Medical Foundation Models (MFMs) are revolutionizing radiography image analysis with scalable and generalized diagnostic capabilities. However, their effectiveness in real-world clinical practice is limited due to insufficient interpretability. To address this limitation, we propose RadLAS, a novel MFM for interpretable Radiographic image analysis by introducing Lesion-Aware Self-supervised pre-training. Unlike conventional MFMs that rely on post-hoc explanations, RadLAS innovates by directly emulating human diagnostic reasoning to first grounding lesion evidence and then making decisions accordingly. Specifically, RadLAS introduces two self-supervised tasks: (I) Lesion-grounded Reconstruction, which learns structured anatomical representations by restoring lesion-aware image patches into their healthy counterparts, thereby facilitating pixel-level grounding of lesion evidence via input-normal contrast. (II) Lesion-discrimination Contrastive Learning, which enhances lesion-aware pattern in representations by explicitly decoupling grounded lesion evidence as clinical cues and aligning them with global semantics, thereby enabling direct lesion-oriented diagnosis while preserving global context. RadLAS demonstrates excellent performance across diverse downstream radiographic datasets, offering verifiable explanations by deriving specific diagnoses (Task II) based on grounded lesion evidence (Task I), while preserving generalized representations essential for high diagnostic accuracy. Extensive experiments demonstrate that RadLAS (i) achieves superior interpretability with highly correlated lesion prediction and localization, surpassing 11 interpretable medical models; (ii) delivers scalable representation learning, outperforming 14 SOTA supervised and self-supervised MFMs.
论文检索
输入标题、作者或关键词，从 1,620 篇学术成果中精准定位
Brain-inspired Spiking Neural Networks (SNNs) have garnered significant attention due to their bio-plausibility and low power consumption advantages compared to Artificial Neural Networks (ANNs). However, the application of SNN in computer vision remains limited, primarily due to their inferior performance. In this work, we aim to bridge the performance gap between ANNs and SNNs in object detection by our Advanced SpikingYOLOX. The proposed approach extends the SpikingYOLOX with two key innovations: PSA-SNN and 2D-Spiking Transformer, both designed to enhance object detection performance. PSA-SNN extends spike-based self-attention by incorporating high-speed partial self-attention with an SNN-based 2D-Spiking Transformer in the deepest layer of the backbone, significantly improving feature extraction. The 2D-Spiking Transformer redefines the role of spiking neurons in Transformer sequences (Key, Query, Value), demonstrating that applying an additional spiking layer solely to the Value sequence yields the best performance while maintaining computational efficiency in spike-driven Transformers. We conduct extensive experiments on static images and the Advanced SpikingYOLOX achieves state-of-the-art performance among other SNN-based object detection methods. This work paves the way for more advanced SNN applications in object detection and broader computer vision tasks.
Masked modeling framework has shown promise in co-speech motion generation. However, it struggles to identify semantically significant frames for effective motion masking. In this work, we propose a speech-queried attention-based mask modeling framework for co-speech motion generation. Our key insight is to leverage motion-aligned speech features to guide the masked motion modeling process, selectively masking rhythm-related and semantically expressive motion frames. Specifically, we first propose a motion-audio alignment module (MAM) to construct a latent motion-audio joint space. In this space, both low-level and high-level speech features are projected, enabling motion-aligned speech representation using learnable speech queries. Then, a speech-queried attention mechanism (SQA) is introduced to compute frame-level attention scores through interactions between motion keys and speech queries, guiding selective masking toward motion frames with high attention scores. Finally, the motion-aligned speech features are also injected into the generation network to facilitate co-speech motion generation. Qualitative and quantitative evaluations confirm that our method outperforms existing state-of-the-art approaches, successfully producing high-quality co-speech motion. Project page: https://xiangyue-zhang.github.io/EchoMask
Real-world time series typically exhibit complex temporal variations, making the time series classification task notably challenging. Recent advancements have demonstrated the potential of multi-scale analysis approaches, which provide an effective solution for capturing these complex temporal patterns. However, existing multi-scale analysis-based time series prediction methods fail to eliminate redundant scale-shared features across multi-scale time series, resulting in the model over- or under-focusing on scale-shared features. To address this issue, we propose a novel end-to-end Disentangled Multi-Scale framework for Time Series classification (DisMS-TS). The core idea of DisMS-TS is to eliminate redundant shared features in multi-scale time series, thereby improving prediction performance. Specifically, we propose a temporal disentanglement module to capture scale-shared and scale-specific temporal representations, respectively. Subsequently, to effectively learn both scale-shared and scale-specific temporal representations, we introduce two regularization terms that ensure the consistency of scale-shared representations and the disparity of scale-specific representations across all temporal scales. Extensive experiments conducted on multiple datasets validate the superiority of DisMS-TS over its competitive baselines, with the accuracy improvement up to 9.71%.
