In this paper, we propose a neural window decoder (NWD) for spatially coupled low-density parity-check (SC-LDPC) codes. The proposed NWD retains the conventional window decoder (WD) process but incorporates trainable neural weights. To train the weights of NWD, we introduce two novel training strategies. First, we restrict the loss function to target variable nodes (VNs) of the window, which prunes the neural network and accordingly enhances training efficiency. Second, we employ the active learning technique with a normalized loss term to prevent the training process from biasing toward specific training regions. Next, we develop a systematic method to derive non-uniform schedules for the NWD based on the training results. We introduce trainable damping factors that reflect the relative importance of check node (CN) updates. By skipping updates with less importance, we can omit $\mathbf{41\%}$ of CN updates without performance degradation compared to the conventional WD. Lastly, we address the error propagation problem inherent in SC-LDPC codes by deploying a complementary weight set, which is activated when an error is detected in the previous window. This adaptive decoding strategy effectively mitigates error propagation without requiring modifications to the code and decoder structures.
Clinical language models are important for many applications in healthcare, but their development depends on access to extensive clinical text for pretraining. However, obtaining clinical notes from electronic health records (EHRs) at scale is challenging due to patient privacy concerns. In this study, we rephrase existing clinical notes using LLMs to generate synthetic pretraining corpora, drawing inspiration from previous work on rephrasing web data. We examine four popular small-sized LLMs (<10B) to create synthetic clinical text to pretrain both decoder-based and encoder-based language models. The method yields better results in language modeling and downstream tasks than previous synthesis approaches without referencing real clinical text. We find that augmenting original clinical notes with synthetic corpora from different LLMs improves performances even at a small token budget, showing the potential of this method to support pretraining at the institutional level or be scaled to synthesize large-scale clinical corpora.
This paper presents a novel deep-learning framework that significantly enhances the transformation of rudimentary face sketches into high-fidelity colour images. Employing a Convolutional Block Attention-based Auto-encoder Network (CA2N), our approach effectively captures and enhances critical facial features through a block attention mechanism within an encoder-decoder architecture. Subsequently, the framework utilises a noise-induced conditional Generative Adversarial Network (cGAN) process that allows the system to maintain high performance even on domains unseen during the training. These enhancements lead to considerable improvements in image realism and fidelity, with our model achieving superior performance metrics that outperform the best method by FID margin of 17, 23, and 38 on CelebAMask-HQ, CUHK, and CUFSF datasets; respectively. The model sets a new state-of-the-art in sketch-to-image generation, can generalize across sketch types, and offers a robust solution for applications such as criminal identification in law enforcement.
Knowledge distillation (KD) is the de facto standard for compressing large-scale models into smaller ones. Prior works have explored ever more complex KD strategies involving different objective functions, teacher-ensembles, and weight inheritance. In this work we explore an alternative, yet simple approach -- active data curation as effective distillation for contrastive multimodal pretraining. Our simple online batch selection method, ACID, outperforms strong KD baselines across various model-, data- and compute-configurations. Further, we find such an active data curation strategy to in fact be complementary to standard KD, and can be effectively combined to train highly performant inference-efficient models. Our simple and scalable pretraining framework, ACED, achieves state-of-the-art results across 27 zero-shot classification and retrieval tasks with upto 11% less inference FLOPs. We further demonstrate that our ACED models yield strong vision-encoders for training generative multimodal models in the LiT-Decoder setting, outperforming larger vision encoders for image-captioning and visual question-answering tasks.
