency, and quality of our approach. Additionally, ablation studies confirm the effectiveness of our proposed components.

Real-time One-Step Diffusion-based Expressive Portrait Videos Generation

Latent diffusion models have made great strides in generating expressive portrait videos with accurate lip-sync and natural motion from a single reference image and audio input. However, these models are far from real-time, often requiring many sampling steps that take minutes to generate even one second of video-significantly limiting practical use. We introduce OSA-LCM (One-Step Avatar Latent Consistency Model), paving the way for real-time diffusion-based avatars. Our method achieves comparable video quality to existing methods but requires only one sampling step, making it more than 10x faster. To accomplish this, we propose a novel avatar discriminator design that guides lip-audio consistency and motion expressiveness to enhance video quality in limited sampling steps. Additionally, we employ a second-stage training architecture using an editing fine-tuned method (EFT), transforming video generation into an editing task during training to effectively address the temporal gap challenge in single-step generation. Experiments demonstrate that OSA-LCM outperforms existing open-source portrait video generation models while operating more efficiently with a single sampling step.

SegMAN: Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation

High-quality semantic segmentation relies on three key capabilities: global context modeling, local detail encoding, and multi-scale feature extraction. However, recent methods struggle to possess all these capabilities simultaneously. Hence, we aim to empower segmentation networks to simultaneously carry out efficient global context modeling, high-quality local detail encoding, and rich multi-scale feature representation for varying input resolutions. In this paper, we introduce SegMAN, a novel linear-time model comprising a hybrid feature encoder dubbed SegMAN Encoder, and a decoder based on state space models. Specifically, the SegMAN Encoder synergistically integrates sliding local attention with dynamic state space models, enabling highly efficient global context modeling while preserving fine-grained local details. Meanwhile, the MMSCopE module in our decoder enhances multi-scale context feature extraction and adaptively scales with the input resolution. We comprehensively evaluate SegMAN on three challenging datasets: ADE20K, Cityscapes, and COCO-Stuff. For instance, SegMAN-B achieves 52.6% mIoU on ADE20K, outperforming SegNeXt-L by 1.6% mIoU while reducing computational complexity by over 15% GFLOPs. On Cityscapes, SegMAN-B attains 83.8% mIoU, surpassing SegFormer-B3 by 2.1% mIoU with approximately half the GFLOPs. Similarly, SegMAN-B improves upon VWFormer-B3 by 1.6% mIoU with lower GFLOPs on the COCO-Stuff dataset. Our code is available at https://github.com/yunxiangfu2001/SegMAN.

Enhanced MRI Representation via Cross-series Masking

Magnetic resonance imaging (MRI) is indispensable for diagnosing and planning treatment in various medical conditions due to its ability to produce multi-series images that reveal different tissue characteristics. However, integrating these diverse series to form a coherent analysis presents significant challenges, such as differing spatial resolutions and contrast patterns meanwhile requiring extensive annotated data, which is scarce in clinical practice. Due to these issues, we introduce a novel Cross-Series Masking (CSM) Strategy for effectively learning MRI representation in a self-supervised manner. Specifically, CSM commences by randomly sampling a subset of regions and series, which are then strategically masked. In the training process, the cross-series representation is learned by utilizing the unmasked data to reconstruct the masked portions. This process not only integrates information across different series but also facilitates the ability to model both intra-series and inter-series correlations and complementarities. With the learned representation, the downstream tasks like segmentation and classification are also enhanced. Taking brain tissue segmentation, breast tumor benign/malignant classification, and prostate cancer diagnosis as examples, our method achieves state-of-the-art performance on both public and in-house datasets.

SparX: A Sparse Cross-Layer Connection Mechanism for Hierarchical Vision Mamba and Transformer Networks

Due to the capability of dynamic state space models (SSMs) in capturing long-range dependencies with near-linear computational complexity, Mamba has shown notable performance in NLP tasks. This has inspired the rapid development of Mamba-based vision models, resulting in promising results in visual recognition tasks. However, such models are not capable of distilling features across layers through feature aggregation, interaction, and selection. Moreover, existing cross-layer feature aggregation methods designed for CNNs or ViTs are not practical in Mamba-based models due to high computational costs. Therefore, this paper aims to introduce an efficient cross-layer feature aggregation mechanism for Mamba-based vision backbone networks. Inspired by the Retinal Ganglion Cells (RGCs) in the human visual system, we propose a new sparse cross-layer connection mechanism termed SparX to effectively improve cross-layer feature interaction and reuse. Specifically, we build two different types of network layers: ganglion layers and normal layers. The former has higher connectivity and complexity, enabling multi-layer feature aggregation and interaction in an input-dependent manner. In contrast, the latter has lower connectivity and complexity. By interleaving these two types of layers, we design a new vision backbone network with sparsely cross-connected layers, achieving an excellent trade-off among model size, computational cost, memory cost, and accuracy in comparison to its counterparts. For instance, with fewer parameters, SparX-Mamba-T improves the top-1 accuracy of VMamba-T from 82.5% to 83.5%, while SparX-Swin-T achieves a 1.3% increase in top-1 accuracy compared to Swin-T. Extensive experimental results demonstrate that our new connection mechanism possesses both superior performance and generalization capabilities on various vision tasks.

Autoregressive Sequence Modeling for 3D Medical Image Representation

Three-dimensional (3D) medical images, such as Computed Tomography (CT) and Magnetic Resonance Imaging (MRI), are essential for clinical applications. However, the need for diverse and comprehensive representations is particularly pronounced when considering the variability across different organs, diagnostic tasks, and imaging modalities. How to effectively interpret the intricate contextual information and extract meaningful insights from these images remains an open challenge to the community. While current self-supervised learning methods have shown potential, they often consider an image as a whole thereby overlooking the extensive, complex relationships among local regions from one or multiple images. In this work, we introduce a pioneering method for learning 3D medical image representations through an autoregressive pre-training framework. Our approach sequences various 3D medical images based on spatial, contrast, and semantic correlations, treating them as interconnected visual tokens within a token sequence. By employing an autoregressive sequence modeling task, we predict the next visual token in the sequence, which allows our model to deeply understand and integrate the contextual information inherent in 3D medical images. Additionally, we implement a random startup strategy to avoid overestimating token relationships and to enhance the robustness of learning. The effectiveness of our approach is demonstrated by the superior performance over others on nine downstream tasks in public datasets.

LaMamba-Diff: Linear-Time High-Fidelity Diffusion Models Based on Local Attention and Mamba

Recent Transformer-based diffusion models have shown remarkable performance, largely attributed to the ability of the self-attention mechanism to accurately capture both global and local contexts by computing all-pair interactions among input tokens. However, their quadratic complexity poses significant computational challenges for long-sequence inputs. Conversely, a recent state space model called Mamba offers linear complexity by compressing a filtered global context into a hidden state. Despite its efficiency, compression inevitably leads to information loss of fine-grained local dependencies among tokens, which are crucial for effective visual generative modeling. Motivated by these observations, we introduce Local Attentional Mamba (LaMamba) blocks that combine the strengths of self-attention and Mamba, capturing both global contexts and local details with linear complexity. Leveraging the efficient U-Net architecture, our model exhibits exceptional scalability and surpasses the performance of DiT across various model scales on ImageNet at 256x256 resolution, all while utilizing substantially fewer GFLOPs and a comparable number of parameters. Compared to state-of-the-art diffusion models on ImageNet 256x256 and 512x512, our largest model presents notable advantages, such as a reduction of up to 62\% GFLOPs compared to DiT-XL/2, while achieving superior performance with comparable or fewer parameters.