The rapid growth of online video platforms, particularly live streaming services, has created an urgent need for real-time video understanding systems. These systems must process continuous video streams and respond to user queries instantaneously, presenting unique challenges for current Video Large Language Models (VideoLLMs). While existing VideoLLMs excel at processing complete videos, they face significant limitations in streaming scenarios due to their inability to handle dense, redundant frames efficiently. We introduce TimeChat-Online, a novel online VideoLLM that revolutionizes real-time video interaction. At its core lies our innovative Differential Token Drop (DTD) module, which addresses the fundamental challenge of visual redundancy in streaming videos. Drawing inspiration from human visual perception's Change Blindness phenomenon, DTD preserves meaningful temporal changes while filtering out static, redundant content between frames. Remarkably, our experiments demonstrate that DTD achieves an 82.8% reduction in video tokens while maintaining 98% performance on StreamingBench, revealing that over 80% of visual content in streaming videos is naturally redundant without requiring language guidance. To enable seamless real-time interaction, we present TimeChat-Online-139K, a comprehensive streaming video dataset featuring diverse interaction patterns including backward-tracing, current-perception, and future-responding scenarios. TimeChat-Online's unique Proactive Response capability, naturally achieved through continuous monitoring of video scene transitions via DTD, sets it apart from conventional approaches. Our extensive evaluation demonstrates TimeChat-Online's superior performance on streaming benchmarks (StreamingBench and OvOBench) and maintaining competitive results on long-form video tasks such as Video-MME and MLVU. Notably, when integrated with Qwen2.5VL-7B, DTD achieves a 5.7-point accuracy improvement on the challenging VideoMME subset containing videos of 30-60 minutes, while reducing video tokens by 84.6%. Project page: https://timechat-online.github.io.
In the rapidly evolving field of Artificial Intelligence Generated Content (AIGC), a central challenge is distinguishing AI-synthesized images from natural images. Despite the impressive capabilities of advanced AI generative models in producing visually compelling content, significant discrepancies remain when compared to natural images. To systematically investigate and quantify these differences, we construct a large-scale multimodal dataset named DANI, comprising 5,000 natural images and over 440,000 AI-generated image (AIGI) samples produced by nine representative models using both unimodal and multimodal prompts, including Text-to-Image (T2I), Text-and-Image-to-Image (I2I), and Text and Image-to-Image (TI2I). We then introduce D-Judge, a benchmark designed to answer the critical question: how far are AI-generated images from truly realistic images? Our fine-grained evaluation framework assesses DANI across five key dimensions: naive visual quality, semantic alignment, aesthetic appeal, downstream task applicability, and coordinated human validation. Extensive experiments reveal substantial discrepancies across these dimensions, highlighting the importance of aligning quantitative metrics with human judgment to achieve a comprehensive understanding of AI-generated image quality. The code and dataset are publicly available at: https://github.com/ryliu68/DJudge, and https://huggingface.co/datasets/Renyang/DANI.
