arXiv subjects

Hongming Shan

Publications and source records attributed to Hongming Shan.

At least 37 records · Page 2Linked to original sources

arXiv2026

Routing Matters in MoE: Scaling Diffusion Transformers with Explicit Routing Guidance

Mixture-of-Experts (MoE) has emerged as a powerful paradigm for scaling model capacity while preserving computational efficiency. Despite its notable success in large language models (LLMs), existing attempts to apply MoE to Diffusion Transformers (DiTs) have yielded limited gains. We attribute this gap to fundamental differences between language and visual tokens. Language tokens are semantically dense with pronounced inter-token variation, while visual tokens exhibit spatial redundancy and functional heterogeneity, hindering expert specialization in vision MoE. To this end, we present ProMoE, an MoE framework featuring a two-step router with explicit routing guidance that promotes expert specialization. Specifically, this guidance encourages the router to partition image tokens into conditional and unconditional sets via conditional routing according to their functional roles, and refine the assignments of conditional image tokens through prototypical routing with learnable prototypes based on semantic content. Moreover, the similarity-based expert allocation in latent space enabled by prototypical routing offers a natural mechanism for incorporating explicit semantic guidance, and we validate that such guidance is crucial for vision MoE. Building on this, we propose a routing contrastive loss that explicitly enhances the prototypical routing process, promoting intra-expert coherence and inter-expert diversity. Extensive experiments on ImageNet benchmark demonstrate that ProMoE surpasses state-of-the-art methods under both Rectified Flow and DDPM training objectives. Code is available at https://github.com/ali-vilab/ProMoE.

cs.CV↗

arXiv2026

Dynamic Differential Linear Attention: Enhancing Linear Diffusion Transformer for High-Quality Image Generation

Diffusion transformers (DiTs) have emerged as a powerful architecture for high-fidelity image generation, yet the quadratic cost of self-attention poses a major scalability bottleneck. To address this, linear attention mechanisms have been adopted to reduce computational cost; unfortunately, the resulting linear diffusion transformers (LiTs) models often come at the expense of generative performance, frequently producing over-smoothed attention weights that limit expressiveness. In this work, we introduce Dynamic Differential Linear Attention (DyDiLA), a novel linear attention formulation that enhances the effectiveness of LiTs by mitigating the oversmoothing issue and improving generation quality. Specifically, the novelty of DyDiLA lies in three key designs: (i) dynamic projection module, which facilitates the decoupling of token representations by learning with dynamically assigned knowledge; (ii) dynamic measure kernel, which provides a better similarity measurement to capture fine-grained semantic distinctions between tokens by dynamically assigning kernel functions for token processing; and (iii) token differential operator, which enables more robust query-to-key retrieval by calculating the differences between the tokens and their corresponding information redundancy produced by dynamic measure kernel. To capitalize on DyDiLA, we introduce a refined LiT, termed DyDi-LiT, that systematically incorporates our advancements. Extensive experiments show that DyDi-LiT consistently outperforms current state-of-the-art (SOTA) models across multiple metrics, underscoring its strong practical potential.

cs.CV↗

arXiv2025

RepLDM: Reprogramming Pretrained Latent Diffusion Models for High-Quality, High-Efficiency, High-Resolution Image Generation

While latent diffusion models (LDMs), such as Stable Diffusion, are designed for high-resolution (HR) image generation, they often struggle with significant structural distortions when generating images at resolutions higher than their training one. Instead of relying on extensive retraining, a more resource-efficient approach is to reprogram the pretrained model for HR image generation; however, existing methods often result in poor image quality and long inference time. We introduce RepLDM, a novel reprogramming framework for pretrained LDMs that enables high-quality, high-efficiency, high-resolution image generation; see Fig. 1. RepLDM consists of two stages: (i) an attention guidance stage, which generates a latent representation of a higher-quality training-resolution image using a novel training-free self-attention mechanism to enhance the structural consistency; and (ii) a progressive upsampling stage, which progressively performs upsampling in pixel space to mitigate the severe artifacts caused by latent space upsampling. The effective initialization from the first stage allows for denoising at higher resolutions with significantly fewer steps, improving the efficiency. Extensive experimental results demonstrate that RepLDM significantly outperforms state-of-the-art methods in both quality and efficiency for HR image generation, underscoring its advantages for real-world applications. Codes: https://github.com/kmittle/RepLDM.