3D Gaussian Splatting (3DGS) has recently created impressive 3D assets for various applications. However, considering security, capacity, invisibility, and training efficiency, the copyright of 3DGS assets is not well protected as existing watermarking methods are unsuited for its rendering pipeline. In this paper, we propose GuardSplat, an innovative and efficient framework for watermarking 3DGS assets. Specifically, 1) We propose a CLIP-guided pipeline for optimizing the message decoder with minimal costs. The key objective is to achieve high-accuracy extraction by leveraging CLIP's aligning capability and rich representations, demonstrating exceptional capacity and efficiency. 2) We tailor a Spherical-Harmonic-aware (SH-aware) Message Embedding module for 3DGS, seamlessly embedding messages into the SH features of each 3D Gaussian while preserving the original 3D structure. This enables watermarking 3DGS assets with minimal fidelity trade-offs and prevents malicious users from removing the watermarks from the model files, meeting the demands for invisibility and security. 3) We present an Anti-distortion Message Extraction module to improve robustness against various distortions. Experiments demonstrate that GuardSplat outperforms state-of-the-art and achieves fast optimization speed. Project page is at https://narcissusex.github.io/GuardSplat, and Code is at https://github.com/NarcissusEx/GuardSplat.
Recently, human motion analysis has experienced great improvement due to inspiring generative models such as the denoising diffusion model and large language model. While the existing approaches mainly focus on generating motions with textual descriptions and overlook the reciprocal task. In this paper, we present~\textbf{MoTe}, a unified multi-modal model that could handle diverse tasks by learning the marginal, conditional, and joint distributions of motion and text simultaneously. MoTe enables us to handle the paired text-motion generation, motion captioning, and text-driven motion generation by simply modifying the input context. Specifically, MoTe is composed of three components: Motion Encoder-Decoder (MED), Text Encoder-Decoder (TED), and Moti-on-Text Diffusion Model (MTDM). In particular, MED and TED are trained for extracting latent embeddings, and subsequently reconstructing the motion sequences and textual descriptions from the extracted embeddings, respectively. MTDM, on the other hand, performs an iterative denoising process on the input context to handle diverse tasks. Experimental results on the benchmark datasets demonstrate the superior performance of our proposed method on text-to-motion generation and competitive performance on motion captioning.
In this paper, we introduce V2SFlow, a novel Video-to-Speech (V2S) framework designed to generate natural and intelligible speech directly from silent talking face videos. While recent V2S systems have shown promising results on constrained datasets with limited speakers and vocabularies, their performance often degrades on real-world, unconstrained datasets due to the inherent variability and complexity of speech signals. To address these challenges, we decompose the speech signal into manageable subspaces (content, pitch, and speaker information), each representing distinct speech attributes, and predict them directly from the visual input. To generate coherent and realistic speech from these predicted attributes, we employ a rectified flow matching decoder built on a Transformer architecture, which models efficient probabilistic pathways from random noise to the target speech distribution. Extensive experiments demonstrate that V2SFlow significantly outperforms state-of-the-art methods, even surpassing the naturalness of ground truth utterances. Code and models are available at: https://github.com/kaistmm/V2SFlow
Removing modeling constraints and unifying architectures across domains has been a key driver of the recent progress in training large multimodal models. However, most of these models still rely on many separately trained components such as modality-specific encoders and decoders. In this work, we further streamline joint generative modeling of images and text. We propose an autoregressive decoder-only transformer - JetFormer - which is trained to directly maximize the likelihood of raw data, without relying on any separately pretrained components, and can understand and generate both text and images. Specifically, we leverage a normalizing flow model to obtain a soft-token image representation that is jointly trained with an autoregressive multimodal transformer. The normalizing flow model serves as both an image encoder for perception tasks and an image decoder for image generation tasks during inference. JetFormer achieves text-to-image generation quality competitive with recent VQ-VAE- and VAE-based baselines. These baselines rely on pretrained image autoencoders, which are trained with a complex mixture of losses, including perceptual ones. At the same time, JetFormer demonstrates robust image understanding capabilities. To the best of our knowledge, JetFormer is the first model that is capable of generating high-fidelity images and producing strong log-likelihood bounds.
