With the rapid evolution of autonomous driving technology and intelligent transportation systems, semantic segmentation has become increasingly critical. Precise interpretation and analysis of real-world environments are indispensable for these advanced applications. However, traditional semantic segmentation approaches frequently face challenges in balancing model performance with computational efficiency, especially regarding the volume of model parameters. To address these constraints, we propose SegRet, a novel model employing the Retentive Network (RetNet) architecture coupled with a lightweight residual decoder that integrates zero-initialization. SegRet offers three distinctive advantages: (1) Lightweight Residual Decoder: by embedding a zero-initialization layer within the residual network structure, the decoder remains computationally streamlined without sacrificing essential information propagation; (2) Robust Feature Extraction: adopting RetNet as its backbone enables SegRet to effectively capture hierarchical image features, thereby enriching the representation quality of extracted features; (3) Parameter Efficiency: SegRet attains state-of-the-art (SOTA) segmentation performance while markedly decreasing the number of parameters, ensuring high accuracy without imposing additional computational burdens. Comprehensive empirical evaluations on prominent benchmarks, such as ADE20K, Citycapes, and COCO-Stuff, highlight the effectiveness and superiority of our method.
Perceptual image compression has shown strong potential for producing visually appealing results at low bitrates, surpassing classical standards and pixel-wise distortion-oriented neural methods. However, existing methods typically improve compression performance by incorporating explicit semantic priors, such as segmentation maps and textual features, into the encoder or decoder, which increases model complexity by adding parameters and floating-point operations. This limits the model's practicality, as image compression often occurs on resource-limited mobile devices. To alleviate this problem, we propose a lightweight perceptual Image Compression method using Implicit Semantic Priors (ICISP). We first develop an enhanced visual state space block that exploits local and global spatial dependencies to reduce redundancy. Since different frequency information contributes unequally to compression, we develop a frequency decomposition modulation block to adaptively preserve or reduce the low-frequency and high-frequency information. We establish the above blocks as the main modules of the encoder-decoder, and to further improve the perceptual quality of the reconstructed images, we develop a semantic-informed discriminator that uses implicit semantic priors from a pretrained DINOv2 encoder. Experiments on popular benchmarks show that our method achieves competitive compression performance and has significantly fewer network parameters and floating point operations than the existing state-of-the-art.
Recent Multimodal Large Language Models (MLLMs) have achieved remarkable performance but face deployment challenges due to their quadratic computational complexity, growing Key-Value cache requirements, and reliance on separate vision encoders. We propose mmMamba, a framework for developing linear-complexity native multimodal state space models through progressive distillation from existing MLLMs using moderate academic computational resources. Our approach enables the direct conversion of trained decoder-only MLLMs to linear-complexity architectures without requiring pre-trained RNN-based LLM or vision encoders. We propose an seeding strategy to carve Mamba from trained Transformer and a three-stage distillation recipe, which can effectively transfer the knowledge from Transformer to Mamba while preserving multimodal capabilities. Our method also supports flexible hybrid architectures that combine Transformer and Mamba layers for customizable efficiency-performance trade-offs. Distilled from the Transformer-based decoder-only HoVLE, mmMamba-linear achieves competitive performance against existing linear and quadratic-complexity VLMs, while mmMamba-hybrid further improves performance significantly, approaching HoVLE's capabilities. At 103K tokens, mmMamba-linear demonstrates 20.6$\times$ speedup and 75.8% GPU memory reduction compared to HoVLE, while mmMamba-hybrid achieves 13.5$\times$ speedup and 60.2% memory savings. Code and models are released at https://github.com/hustvl/mmMamba
This paper proposes a new compression paradigm -- Guaranteed Conditional Diffusion with Tensor Correction (GCDTC) -- for lossy scientific data compression. The framework is based on recent conditional diffusion (CD) generative models, and it consists of a conditional diffusion model, tensor correction, and error guarantee. Our diffusion model is a mixture of 3D conditioning and 2D denoising U-Net. The approach leverages a 3D block-based compressing module to address spatiotemporal correlations in structured scientific data. Then, the reverse diffusion process for 2D spatial data is conditioned on the ``slices'' of content latent variables produced by the compressing module. After training, the denoising decoder reconstructs the data with zero noise and content latent variables, and thus it is entirely deterministic. The reconstructed outputs of the CD model are further post-processed by our tensor correction and error guarantee steps to control and ensure a maximum error distortion, which is an inevitable requirement in lossy scientific data compression. Our experiments involving two datasets generated by climate and chemical combustion simulations show that our framework outperforms standard convolutional autoencoders and yields competitive compression quality with an existing scientific data compression algorithm.