cs.CV↗

arXiv2025

A Survey of Scientific Large Language Models: From Data Foundations to Agent Frontiers

Scientific Large Language Models (Sci-LLMs) are transforming how knowledge is represented, integrated, and applied in scientific research, yet their progress is shaped by the complex nature of scientific data. This survey presents a comprehensive, data-centric synthesis that reframes the development of Sci-LLMs as a co-evolution between models and their underlying data substrate. We formulate a unified taxonomy of scientific data and a hierarchical model of scientific knowledge, emphasizing the multimodal, cross-scale, and domain-specific challenges that differentiate scientific corpora from general natural language processing datasets. We systematically review recent Sci-LLMs, from general-purpose foundations to specialized models across diverse scientific disciplines, alongside an extensive analysis of over 270 pre-/post-training datasets, showing why Sci-LLMs pose distinct demands -- heterogeneous, multi-scale, uncertainty-laden corpora that require representations preserving domain invariance and enabling cross-modal reasoning. On evaluation, we examine over 190 benchmark datasets and trace a shift from static exams toward process- and discovery-oriented assessments with advanced evaluation protocols. These data-centric analyses highlight persistent issues in scientific data development and discuss emerging solutions involving semi-automated annotation pipelines and expert validation. Finally, we outline a paradigm shift toward closed-loop systems where autonomous agents based on Sci-LLMs actively experiment, validate, and contribute to a living, evolving knowledge base. Collectively, this work provides a roadmap for building trustworthy, continually evolving artificial intelligence (AI) systems that function as a true partner in accelerating scientific discovery.

cs.CL↗

arXiv2025

Towards Interpretable Counterfactual Generation via Multimodal Autoregression

Counterfactual medical image generation enables clinicians to explore clinical hypotheses, such as predicting disease progression, facilitating their decision-making. While existing methods can generate visually plausible images from disease progression prompts, they produce silent predictions that lack interpretation to verify how the generation reflects the hypothesized progression -- a critical gap for medical applications that require traceable reasoning. In this paper, we propose Interpretable Counterfactual Generation (ICG), a novel task requiring the joint generation of counterfactual images that reflect the clinical hypothesis and interpretation texts that outline the visual changes induced by the hypothesis. To enable ICG, we present ICG-CXR, the first dataset pairing longitudinal medical images with hypothetical progression prompts and textual interpretations. We further introduce ProgEmu, an autoregressive model that unifies the generation of counterfactual images and textual interpretations. We demonstrate the superiority of ProgEmu in generating progression-aligned counterfactuals and interpretations, showing significant potential in enhancing clinical decision support and medical education. Project page: https://progemu.github.io.

eess.IV↗

arXiv2025

UniSino: Physics-Driven Foundational Model for Universal CT Sinogram Standardization

During raw-data acquisition in CT imaging, diverse factors can degrade the collected sinograms, with undersampling and noise leading to severe artifacts and noise in reconstructed images and compromising diagnostic accuracy. Conventional correction methods rely on manually designed algorithms or fixed empirical parameters, but these approaches often lack generalizability across heterogeneous artifact types. To address these limitations, we propose UniSino, a foundation model for universal CT sinogram standardization. Unlike existing foundational models that operate in image domain, UniSino directly standardizes data in the projection domain, which enables stronger generalization across diverse undersampling scenarios. Its training framework incorporates the physical characteristics of sinograms, enhancing generalization and enabling robust performance across multiple subtasks spanning four benchmark datasets. Experimental results demonstrate thatUniSino achieves superior reconstruction quality both single and mixed undersampling case, demonstrating exceptional robustness and generalization in sinogram enhancement for CT imaging. The code is available at: https://github.com/yqx7150/UniSino.

cs.CV↗

arXiv2025

LangMamba: A Language-driven Mamba Framework for Low-dose CT Denoising with Vision-language Models