Sparse Multi-view Images can be Learned to predict explicit radiance fields via Generalizable Gaussian Splatting approaches, which can achieve wider application prospects in real-life when ground-truth camera parameters are not required as inputs. In this paper, a novel generalizable Gaussian Splatting method, SmileSplat, is proposed to reconstruct pixel-aligned Gaussian surfels for diverse scenarios only requiring unconstrained sparse multi-view images. First, Gaussian surfels are predicted based on the multi-head Gaussian regression decoder, which can are represented with less degree-of-freedom but have better multi-view consistency. Furthermore, the normal vectors of Gaussian surfel are enhanced based on high-quality of normal priors. Second, the Gaussians and camera parameters (both extrinsic and intrinsic) are optimized to obtain high-quality Gaussian radiance fields for novel view synthesis tasks based on the proposed Bundle-Adjusting Gaussian Splatting module. Extensive experiments on novel view rendering and depth map prediction tasks are conducted on public datasets, demonstrating that the proposed method achieves state-of-the-art performance in various 3D vision tasks. More information can be found on our project page (https://yanyan-li.github.io/project/gs/smilesplat)
The remarkable progress in deep learning (DL) showcases outstanding results in various computer vision tasks. However, adaptation to real-time variations in data distributions remains an important challenge. Test-Time Training (TTT) was proposed as an effective solution to this issue, which increases the generalization ability of trained models by adding an auxiliary task at train time and then using its loss at test time to adapt the model. Inspired by the recent achievements of contrastive representation learning in unsupervised tasks, we propose ReC-TTT, a test-time training technique that can adapt a DL model to new unseen domains by generating discriminative views of the input data. ReC-TTT uses cross-reconstruction as an auxiliary task between a frozen encoder and two trainable encoders, taking advantage of a single shared decoder. This enables, at test time, to adapt the encoders to extract features that will be correctly reconstructed by the decoder that, in this phase, is frozen on the source domain. Experimental results show that ReC-TTT achieves better results than other state-of-the-art techniques in most domain shift classification challenges.
Computational protein design (CPD) offers transformative potential for bioengineering, but current deep CPD models, focused on universal domains, struggle with function-specific designs. This work introduces a novel CPD paradigm tailored for functional design tasks, particularly for enzymes-a key protein class often lacking specific application efficiency. To address structural data scarcity, we present CrossDesign, a domain-adaptive framework that leverages pretrained protein language models (PPLMs). By aligning protein structures with sequences, CrossDesign transfers pretrained knowledge to structure models, overcoming the limitations of limited structural data. The framework combines autoregressive (AR) and non-autoregressive (NAR) states in its encoder-decoder architecture, applying it to enzyme datasets and pan-proteins. Experimental results highlight CrossDesign's superior performance and robustness, especially with out-of-domain enzymes. Additionally, the model excels in fitness prediction when tested on large-scale mutation data, showcasing its stability.
Current language models (LMs) use a fixed, static subword tokenizer. This default choice typically results in degraded efficiency and language capabilities, especially in languages other than English. To address this issue, we challenge the static design and propose retrofitting LMs with dynamic tokenization: a way to dynamically decide on token boundaries based on the input text via a subword-merging algorithm inspired by byte-pair encoding. We merge frequent subword sequences in a batch, then apply a pre-trained embedding-prediction hypernetwork to compute the token embeddings on-the-fly. For encoder-style models (e.g., XLM-R), this on average reduces token sequence lengths by >20% across 14 languages while degrading performance by less than 2%. The same method applied to pre-filling and scoring in decoder-style models (e.g., Mistral-7B; evaluated on English) results in minimal performance degradation at up to 6% reduction in sequence length. Overall, we find that dynamic tokenization can mitigate the limitations of static tokenization by substantially improving inference speed and promoting fairness across languages, enabling more equitable and adaptable LMs.