Human beings perceive the real world through a spectrum of sensory modalities, encompassing auditory, visual, and linguistic faculties. This work proposes an industry-level omni-modal large language model (LLM) pipeline that integrates auditory, visual, and linguistic modalities to overcome challenges such as limited tri-modal datasets, high computational costs, and complex feature alignments. Our pipeline consists of three main components: First, a modular, end-to-end framework enabling flexible configuration of various encoder-LLM-decoder architectures. Second, a lightweight training strategy that pre-trains audio-language alignment on the state-of-the-art vision-language model Qwen2.5-VL, thus avoiding the costly pre-training of vision-specific modalities. Third, an audio synthesis pipeline that generates high-quality audio-text data from diverse real-world scenarios, supporting applications such as Automatic Speech Recognition and Speech-to-Speech chat. To this end, we introduce an industry-level omni-modal LLM, NEXUS-O. Extensive experiments validate the efficacy of our pipeline, yielding the following key findings: (1) In the visual understanding task, NEXUSO exhibits superior performance compared with its backbone model - Qwen2.5-VL-7B, validating the efficiency of our training strategy. (2) Within the English Spoken Question-Answering task, the model achieves better accuracy than the same-period competitor (i.e, MiniCPM-o2.6-7B) in the LLaMA Q. benchmark. (3) In our realworld ASR testset, NEXUS-O achieves outstanding performance, indicating its robustness in real scenarios. (4) In the Speech-to-Text Translation task, our model outperforms Qwen2-Audio-Instruct-7B. (5) In the Text-to-Speech task, based on pretrained vocoder (e.g., Fishspeech1.4 or CosyVoice2.0), NEXUS-O is comparable to its backbone vocoder on Seed-TTS benchmark. (6) An in-depth analysis of tri-modal alignment reveals that incorporating the audio modality enhances representational alignment between vision and language.
Automated vector floorplan generation is valuable for designers to explore potential spatial designs. However, existing learning-based methods rely on complex post-processing or optimization to obtain plausible vector floorplans, which disrupts the end-to-end design flow. In this paper, we propose FloorplanSBS, a patch-based segmentation framework for directly synthesizing vector floorplans. Our method leverages the strengths of box-based representation and segmentation-based generation, following a division-and-labeling scheme. The framework operates in two stages: given input design constraints, a division model first divides the design space into rectangular patches, followed by a labelling model that assigns semantic labels to each patch. FloorplanSBS supports constraints such as boundaries and layout graphs. Extensive evaluations show that it surpasses state-of-the-art methods in generating high-quality vector floorplans. With its end-to-end neural framework, FloorplanSBS eliminates the need for post-processing, offering a simple, efficient, and user-friendly tool for vector floorplan design.
Diffusion models have shown remarkable promise for image restoration by leveraging powerful priors. Prominent methods typically frame the restoration problem within a Bayesian inference framework, which iteratively combines a denoising step with a likelihood guidance step. However, the interactions between these two components in the generation process remain underexplored. In this paper, we analyze the underlying gradient dynamics of these components and identify significant instabilities. Specifically, we demonstrate conflicts between the prior and likelihood gradient directions, alongside temporal fluctuations in the likelihood gradient itself. We show that these instabilities disrupt the generative process and compromise restoration performance. To address these issues, we propose Stabilized Progressive Gradient Diffusion (SPGD), a novel gradient management technique. SPGD integrates two synergistic components: (1) a progressive likelihood warm-up strategy to mitigate gradient conflicts; and (2) adaptive directional momentum (ADM) smoothing to reduce fluctuations in the likelihood gradient. Extensive experiments across diverse restoration tasks demonstrate that SPGD significantly enhances generation stability, leading to state-of-the-art performance in quantitative metrics and visually superior results. Code is available at https://github.com/74587887/SPGD.
In this paper, we address the task of multimodal-to-speech generation, which aims to synthesize high-quality speech from multiple input modalities: text, video, and reference audio. This task has gained increasing attention due to its wide range of applications, such as film production, dubbing, and virtual avatars. Despite recent progress, existing methods still suffer from limitations in speech intelligibility, audio-video synchronization, speech naturalness, and voice similarity to the reference speaker. To address these challenges, we propose AlignDiT, a multimodal Aligned Diffusion Transformer that generates accurate, synchronized, and natural-sounding speech from aligned multimodal inputs. Built upon the in-context learning capability of the DiT architecture, AlignDiT explores three effective strategies to align multimodal representations. Furthermore, we introduce a novel multimodal classifier-free guidance mechanism that allows the model to adaptively balance information from each modality during speech synthesis. Extensive experiments demonstrate that AlignDiT significantly outperforms existing methods across multiple benchmarks in terms of quality, synchronization, and speaker similarity. Moreover, AlignDiT exhibits strong generalization capability across various multimodal tasks, such as video-to-speech synthesis and visual forced alignment, consistently achieving state-of-the-art performance. The demo page is available at https://mm.kaist.ac.kr/projects/AlignDiT.