Low-dose computed tomography (LDCT) reduces radiation exposure but often degrades image quality, potentially compromising diagnostic accuracy. Existing deep learning-based denoising methods focus primarily on pixel-level mappings, overlooking the potential benefits of high-level semantic guidance. Recent advances in vision-language models (VLMs) suggest that language can serve as a powerful tool for capturing structured semantic information, offering new opportunities to improve LDCT reconstruction. In this paper, we introduce LangMamba, a Language-driven Mamba framework for LDCT denoising that leverages VLM-derived representations to enhance supervision from normal-dose CT (NDCT). LangMamba follows a two-stage learning strategy. First, we pre-train a Language-guided AutoEncoder (LangAE) that leverages frozen VLMs to map NDCT images into a semantic space enriched with anatomical information. Second, we synergize LangAE with two key components to guide LDCT denoising: Semantic-Enhanced Efficient Denoiser (SEED), which enhances NDCT-relevant local semantic while capturing global features with efficient Mamba mechanism, and Language-engaged Dual-space Alignment (LangDA) Loss, which ensures that denoised images align with NDCT in both perceptual and semantic spaces. Extensive experiments on two public datasets demonstrate that LangMamba outperforms conventional state-of-the-art methods, significantly improving detail preservation and visual fidelity. Remarkably, LangAE exhibits strong generalizability to unseen datasets, thereby reducing training costs. Furthermore, LangDA loss improves explainability by integrating language-guided insights into image reconstruction and offers a plug-and-play fashion. Our findings shed new light on the potential of language as a supervisory signal to advance LDCT denoising. The code is publicly available on https://github.com/hao1635/LangMamba.

eess.IV↗

arXiv2025

Noise-Inspired Diffusion Model for Generalizable Low-Dose CT Reconstruction

The generalization of deep learning-based low-dose computed tomography (CT) reconstruction models to doses unseen in the training data is important and remains challenging. Previous efforts heavily rely on paired data to improve the generalization performance and robustness through collecting either diverse CT data for re-training or a few test data for fine-tuning. Recently, diffusion models have shown promising and generalizable performance in low-dose CT (LDCT) reconstruction, however, they may produce unrealistic structures due to the CT image noise deviating from Gaussian distribution and imprecise prior information from the guidance of noisy LDCT images. In this paper, we propose a noise-inspired diffusion model for generalizable LDCT reconstruction, termed NEED, which tailors diffusion models for noise characteristics of each domain. First, we propose a novel shifted Poisson diffusion model to denoise projection data, which aligns the diffusion process with the noise model in pre-log LDCT projections. Second, we devise a doubly guided diffusion model to refine reconstructed images, which leverages LDCT images and initial reconstructions to more accurately locate prior information and enhance reconstruction fidelity. By cascading these two diffusion models for dual-domain reconstruction, our NEED requires only normal-dose data for training and can be effectively extended to various unseen dose levels during testing via a time step matching strategy. Extensive qualitative, quantitative, and segmentation-based evaluations on two datasets demonstrate that our NEED consistently outperforms state-of-the-art methods in reconstruction and generalization performance. Source code is made available at https://github.com/qgao21/NEED.

eess.IV↗

arXiv2025

PROTOCOL: Partial Optimal Transport-enhanced Contrastive Learning for Imbalanced Multi-view Clustering

While contrastive multi-view clustering has achieved remarkable success, it implicitly assumes balanced class distribution. However, real-world multi-view data primarily exhibits class imbalance distribution. Consequently, existing methods suffer performance degradation due to their inability to perceive and model such imbalance. To address this challenge, we present the first systematic study of imbalanced multi-view clustering, focusing on two fundamental problems: i. perceiving class imbalance distribution, and ii. mitigating representation degradation of minority samples. We propose PROTOCOL, a novel PaRtial Optimal TranspOrt-enhanced COntrastive Learning framework for imbalanced multi-view clustering. First, for class imbalance perception, we map multi-view features into a consensus space and reformulate the imbalanced clustering as a partial optimal transport (POT) problem, augmented with progressive mass constraints and weighted KL divergence for class distributions. Second, we develop a POT-enhanced class-rebalanced contrastive learning at both feature and class levels, incorporating logit adjustment and class-sensitive learning to enhance minority sample representations. Extensive experiments demonstrate that PROTOCOL significantly improves clustering performance on imbalanced multi-view data, filling a critical research gap in this field.