The Vision Transformer (ViT) has achieved notable success in computer vision, with its variants extensively validated across various downstream tasks, including semantic segmentation. However, designed as general-purpose visual encoders, ViT backbones often overlook the specific needs of task decoders, revealing opportunities to design decoders tailored to efficient semantic segmentation. This paper proposes Strip Cross-Attention (SCASeg), an innovative decoder head explicitly designed for semantic segmentation. Instead of relying on the simple conventional skip connections, we employ lateral connections between the encoder and decoder stages, using encoder features as Queries for the cross-attention modules. Additionally, we introduce a Cross-Layer Block that blends hierarchical feature maps from different encoder and decoder stages to create a unified representation for Keys and Values. To further boost computational efficiency, SCASeg compresses queries and keys into strip-like patterns to optimize memory usage and inference speed over the traditional vanilla cross-attention. Moreover, the Cross-Layer Block incorporates the local perceptual strengths of convolution, enabling SCASeg to capture both global and local context dependencies across multiple layers. This approach facilitates effective feature interaction at different scales, improving the overall performance. Experiments show that the adaptable decoder of SCASeg produces competitive performance across different setups, surpassing leading segmentation architectures on all benchmark datasets, including ADE20K, Cityscapes, COCO-Stuff 164k, and Pascal VOC2012, even under varying computational limitations.
This article describes the development of a novel U-Net-enhanced Wavelet Neural Operator (U-WNO),which combines wavelet decomposition, operator learning, and an encoder-decoder mechanism. This approach harnesses the superiority of the wavelets in time frequency localization of the functions, and the combine down-sampling and up-sampling operations to generate the segmentation map to enable accurate tracking of patterns in spatial domain and effective learning of the functional mappings to perform regional segmentation. By bridging the gap between theoretical advancements and practical applications, the U-WNO holds potential for significant impact in multiple science and industrial fields, facilitating more accurate decision-making and improved operational efficiencies. The operator is demonstrated for different pregnancy trimesters, utilizing two-dimensional ultrasound images.
Image-based relighting of indoor rooms creates an immersive virtual understanding of the space, which is useful for interior design, virtual staging, and real estate. Relighting indoor rooms from a single image is especially challenging due to complex illumination interactions between multiple lights and cluttered objects featuring a large variety in geometrical and material complexity. Recently, generative models have been successfully applied to image-based relighting conditioned on a target image or a latent code, albeit without detailed local lighting control. In this paper, we introduce ScribbleLight, a generative model that supports local fine-grained control of lighting effects through scribbles that describe changes in lighting. Our key technical novelty is an Albedo-conditioned Stable Image Diffusion model that preserves the intrinsic color and texture of the original image after relighting and an encoder-decoder-based ControlNet architecture that enables geometry-preserving lighting effects with normal map and scribble annotations. We demonstrate ScribbleLight's ability to create different lighting effects (e.g., turning lights on/off, adding highlights, cast shadows, or indirect lighting from unseen lights) from sparse scribble annotations.
The rapid progress of foundation models and large language models (LLMs) has fueled significantly improvement in the capabilities of machine learning systems that benefit from mutlimodal input data. However, existing multimodal models are predominantly built on top of pre-trained LLMs, which can limit accurate modeling of temporal dependencies across other modalities and thus limit the model's ability to jointly process and leverage multimodal inputs. To specifically investigate the alignment of text, video, and speech modalities in LLM-style (decoder-only) models, we consider a simplified multimodal generation task, Video-Text to Speech (VTTS): speech generation conditioned on both its corresponding text and video of talking people. The ultimate goal is to generate speech that not only follows the text but also aligns temporally with the video and is consistent with the facial expressions. In this paper, we first introduce Visatronic, a unified multimodal decoder-only transformer model that adopts an LLM-style architecture to embed visual, textual, and speech inputs into a shared subspace, treating all modalities as temporally aligned token streams. Next, we carefully explore different token mixing strategies to understand the best way to propagate information from the steps where video and text conditioning is input to the steps where the audio is generated. We extensively evaluate Visatronic on the challenging VoxCeleb2 dataset and demonstrate zero-shot generalization to LRS3, where Visatronic, trained on VoxCeleb2, achieves a 4.5% WER, outperforming prior SOTA methods trained only on LRS3, which report a 21.4% WER. Additionally, we propose a new objective metric, TimeSync, specifically designed to measure phoneme-level temporal alignment between generated and reference speech, further ensuring synchronization quality. Demo: https://apple.github.io/visatronic-demo/