Human speech goes beyond the mere transfer of information; it is a profound exchange of emotions and a connection between individuals. While Text-to-Speech (TTS) models have made huge progress, they still face challenges in controlling the emotional expression in the generated speech. In this work, we propose EmoVoice, a novel emotion-controllable TTS model that exploits large language models (LLMs) to enable fine-grained freestyle natural language emotion control, and a phoneme boost variant design that makes the model output phoneme tokens and audio tokens in parallel to enhance content consistency, inspired by chain-of-thought (CoT) and modality-of-thought (CoM) techniques. Besides, we introduce EmoVoice-DB, a high-quality 40-hour English emotion dataset featuring expressive speech and fine-grained emotion labels with natural language descriptions. EmoVoice achieves state-of-the-art performance on the English EmoVoice-DB test set using only synthetic training data, and on the Chinese Secap test set using our in-house data. We further investigate the reliability of existing emotion evaluation metrics and their alignment with human perceptual preferences, and explore using SOTA multimodal LLMs GPT-4o-audio and Gemini to assess emotional speech. Dataset, code, checkpoints and demo samples are available at https://github.com/yanghaha0908/EmoVoice.
Generating highly realistic 4D interaction in real time is significant for visual content generation. Although existing works have validated to produce impressive dynamics by employing physical simulation and learned material mainly from pre-trained video diffusion models, it is still challenging to generate real-time 4D interaction with high-quality motion due to the heavy time consumption of the simulation solver and indirect material learning strategy. This paper proposes a novel physics-based 4D generation method, Phys4DRT, for arbitrary realistic real-time interaction on 3D Gaussian Splatting (3DGS) objects with direct motion supervision in time-frequency domain. Specifically, we devise a fast and differentiable eXtended Position Based Dynamics (XPBD) simulator as the light-weight controller for efficient physical evolution on a quasi-regular tetrahedral proxy mesh, into which we immerse the static 3DGS for efficient and stable deformation simulation. In addition, to learn the heterogeneous material for realistic motion, we directly supervise the generated dynamic 3D behavior by the motion representation of the optical flow and spectral volume extracted from the generated reference video, rather than indirect supervision in the color space used in previous approaches. We thoroughly conduct experiments on the public benchmarks to demonstrate the efficiency and effectiveness of our method. Our model can accelerate real-time 4D interaction generation by approximately x20 faster than the current Material Point Method (MPM) based approaches while achieving competitive visual quality compared with the state-of-the-art baselines.
In recent years, Text-to-Music (T2M) generation models have rapidly emerged as powerful tools in content creation across fields. While existing models have made notable progress in sound quality, instrument identification, and stylistic alignment, they still exhibit clear limitations in modeling musical structure and musicality-particularly in terms of harmonic coherence and rhythmic alignment. To address these issues, we propose a Temporal-Conditioned Symbolic Alignment for Controllable Text-to-Music Generation(TCSA), which introduces explicit local condition controls to enhance structural fidelity in music generation. Specifically, we design a music theory enrichment strategy based on GPT-2 that transforms input text into detailed descriptions with embedded music theory knowledge, from which accurate chord progressions and rhythmic patterns are extracted as generation conditions. To synchronize these local features effectively, we develop a temporal alignment feature fusion mechanism. Additionally, we propose a layer-skipping fine-tuning strategy to avoid overfitting and enable fine-grained structural modeling. Finally, we introduce a perception-driven loss function based on Mel spectrograms to optimize the harmonic consistency and structural coherence of the generated music. Experimental results demonstrate that TCSA achieves competitive generation quality while offering significantly improved controllability over musical structure, making it well-suited for professional music production and refined content creation.