cs.LG↗

arXiv2025

Shushing! Let's Imagine an Authentic Speech from the Silent Video

Vision-guided speech generation aims to produce authentic speech from facial appearance or lip motions without relying on auditory signals, offering significant potential for applications such as dubbing in filmmaking and assisting individuals with aphonia. Despite recent progress, existing methods struggle to achieve unified cross-modal alignment across semantics, timbre, and emotional prosody from visual cues, prompting us to propose Consistent Video-to-Speech (CV2S) as an extended task to enhance cross-modal consistency. To tackle emerging challenges, we introduce ImaginTalk, a novel cross-modal diffusion framework that generates faithful speech using only visual input, operating within a discrete space. Specifically, we propose a discrete lip aligner that predicts discrete speech tokens from lip videos to capture semantic information, while an error detector identifies misaligned tokens, which are subsequently refined through masked language modeling with BERT. To further enhance the expressiveness of the generated speech, we develop a style diffusion transformer equipped with a face-style adapter that adaptively customizes identity and prosody dynamics across both the channel and temporal dimensions while ensuring synchronization with lip-aware semantic features. Extensive experiments demonstrate that ImaginTalk can generate high-fidelity speech with more accurate semantic details and greater expressiveness in timbre and emotion compared to state-of-the-art baselines. Demos are shown at our project page: https://imagintalk.github.io.

cs.CV↗

arXiv2025

DreamRelation: Relation-Centric Video Customization

Relational video customization refers to the creation of personalized videos that depict user-specified relations between two subjects, a crucial task for comprehending real-world visual content. While existing methods can personalize subject appearances and motions, they still struggle with complex relational video customization, where precise relational modeling and high generalization across subject categories are essential. The primary challenge arises from the intricate spatial arrangements, layout variations, and nuanced temporal dynamics inherent in relations; consequently, current models tend to overemphasize irrelevant visual details rather than capturing meaningful interactions. To address these challenges, we propose DreamRelation, a novel approach that personalizes relations through a small set of exemplar videos, leveraging two key components: Relational Decoupling Learning and Relational Dynamics Enhancement. First, in Relational Decoupling Learning, we disentangle relations from subject appearances using relation LoRA triplet and hybrid mask training strategy, ensuring better generalization across diverse relationships. Furthermore, we determine the optimal design of relation LoRA triplet by analyzing the distinct roles of the query, key, and value features within MM-DiT's attention mechanism, making DreamRelation the first relational video generation framework with explainable components. Second, in Relational Dynamics Enhancement, we introduce space-time relational contrastive loss, which prioritizes relational dynamics while minimizing the reliance on detailed subject appearances. Extensive experiments demonstrate that DreamRelation outperforms state-of-the-art methods in relational video customization. Code and models will be made publicly available.

cs.CV↗

arXiv2025

Patient-Level Anatomy Meets Scanning-Level Physics: Personalized Federated Low-Dose CT Denoising Empowered by Large Language Model

Reducing radiation doses benefits patients, however, the resultant low-dose computed tomography (LDCT) images often suffer from clinically unacceptable noise and artifacts. While deep learning (DL) shows promise in LDCT reconstruction, it requires large-scale data collection from multiple clients, raising privacy concerns. Federated learning (FL) has been introduced to address these privacy concerns; however, current methods are typically tailored to specific scanning protocols, which limits their generalizability and makes them less effective for unseen protocols. To address these issues, we propose SCAN-PhysFed, a novel SCanning- and ANatomy-level personalized Physics-Driven Federated learning paradigm for LDCT reconstruction. Since the noise distribution in LDCT data is closely tied to scanning protocols and anatomical structures being scanned, we design a dual-level physics-informed way to address these challenges. Specifically, we incorporate physical and anatomical prompts into our physics-informed hypernetworks to capture scanning- and anatomy-specific information, enabling dual-level physics-driven personalization of imaging features. These prompts are derived from the scanning protocol and the radiology report generated by a medical large language model (MLLM), respectively. Subsequently, client-specific decoders project these dual-level personalized imaging features back into the image domain. Besides, to tackle the challenge of unseen data, we introduce a novel protocol vector-quantization strategy (PVQS), which ensures consistent performance across new clients by quantifying the unseen scanning code as one of the codes in the scanning codebook. Extensive experimental results demonstrate the superior performance of SCAN-PhysFed on public datasets.

