Token-based video representation has emerged as a promising approach for enabling large language models to interpret video content. However, existing token reduction techniques, such as token pruning and token merging, often disrupt essential spatial-temporal positional embeddings, failing to adequately balance computational efficiency with fewer tokens. Consequently, these methods result in relatively lengthy token sequences, limiting their applicability in scenarios requiring extreme token compression, such as video large language models. In this paper, we introduce the novel task of extreme short token reduction, aiming to represent extensive video sequences with a minimal number of tokens. To address this challenge, we propose Token Dynamics, a new video representation framework that dynamically reduces token count while preserving spatial-temporal coherence. Specifically, we disentangle video representations by separating visual embeddings from grid-level motion information, structuring them into: 1. a concise token base, created by clustering tokens that describe object-level content; 2. a token dynamics map, capturing detailed spatial-temporal motion patterns across grids. Furthermore, we introduce a cross-dynamics attention mechanism that integrates motion features into the token base without increasing token length, thereby maintaining compactness and spatial-temporal integrity. The experiments demonstrate a reduction of token count to merely 0.07% of the original tokens, with only a minor performance drop of 1.13%. Additionally, we propose two novel subtasks within extreme token reduction (fixed-length and adaptive-length compression), both effectively representing long token sequences for video-language tasks. Our method offers significantly lower theoretical complexity, fewer tokens, and enhanced throughput, thus providing an efficient solution for video LLMs.

Foundational models (FMs) have made significant strides in the healthcare domain. Yet the data silo challenge and privacy concern remain in healthcare systems, hindering safe medical data sharing and collaborative model development among institutions. The collection and curation of scalable clinical datasets increasingly become the bottleneck for training strong FMs. In this study, we propose Medical Foundation Models Merging (MedForge), a cooperative framework enabling a community-driven medical foundation model development, meanwhile preventing the information leakage of raw patient data and mitigating synchronization model development issues across clinical institutions. MedForge offers a bottom-up model construction mechanism by flexibly merging task-specific Low-Rank Adaptation (LoRA) modules, which can adapt to downstream tasks while retaining original model parameters. Through an asynchronous LoRA module integration scheme, the resulting composite model can progressively enhance its comprehensive performance on various clinical tasks. MedForge shows strong performance on multiple clinical datasets (e.g., breast cancer, lung cancer, and colon cancer) collected from different institutions. Our major findings highlight the value of collaborative foundation models in advancing multi-center clinical collaboration effectively and cohesively. Our code is publicly available at https://github.com/TanZheling/MedForge.

Consistency models are a new family of generative models capable of producing high-quality samples in either a single step or multiple steps. Recently, consistency models have demonstrated impressive performance, achieving results on par with diffusion models in the pixel space. However, the success of scaling consistency training to large-scale datasets, particularly for text-to-image and video generation tasks, is determined by performance in the latent space. In this work, we analyze the statistical differences between pixel and latent spaces, discovering that latent data often contains highly impulsive outliers, which significantly degrade the performance of iCT in the latent space. To address this, we replace Pseudo-Huber losses with Cauchy losses, effectively mitigating the impact of outliers. Additionally, we introduce a diffusion loss at early timesteps and employ optimal transport (OT) coupling to further enhance performance. Lastly, we introduce the adaptive scaling-$c$ scheduler to manage the robust training process and adopt Non-scaling LayerNorm in the architecture to better capture the statistics of the features and reduce outlier impact. With these strategies, we successfully train latent consistency models capable of high-quality sampling with one or two steps, significantly narrowing the performance gap between latent consistency and diffusion models. The implementation is released here: https://github.com/quandao10/sLCT/

Automated chest radiographs interpretation requires both accurate disease classification and detailed radiology report generation, presenting a significant challenge in the clinical workflow. Current approaches either focus on classification accuracy at the expense of interpretability or generate detailed but potentially unreliable reports through image captioning techniques. In this study, we present RadAlign, a novel framework that combines the predictive accuracy of vision-language models (VLMs) with the reasoning capabilities of large language models (LLMs). Inspired by the radiologist's workflow, RadAlign first employs a specialized VLM to align visual features with key medical concepts, achieving superior disease classification with an average AUC of 0.885 across multiple diseases. These recognized medical conditions, represented as text-based concepts in the aligned visual-language space, are then used to prompt LLM-based report generation. Enhanced by a retrieval-augmented generation mechanism that grounds outputs in similar historical cases, RadAlign delivers superior report quality with a GREEN score of 0.678, outperforming state-of-the-art methods' 0.634. Our framework maintains strong clinical interpretability while reducing hallucinations, advancing automated medical imaging and report analysis through integrated predictive and generative AI. Code is available at https://github.com/difeigu/RadAlign.

Cardiovascular disease (CVD) and cardiac dyssynchrony are major public health problems in the United States. Precise cardiac image segmentation is crucial for extracting quantitative measures that help categorize cardiac dyssynchrony. However, achieving high accuracy often depends on centralizing large datasets from different hospitals, which can be challenging due to privacy concerns. To solve this problem, Federated Learning (FL) is proposed to enable decentralized model training on such data without exchanging sensitive information. However, bandwidth limitations and data heterogeneity remain as significant challenges in conventional FL algorithms. In this paper, we propose a novel efficient and adaptive federate learning method for cardiac segmentation that improves model performance while reducing the bandwidth requirement. Our method leverages the low-rank adaptation (LoRA) to regularize model weight update and reduce communication overhead. We also propose a \mymethod{} aggregation technique to address data heterogeneity among clients. This technique adaptively penalizes the aggregated weights from different clients by comparing the validation accuracy in each client, allowing better generalization performance and fast local adaptation. In-client and cross-client evaluations on public cardiac MR datasets demonstrate the superiority of our method over other LoRA-based federate learning approaches.