Accurate 3D shape representation is essential in engineering applications such as design, optimization, and simulation. In practice, engineering workflows require structured, part-aware representations, as objects are inherently designed as assemblies of distinct components. However, most existing methods either model shapes holistically or decompose them without predefined part structures, limiting their applicability in real-world design tasks. We propose PartSDF, a supervised implicit representation framework that explicitly models composite shapes with independent, controllable parts while maintaining shape consistency. Despite its simple single-decoder architecture, PartSDF outperforms both supervised and unsupervised baselines in reconstruction and generation tasks. We further demonstrate its effectiveness as a structured shape prior for engineering applications, enabling precise control over individual components while preserving overall coherence. Code available at https://github.com/cvlab-epfl/PartSDF.
While neural vocoders have made significant progress in high-fidelity speech synthesis, their application on polyphonic music has remained underexplored. In this work, we propose DisCoder, a neural vocoder that leverages a generative adversarial encoder-decoder architecture informed by a neural audio codec to reconstruct high-fidelity 44.1 kHz audio from mel spectrograms. Our approach first transforms the mel spectrogram into a lower-dimensional representation aligned with the Descript Audio Codec (DAC) latent space before reconstructing it to an audio signal using a fine-tuned DAC decoder. DisCoder achieves state-of-the-art performance in music synthesis on several objective metrics and in a MUSHRA listening study. Our approach also shows competitive performance in speech synthesis, highlighting its potential as a universal vocoder.
Nonlinear time-dependent partial differential equations are essential in modeling complex phenomena across diverse fields, yet they pose significant challenges due to their computational complexity, especially in higher dimensions. This study explores Quantum Recurrent Neural Networks within an encoder-decoder framework, integrating Variational Quantum Circuits into Gated Recurrent Units and Long Short-Term Memory networks. Using this architecture, the model efficiently compresses high-dimensional spatiotemporal data into a compact latent space, facilitating more efficient temporal evolution. We evaluate the algorithms on the Hamilton-Jacobi-Bellman equation, Burgers' equation, the Gray-Scott reaction-diffusion system, and the three dimensional Michaelis-Menten reaction-diffusion equation. The results demonstrate the superior performance of the quantum-based algorithms in capturing nonlinear dynamics, handling high-dimensional spaces, and providing stable solutions, highlighting their potential as an innovative tool in solving challenging and complex systems.
Accurately restoring topology is both challenging and crucial in tubular structure extraction tasks, such as blood vessel segmentation and road network extraction. Diverging from traditional approaches based on pixel-level classification, our proposed method, named GraphMorph, focuses on branch-level features of tubular structures to achieve more topologically accurate predictions. GraphMorph comprises two main components: a Graph Decoder and a Morph Module. Utilizing multi-scale features extracted from an image patch by the segmentation network, the Graph Decoder facilitates the learning of branch-level features and generates a graph that accurately represents the tubular structure in this patch. The Morph Module processes two primary inputs: the graph and the centerline probability map, provided by the Graph Decoder and the segmentation network, respectively. Employing a novel SkeletonDijkstra algorithm, the Morph Module produces a centerline mask that aligns with the predicted graph. Furthermore, we observe that employing centerline masks predicted by GraphMorph significantly reduces false positives in the segmentation task, which is achieved by a simple yet effective post-processing strategy. The efficacy of our method in the centerline extraction and segmentation tasks has been substantiated through experimental evaluations across various datasets. Source code will be released soon.