eess.IV↗

arXiv2025

Autoregressive Medical Image Segmentation via Next-Scale Mask Prediction

While deep learning has significantly advanced medical image segmentation, most existing methods still struggle with handling complex anatomical regions. Cascaded or deep supervision-based approaches attempt to address this challenge through multi-scale feature learning but fail to establish sufficient inter-scale dependencies, as each scale relies solely on the features of the immediate predecessor. To this end, we propose the AutoRegressive Segmentation framework via next-scale mask prediction, termed AR-Seg, which progressively predicts the next-scale mask by explicitly modeling dependencies across all previous scales within a unified architecture. AR-Seg introduces three innovations: (1) a multi-scale mask autoencoder that quantizes the mask into multi-scale token maps to capture hierarchical anatomical structures, (2) a next-scale autoregressive mechanism that progressively predicts next-scale masks to enable sufficient inter-scale dependencies, and (3) a consensus-aggregation strategy that combines multiple sampled results to generate a more accurate mask, further improving segmentation robustness. Extensive experimental results on two benchmark datasets with different modalities demonstrate that AR-Seg outperforms state-of-the-art methods while explicitly visualizing the intermediate coarse-to-fine segmentation process.

eess.IV↗

arXiv2025

Emotional Face-to-Speech

How much can we infer about an emotional voice solely from an expressive face? This intriguing question holds great potential for applications such as virtual character dubbing and aiding individuals with expressive language disorders. Existing face-to-speech methods offer great promise in capturing identity characteristics but struggle to generate diverse vocal styles with emotional expression. In this paper, we explore a new task, termed emotional face-to-speech, aiming to synthesize emotional speech directly from expressive facial cues. To that end, we introduce DEmoFace, a novel generative framework that leverages a discrete diffusion transformer (DiT) with curriculum learning, built upon a multi-level neural audio codec. Specifically, we propose multimodal DiT blocks to dynamically align text and speech while tailoring vocal styles based on facial emotion and identity. To enhance training efficiency and generation quality, we further introduce a coarse-to-fine curriculum learning algorithm for multi-level token processing. In addition, we develop an enhanced predictor-free guidance to handle diverse conditioning scenarios, enabling multi-conditional generation and disentangling complex attributes effectively. Extensive experimental results demonstrate that DEmoFace generates more natural and consistent speech compared to baselines, even surpassing speech-driven methods. Demos are shown at https://demoface-ai.github.io/.

cs.SD↗

arXiv2025

Radiologist-in-the-Loop Self-Training for Generalizable CT Metal Artifact Reduction

Metal artifacts in computed tomography (CT) images can significantly degrade image quality and impede accurate diagnosis. Supervised metal artifact reduction (MAR) methods, trained using simulated datasets, often struggle to perform well on real clinical CT images due to a substantial domain gap. Although state-of-the-art semi-supervised methods use pseudo ground-truths generated by a prior network to mitigate this issue, their reliance on a fixed prior limits both the quality and quantity of these pseudo ground-truths, introducing confirmation bias and reducing clinical applicability. To address these limitations, we propose a novel Radiologist-In-the-loop SElf-training framework for MAR, termed RISE-MAR, which can integrate radiologists' feedback into the semi-supervised learning process, progressively improving the quality and quantity of pseudo ground-truths for enhanced generalization on real clinical CT images. For quality assurance, we introduce a clinical quality assessor model that emulates radiologist evaluations, effectively selecting high-quality pseudo ground-truths for semi-supervised training. For quantity assurance, our self-training framework iteratively generates additional high-quality pseudo ground-truths, expanding the clinical dataset and further improving model generalization. Extensive experimental results on multiple clinical datasets demonstrate the superior generalization performance of our RISE-MAR over state-of-the-art methods, advancing the development of MAR models for practical application. Code is available at https://github.com/Masaaki-75/rise-mar.