Emerging immersive display technologies efficiently utilize resources with perceptual graphics methods such as foveated rendering and denoising. Running multiple perceptual graphics methods challenges devices with limited power and computational resources. We propose a computationally-lightweight learned multitasking perceptual graphics model. Given RGB images and text-prompts, our model performs text-described perceptual tasks in a single inference step. Simply daisy-chaining multiple models or training dedicated models can lead to model management issues and exhaust computational resources. In contrast, our flexible method unlocks consistent high quality perceptual effects with reasonable compute, supporting various permutations at varied intensities using adjectives in text prompts (e.g., ''mildly'', ''lightly''). Text-guidance provides ease of use for dynamic requirements such as creative processes. To train our model, we propose a dataset containing source and perceptually enhanced images with corresponding text prompts. We evaluate our model on desktop and embedded platforms and validate perceptual quality through a user study.
Large Vision-Language Models (LVLMs) have achieved impressive progress in multi-modal understanding and generation. However, they still tend to produce hallucinated content that is inconsistent with the visual input, which limits their reliability in real-world applications. We propose CoFi-Dec, a training-free decoding framework that mitigates hallucinations by integrating generative self-feedback with coarse-to-fine visual conditioning. Inspired by the human visual process from global scene perception to detailed inspection, CoFi-Dec first generates two intermediate textual responses conditioned on coarse- and fine-grained views of the original image. These responses are then transformed into synthetic images using a text-to-image model, forming multi-level visual hypotheses that enrich grounding cues. To unify the predictions from these multiple visual conditions, we introduce a Wasserstein-based fusion mechanism that aligns their predictive distributions into a geometrically consistent decoding trajectory. This principled fusion reconciles high-level semantic consistency with fine-grained visual grounding, leading to more robust and faithful outputs. Extensive experiments on six hallucination-focused benchmarks show that CoFi-Dec substantially reduces both entity-level and semantic-level hallucinations, outperforming existing decoding strategies. The framework is model-agnostic, requires no additional training, and can be seamlessly applied to a wide range of LVLMs.
Emotions are fundamental to the creation and perception of music performances. However, achieving human-like expression and emotion through machine learning models for performance rendering remains a challenging task. In this work, we present SyMuPe, a novel framework for developing and training affective and controllable symbolic piano performance models. Our flagship model, PianoFlow, uses conditional flow matching trained to solve diverse multi-mask performance inpainting tasks. By design, it supports both unconditional generation and infilling of music performance features. For training, we use a curated, cleaned dataset of 2,968 hours of aligned musical scores and expressive MIDI performances. For text and emotion control, we integrate a piano performance emotion classifier and tune PianoFlow with the emotion-weighted Flan-T5 text embeddings provided as conditional inputs. Objective and subjective evaluations against transformer-based baselines and existing models show that PianoFlow not only outperforms other approaches, but also achieves performance quality comparable to that of human-recorded and transcribed MIDI samples. For emotion control, we present and analyze samples generated under different text conditioning scenarios. The developed model can be integrated into interactive applications, contributing to the creation of more accessible and engaging music performance systems.
Identity-Preserving Facial Aesthetic Enhancement via Hierarchical Prompt Learning and Pivotal Tuning
The demand for identity-preserving Facial Aesthetic Enhancement (FAE) has surged in social media and digital entertainment. However, existing methods based on deep generative models encounter difficulties in striking a balance between fine-grained detail enhancement and preserving the unique identities of individuals from diverse ethnic and gender backgrounds. To tackle this issue, this paper proposes a novel tuning-based framework that integrates prototype-based hierarchical prompt learning within a CLIP model and a StyleGAN-based inversion model. Our approach first adapts a pre-trained StyleGAN to the input face via pivotal tuning, optimizing around pivotal latent codes to minimize reconstruction distortion while retaining editability. Then, a prototype-based hierarchical prompt learning module is designed for learning multigrained facial features to achieve comprehensive and fine-grained facial descriptions for FAE. Specifically, we propose a prototypical similarity measure based on a multi-ethnic dataset to select geometrically similar faces with high aesthetic scores as reference faces. This selection is guided by ArcFace regularization within categorized gender and ethnic groups to minimize identity loss. Additionally, we design a novel aesthetic attribute selection algorithm to generate generic fine-grained aesthetic attributes from these reference faces for detailed facial descriptions. These components work synergistically through dynamic weight modulation, prioritizing features with high aesthetic contributions (such as enhancing lip fullness) while ensuring semantic consistency through CLIP-driven optimization for pivotal latent codes. Extensive experiments demonstrate that our method outperforms state-of-the-art techniques in both aesthetic quality and identity preservation, especially for out-of-domain faces.