We introduce a novel state-space architecture for diffusion models, effectively harnessing spatial and frequency information to enhance the inductive bias towards local features in input images for image generation tasks. While state-space networks, including Mamba, a revolutionary advancement in recurrent neural networks, typically scan input sequences from left to right, they face difficulties in designing effective scanning strategies, especially in the processing of image data. Our method demonstrates that integrating wavelet transformation into Mamba enhances the local structure awareness of visual inputs and better captures long-range relations of frequencies by disentangling them into wavelet subbands, representing both low- and high-frequency components. These wavelet-based outputs are then processed and seamlessly fused with the original Mamba outputs through a cross-attention fusion layer, combining both spatial and frequency information to optimize the order awareness of state-space models which is essential for the details and overall quality of image generation. Besides, we introduce a globally-shared transformer to supercharge the performance of Mamba, harnessing its exceptional power to capture global relationships. Through extensive experiments on standard benchmarks, our method demonstrates superior results compared to DiT and DIFFUSSM, achieving faster training convergence and delivering high-quality outputs. The codes and pretrained models are released at https://github.com/VinAIResearch/DiMSUM.git.

Despite the advances in learning-based image segmentation approach, the accurate segmentation of cardiac structures from magnetic resonance imaging (MRI) remains a critical challenge. While existing automatic segmentation methods have shown promise, they still require extensive manual corrections of the segmentation results by human experts, particularly in complex regions such as the basal and apical parts of the heart. Recent efforts have been made on developing interactive image segmentation methods that enable human-in-the-loop learning. However, they are semi-automatic and inefficient, due to their reliance on click-based prompts, especially for 3D cardiac MRI volumes. To address these limitations, we propose VerSe, a Versatile Segmentation framework to unify automatic and interactive segmentation through mutiple queries. Our key innovation lies in the joint learning of object and click queries as prompts for a shared segmentation backbone. VerSe supports both fully automatic segmentation, through object queries, and interactive mask refinement, by providing click queries when needed. With the proposed integrated prompting scheme, VerSe demonstrates significant improvement in performance and efficiency over existing methods, on both cardiac MRI and out-of-distribution medical imaging datasets. The code is available at https://github.com/bangwayne/Verse.

We have witnessed the unprecedented success of diffusion-based video generation over the past year. Recently proposed models from the community have wielded the power to generate cinematic and high-resolution videos with smooth motions from arbitrary input prompts. However, as a supertask of image generation, video generation models require more computation and are thus hosted mostly on cloud servers, limiting broader adoption among content creators. In this work, we propose a comprehensive acceleration framework to bring the power of the large-scale video diffusion model to the hands of edge users. From the network architecture scope, we initialize from a compact image backbone and search out the design and arrangement of temporal layers to maximize hardware efficiency. In addition, we propose a dedicated adversarial fine-tuning algorithm for our efficient model and reduce the denoising steps to 4. Our model, with only 0.6B parameters, can generate a 5-second video on an iPhone 16 PM within 5 seconds. Compared to server-side models that take minutes on powerful GPUs to generate a single video, we accelerate the generation by magnitudes while delivering on-par quality.

Discrete diffusion models have achieved success in tasks like image generation and masked language modeling but face limitations in controlled content editing. We introduce DICE (Discrete Inversion for Controllable Editing), the first approach to enable precise inversion for discrete diffusion models, including multinomial diffusion and masked generative models. By recording noise sequences and masking patterns during the reverse diffusion process, DICE enables accurate reconstruction and flexible editing of discrete data without the need for predefined masks or attention manipulation. We demonstrate the effectiveness of DICE across both image and text domains, evaluating it on models such as VQ-Diffusion, Paella, and RoBERTa. Our results show that DICE preserves high data fidelity while enhancing editing capabilities, offering new opportunities for fine-grained content manipulation in discrete spaces. For project webpage, see https://hexiaoxiao-cs.github.io/DICE/.

Looking up an unknown sign in an ASL dictionary can be difficult. Most ASL dictionaries are organized based on English glosses, despite the fact that (1) there is no convention for assigning English-based glosses to ASL signs; and (2) there is no 1-1 correspondence between ASL signs and English words. Furthermore, what if the user does not know either the meaning of the target sign or its possible English translation(s)? Some ASL dictionaries enable searching through specification of articulatory properties, such as handshapes, locations, movement properties, etc. However, this is a cumbersome process and does not always result in successful lookup. Here we describe a new system, publicly shared on the Web, to enable lookup of a video of an ASL sign (e.g., a webcam recording or a clip from a continuous signing video). The user submits a video for analysis and is presented with the five most likely sign matches, in decreasing order of likelihood, so that the user can confirm the selection and then be taken to our ASLLRP Sign Bank entry for that sign. Furthermore, this video lookup is also integrated into our newest version of SignStream(R) software to facilitate linguistic annotation of ASL video data, enabling the user to directly look up a sign in the video being annotated, and, upon confirmation of the match, to directly enter into the annotation the gloss and features of that sign, greatly increasing the efficiency and consistency of linguistic annotations of ASL video data.