eess.IV↗

arXiv2025

Information-Maximized Soft Variable Discretization for Self-Supervised Image Representation Learning

Self-supervised learning (SSL) has emerged as a crucial technique in image processing, encoding, and understanding, especially for developing today's vision foundation models that utilize large-scale datasets without annotations to enhance various downstream tasks. This study introduces a novel SSL approach, Information-Maximized Soft Variable Discretization (IMSVD), for image representation learning. Specifically, IMSVD softly discretizes each variable in the latent space, enabling the estimation of their probability distributions over training batches and allowing the learning process to be directly guided by information measures. Motivated by the MultiView assumption, we propose an information-theoretic objective function to learn transform-invariant, non-travail, and redundancy-minimized representation features. We then derive a joint-cross entropy loss function for self-supervised image representation learning, which theoretically enjoys superiority over the existing methods in reducing feature redundancy. Notably, our non-contrastive IMSVD method statistically performs contrastive learning. Extensive experimental results demonstrate the effectiveness of IMSVD on various downstream tasks in terms of both accuracy and efficiency. Thanks to our variable discretization, the embedding features optimized by IMSVD offer unique explainability at the variable level. IMSVD has the potential to be adapted to other learning paradigms. Our code is publicly available at https://github.com/niuchuangnn/IMSVD.

cs.CV↗

arXiv2024

DreamVideo-2: Zero-Shot Subject-Driven Video Customization with Precise Motion Control

Recent advances in customized video generation have enabled users to create videos tailored to both specific subjects and motion trajectories. However, existing methods often require complicated test-time fine-tuning and struggle with balancing subject learning and motion control, limiting their real-world applications. In this paper, we present DreamVideo-2, a zero-shot video customization framework capable of generating videos with a specific subject and motion trajectory, guided by a single image and a bounding box sequence, respectively, and without the need for test-time fine-tuning. Specifically, we introduce reference attention, which leverages the model's inherent capabilities for subject learning, and devise a mask-guided motion module to achieve precise motion control by fully utilizing the robust motion signal of box masks derived from bounding boxes. While these two components achieve their intended functions, we empirically observe that motion control tends to dominate over subject learning. To address this, we propose two key designs: 1) the masked reference attention, which integrates a blended latent mask modeling scheme into reference attention to enhance subject representations at the desired positions, and 2) a reweighted diffusion loss, which differentiates the contributions of regions inside and outside the bounding boxes to ensure a balance between subject and motion control. Extensive experimental results on a newly curated dataset demonstrate that DreamVideo-2 outperforms state-of-the-art methods in both subject customization and motion control. The dataset, code, and models will be made publicly available.

cs.CV↗

arXiv2024

DepMamba: Progressive Fusion Mamba for Multimodal Depression Detection

Depression is a common mental disorder that affects millions of people worldwide. Although promising, current multimodal methods hinge on aligned or aggregated multimodal fusion, suffering two significant limitations: (i) inefficient long-range temporal modeling, and (ii) sub-optimal multimodal fusion between intermodal fusion and intramodal processing. In this paper, we propose an audio-visual progressive fusion Mamba for multimodal depression detection, termed DepMamba. DepMamba features two core designs: hierarchical contextual modeling and progressive multimodal fusion. On the one hand, hierarchical modeling introduces convolution neural networks and Mamba to extract the local-to-global features within long-range sequences. On the other hand, the progressive fusion first presents a multimodal collaborative State Space Model (SSM) extracting intermodal and intramodal information for each modality, and then utilizes a multimodal enhanced SSM for modality cohesion. Extensive experimental results on two large-scale depression datasets demonstrate the superior performance of our DepMamba over existing state-of-the-art methods. Code is available at https://github.com/Jiaxin-Ye/DepMamba.

cs.CY↗