Generative models are spearheading recent progress in deep learning, showcasing strong promise for trajectory sampling in dynamical systems as well. However, whereas latent space modeling paradigms have transformed image and video generation, similar approaches are more difficult for most dynamical systems. Such systems -- from chemical molecule structures to collective human behavior -- are described by interactions of entities, making them inherently linked to connectivity patterns, entity conservation, and the traceability of entities over time. Our approach, LaM-SLidE (Latent Space Modeling of Spatial Dynamical Systems via Linked Entities), bridges the gap between: (1) keeping the traceability of individual entities in a latent system representation, and (2) leveraging the efficiency and scalability of recent advances in image and video generation, where pre-trained encoder and decoder enable generative modeling directly in latent space. The core idea of LaM-SLidE is the introduction of identifier representations (IDs) that enable the retrieval of entity properties and entity composition from latent system representations, thus fostering traceability. Experimentally, across different domains, we show that LaM-SLidE performs favorably in terms of speed, accuracy, and generalizability. Code is available at https://github.com/ml-jku/LaM-SLidE .
This paper presents ThinkDiff, a novel alignment paradigm that empowers text-to-image diffusion models with multimodal in-context understanding and reasoning capabilities by integrating the strengths of vision-language models (VLMs). Existing multimodal diffusion finetuning methods largely focus on pixel-level reconstruction rather than in-context reasoning, and are constrained by the complexity and limited availability of reasoning-based datasets. ThinkDiff addresses these challenges by leveraging vision-language training as a proxy task, aligning VLMs with the decoder of an encoder-decoder large language model (LLM) instead of a diffusion decoder. This proxy task builds on the observation that the $\textbf{LLM decoder}$ shares the same input feature space with $\textbf{diffusion decoders}$ that use the corresponding $\textbf{LLM encoder}$ for prompt embedding. As a result, aligning VLMs with diffusion decoders can be simplified through alignment with the LLM decoder. Without complex training and datasets, ThinkDiff effectively unleashes understanding, reasoning, and composing capabilities in diffusion models. Experiments demonstrate that ThinkDiff significantly improves accuracy from 19.2% to 46.3% on the challenging CoBSAT benchmark for multimodal in-context reasoning generation, with only 5 hours of training on 4 A100 GPUs. Additionally, ThinkDiff demonstrates exceptional performance in composing multiple images and texts into logically coherent images. Project page: https://mizhenxing.github.io/ThinkDiff.
Vision Language Models (VLMs) have achieved remarkable progress in multimodal tasks, yet they often struggle with visual arithmetic, seemingly simple capabilities like object counting or length comparison, which are essential for relevant complex tasks like chart understanding and geometric reasoning. In this work, we first investigate the root causes of this deficiency through a suite of probing tasks focusing on basic visual arithmetic. Our analysis reveals that while pre-trained vision encoders typically capture sufficient information, the text decoder often fails to decode it correctly for arithmetic reasoning. To address this, we propose CogAlign, a novel post-training strategy inspired by Piaget's theory of cognitive development. CogAlign trains VLMs to recognize invariant properties under visual transformations. We demonstrate that this approach significantly improves the performance of three diverse VLMs on our proposed probing tasks. Furthermore, CogAlign enhances performance by an average of 4.6% on CHOCOLATE and 2.9% on MATH-VISION, outperforming or matching supervised fine-tuning methods while requiring only 60% less training data. These results highlight the effectiveness and generalizability of CogAlign in improving fundamental visual arithmetic capabilities and their transfer to downstream tasks.
Lexico-semantic networks represent words as nodes and their semantic relatedness as edges. While such networks are traditionally constructed using embeddings from encoder-based models or static vectors, embeddings from decoder-only large language models (LLMs) remain underexplored. Unlike encoder models, LLMs are trained with a next-token prediction objective, which does not directly encode the meaning of the current token. In this paper, we construct lexico-semantic networks from the input embeddings of LLMs with varying parameter scales and conduct a comparative analysis of their global and local structures. Our results show that these networks exhibit small-world properties, characterized by high clustering and short path lengths. Moreover, larger LLMs yield more intricate networks with less small-world effects and longer paths, reflecting richer semantic structures and relations. We further validate our approach through analyses of common conceptual pairs, structured lexical relations derived from WordNet, and a cross-lingual semantic network for qualitative words.