Face cartoonization remains a challenging task due to significant geometric deformations between facial photos and cartoons, as well as the absence of paired training data for supervised learning. Existing methods struggle to generate high-quality cartoonized avatars with attribute consistency. To address this challenge, this paper proposes an unsupervised facial cartoonization method based on cross-domain aligned and deformable vector quantization (CADQ). Firstly, we construct textual descriptions with facial attributes for both photo datasets and cartoon collections. Attribute consistency during transformation is enforced through individually contrastive learning between image-text cross-modal features and globally distribution alignment across photo-cartoon domains. Secondly, a deformable Transformer with dual attention is introduced during the transformation process, which queries corresponding cartoon codebook entries based on image features to simulate cross-domain geometric deformations. Experimental results demonstrate that the proposed method can convert facial photos into high-quality cartoons with attribute consistency, outperforming existing state-of-the-art approaches. Furthermore, the method can be effectively extended to unsupervised cross-domain generation of other artistic portrait styles, achieving superior or highly competitive performance. Our code has been released at: https://github.com/IIP-Lab-XDU/CADQ.
While recent video-to-audio (V2A) models can generate realistic background audio from visual input, they largely overlook speech, an essential part of many video soundtracks. This paper proposes a new task, video-to-soundtrack (V2ST) generation, which aims to jointly produce synchronized background audio and speech within a unified framework. To tackle V2ST, we introduce DualDub, a unified framework built on a multimodal language model that integrates a multimodal encoder, a cross-modal aligner, and dual decoding heads for simultaneous background audio and speech generation. Specifically, our proposed cross-modal aligner employs causal and non-causal attention mechanisms to improve synchronization and acoustic harmony. Besides, to handle data scarcity, we design a curriculum learning strategy that progressively builds the multimodal capability. Finally, we introduce DualBench, the first benchmark for V2ST evaluation with a carefully curated test set and comprehensive metrics. Experimental results demonstrate that DualDub achieves state-of-the-art performance, generating high-quality and well-synchronized soundtracks with both speech and background audio. DualBench and generated samples of DualDub are available at https://github.com/wjtian-wonderful/DualBench.
High-fidelity hand gesture generation represents a significant challenge in human-centric generation tasks. Existing methods typically employ a single-view mesh-rendered image prior to enhancing gesture generation quality. However, the spatial complexity of hand gestures and the inherent limitations of single-view rendering make it difficult to capture complete gesture information, particularly when fingers are occluded. The fundamental contradiction lies in the loss of 3D topological relationships through 2D projection and the incomplete spatial coverage inherent to single-view representations. Diverging from single-view prior approaches, we propose a multi-view prior framework, named Multi-Modal UNet-based Feature Encoder (MUFEN), to guide diffusion models in learning comprehensive 3D hand information. Specifically, we extend conventional front-view rendering to include rear, left, right, top, and bottom perspectives, selecting the most information-rich view combination as training priors to address occlusion. This multi-view prior with a dedicated dual stream encoder significantly improves the model's understanding of complete hand features. Furthermore, we design a bounding box feature fusion module, which can fuse the gesture localization features and multi-modal features to enhance the location-awareness of the MUFEN features to the gesture-related features. Experiments demonstrate that our method achieves state-of-the-art performance on quantitative metrics and exhibits superior qualitative results. The source code is available at https://github.com/fuqifan/MUFEN.