While multilingual language models like XLM-R have advanced multilingualism in NLP, they still perform poorly in extremely low-resource languages. This situation is exacerbated by the fact that modern LLMs such as LLaMA and Qwen support far fewer languages than XLM-R, making text generation models non-existent for many languages in the world. To tackle this challenge, we propose a novel framework for adapting multilingual encoders to text generation in extremely low-resource languages. By reusing the weights between the encoder and the decoder, our framework allows the model to leverage the learned semantic space of the encoder, enabling efficient learning and effective generalization in low-resource languages. Applying this framework to four Chinese minority languages, we present XLM-SWCM, and demonstrate its superior performance on various downstream tasks even when compared with much larger models.
Molecules have various computational representations, including numerical descriptors, strings, graphs, point clouds, and surfaces. Each representation method enables the application of various machine learning methodologies from linear regression to graph neural networks paired with large language models. To complement existing representations, we introduce the representation of molecules through vector-valued functions, or $n$-dimensional vector fields, that are parameterized by neural networks, which we denote molecular neural fields. Unlike surface representations, molecular neural fields capture external features and the hydrophobic core of macromolecules such as proteins. Compared to discrete graph or point representations, molecular neural fields are compact, resolution independent and inherently suited for interpolation in spatial and temporal dimensions. These properties inherited by molecular neural fields lend themselves to tasks including the generation of molecules based on their desired shape, structure, and composition, and the resolution-independent interpolation between molecular conformations in space and time. Here, we provide a framework and proofs-of-concept for molecular neural fields, namely, the parametrization and superresolution reconstruction of a protein-ligand complex using an auto-decoder architecture and the embedding of molecular volumes in latent space using an auto-encoder architecture.
Reduced Order Modeling is of paramount importance for efficiently inferring high-dimensional spatio-temporal fields in parametric contexts, enabling computationally tractable parametric analyses, uncertainty quantification and control. However, conventional dimensionality reduction techniques are typically limited to known and constant parameters, inefficient for nonlinear and chaotic dynamics, and uninformed to the actual system behavior. In this work, we propose sensor-driven SHallow REcurrent Decoder networks for Reduced Order Modeling (SHRED-ROM). Specifically, we consider the composition of a long short-term memory network, which encodes the temporal dynamics of limited sensor data in multiple scenarios, and a shallow decoder, which reconstructs the corresponding high-dimensional states. SHRED-ROM is a robust decoding-only strategy that circumvents the numerically unstable approximation of an inverse which is required by encoding-decoding schemes. To enhance computational efficiency and memory usage, the full-order state snapshots are reduced by, e.g., proper orthogonal decomposition, allowing for compressive training of the networks with minimal hyperparameter tuning. Through applications on chaotic and nonlinear fluid dynamics, we show that SHRED-ROM (i) accurately reconstructs the state dynamics for new parameter values starting from limited fixed or mobile sensors, independently on sensor placement, (ii) can cope with both physical, geometrical and time-dependent parametric dependencies, while being agnostic to their actual values, (iii) can accurately estimate unknown parameters, and (iv) can deal with different data sources, such as high-fidelity simulations, coupled fields and videos.
Current image tampering localization methods primarily rely on Convolutional Neural Networks (CNNs) and Transformers. While CNNs suffer from limited local receptive fields, Transformers offer global context modeling at the expense of quadratic computational complexity. Recently, the state space model Mamba has emerged as a competitive alternative, enabling linear-complexity global dependency modeling. Inspired by it, we propose a lightweight and effective FORensic network based on vision MAmba (ForMa) for blind image tampering localization. Firstly, ForMa captures multi-scale global features that achieves efficient global dependency modeling through linear complexity. Then the pixel-wise localization map is generated by a lightweight decoder, which employs a parameter-free pixel shuffle layer for upsampling. Additionally, a noise-assisted decoding strategy is proposed to integrate complementary manipulation traces from tampered images, boosting decoder sensitivity to forgery cues. Experimental results on 10 standard datasets demonstrate that ForMa achieves state-of-the-art generalization ability and robustness, while maintaining the lowest computational complexity. Code is available at https://github.com/multimediaFor/ForMa.
The deployment of advanced deep learning models for medical image segmentation is often constrained by the requirement for extensively annotated datasets. Weakly-supervised learning, which allows less precise labels, has become a promising solution to this challenge. Building on this approach, we propose QMaxViT-Unet+, a novel framework for scribble-supervised medical image segmentation. This framework is built on the U-Net architecture, with the encoder and decoder replaced by Multi-Axis Vision Transformer (MaxViT) blocks. These blocks enhance the model's ability to learn local and global features efficiently. Additionally, our approach integrates a query-based Transformer decoder to refine features and an edge enhancement module to compensate for the limited boundary information in the scribble label. We evaluate the proposed QMaxViT-Unet+ on four public datasets focused on cardiac structures, colorectal polyps, and breast cancer: ACDC, MS-CMRSeg, SUN-SEG, and BUSI. Evaluation metrics include the Dice similarity coefficient (DSC) and the 95th percentile of Hausdorff distance (HD95). Experimental results show that QMaxViT-Unet+ achieves 89.1\% DSC and 1.316mm HD95 on ACDC, 88.4\% DSC and 2.226mm HD95 on MS-CMRSeg, 71.4\% DSC and 4.996mm HD95 on SUN-SEG, and 69.4\% DSC and 50.122mm HD95 on BUSI. These results demonstrate that our method outperforms existing approaches in terms of accuracy, robustness, and efficiency while remaining competitive with fully-supervised learning approaches. This makes it ideal for medical image analysis, where high-quality annotations are often scarce and require significant effort and expense. The code is available at: https://github.com/anpc849/QMaxViT-Unet
An emerging research direction in NMT involves the use of Quality Estimation (QE) models, which have demonstrated high correlations with human judgment and can enhance translations through Quality-Aware Decoding. Although several approaches have been proposed based on sampling multiple candidate translations, none have integrated these models directly into the decoding process. In this paper, we address this by proposing a novel token-level QE model capable of reliably scoring partial translations. We build a uni-directional QE model for this, as decoder models are inherently trained and efficient on partial sequences. We then present a decoding strategy that integrates the QE model for Quality-Aware decoding and demonstrate that the translation quality improves when compared to the N-best list re-ranking with state-of-the-art QE models (upto $1.39$ XCOMET-XXL $\uparrow$). Finally, we show that our approach provides significant benefits in document translation tasks, where the quality of N-best lists is typically suboptimal.
This paper introduces the Large Memory Model (LM2), a decoder-only Transformer architecture enhanced with an auxiliary memory module that aims to address the limitations of standard Transformers in multi-step reasoning, relational argumentation, and synthesizing information distributed over long contexts. The proposed LM2 incorporates a memory module that acts as a contextual representation repository, interacting with input tokens via cross attention and updating through gating mechanisms. To preserve the Transformers general-purpose capabilities, LM2 maintains the original information flow while integrating a complementary memory pathway. Experimental results on the BABILong benchmark demonstrate that the LM2model outperforms both the memory-augmented RMT model by 37.1% and the baseline Llama-3.2 model by 86.3% on average across tasks. LM2 exhibits exceptional capabilities in multi-hop inference, numerical reasoning, and large-context question-answering. On the MMLU dataset, it achieves a 5.0% improvement over a pre-trained vanilla model, demonstrating that its memory module does not degrade performance on general tasks. Further, in our analysis, we explore the memory interpretability, effectiveness of memory modules, and test-time behavior. Our findings emphasize the importance of explicit memory in enhancing Transformer architectures.
Recent advancements in large vision-language models have enabled highly expressive and diverse vector sketch generation. However, state-of-the-art methods rely on a time-consuming optimization process involving repeated feedback from a pretrained model to determine stroke placement. Consequently, despite producing impressive sketches, these methods are limited in practical applications. In this work, we introduce SwiftSketch, a diffusion model for image-conditioned vector sketch generation that can produce high-quality sketches in less than a second. SwiftSketch operates by progressively denoising stroke control points sampled from a Gaussian distribution. Its transformer-decoder architecture is designed to effectively handle the discrete nature of vector representation and capture the inherent global dependencies between strokes. To train SwiftSketch, we construct a synthetic dataset of image-sketch pairs, addressing the limitations of existing sketch datasets, which are often created by non-artists and lack professional quality. For generating these synthetic sketches, we introduce ControlSketch, a method that enhances SDS-based techniques by incorporating precise spatial control through a depth-aware ControlNet. We demonstrate that SwiftSketch generalizes across diverse concepts, efficiently producing sketches that combine high fidelity with a natural and visually appealing style.