Incorporating modern computer vision techniques into clinical protocols shows promise in improving skin lesion segmentation. The U-Net architecture has been a key model in this area, iteratively improved to address challenges arising from the heterogeneity of dermatologic images due to varying clinical settings, lighting, patient attributes, and hair density. To further improve skin lesion segmentation, we developed TAFM-Net, an innovative model leveraging self-adaptive transformer attention (TA) coupled with focal modulation (FM). Our model integrates an EfficientNetV2B1 encoder, which employs TA to enhance spatial and channel-related saliency, while a densely connected decoder integrates FM within skip connections, enhancing feature emphasis, segmentation performance, and interpretability crucial for medical image analysis. A novel dynamic loss function amalgamates region and boundary information, guiding effective model training. Our model achieves competitive performance, with Jaccard coefficients of 93.64\%, 86.88\% and 92.88\% in the ISIC2016, ISIC2017 and ISIC2018 datasets, respectively, demonstrating its potential in real-world scenarios.
Multiple-in-one image restoration (IR) has made significant progress, aiming to handle all types of single degraded image restoration with a single model. However, in real-world scenarios, images often suffer from combinations of multiple degradation factors. Existing multiple-in-one IR models encounter challenges related to degradation diversity and prompt singularity when addressing this issue. In this paper, we propose a novel multiple-in-one IR model that can effectively restore images with both single and mixed degradations. To address degradation diversity, we design a Local Dynamic Optimization (LDO) module which dynamically processes degraded areas of varying types and granularities. To tackle the prompt singularity issue, we develop an efficient Conditional Feature Embedding (CFE) module that guides the decoder in leveraging degradation-type-related features, significantly improving the model's performance in mixed degradation restoration scenarios. To validate the effectiveness of our model, we introduce a new dataset containing both single and mixed degradation elements. Experimental results demonstrate that our proposed model achieves state-of-the-art (SOTA) performance not only on mixed degradation tasks but also on classic single-task restoration benchmarks.
The emergence of the metaverse has boosted productivity and creativity, driving real-time updates and personalized content, which will substantially increase data traffic. However, current bit-oriented communication networks struggle to manage this high volume of dynamic information, restricting metaverse applications interactivity. To address this research gap, we propose a goal-oriented semantic communication (GSC) framework for metaverse. Building on an existing metaverse wireless construction task, our proposed GSC framework includes an hourglass network-based (HgNet) encoder to extract semantic information of objects in the metaverse; and a semantic decoder that uses this extracted information to reconstruct the metaverse content after wireless transmission, enabling efficient communication and real-time object behaviour updates to the scenery for metaverse construction task. To overcome the wireless channel noise at the receiver, we design an optimal transport (OT)-enabled semantic denoiser, which enhances the accuracy of metaverse scenery through wireless communication. Experimental results show that compared to the conventional metaverse construction, our proposed GSC framework significantly reduces wireless metaverse construction latency by 92.6\%, while improving metaverse object status accuracy and viewing experience by 45.6\% and 44.7\%, respectively.
Wide-FoV cameras, like fisheye and panoramic setups, are essential for broader perception but introduce significant distortions in 180{\deg} and 360{\deg} images, complicating dense prediction tasks. For instance, existing MAMBA models lacking distortion-aware capacity cannot perform well in panoramic semantic segmentation. To address this problem, this work presents Deformable Mamba, a unified framework specifically designed to address imaging distortions within the context of panoramic and fisheye semantic segmentation. At the core is a decoder constructed with a series of Deformable Mamba Fusion (DMF) blocks, making the whole framework more deformable, efficient, and accurate, when handling extreme distortions. Extensive evaluations across five datasets demonstrate that our method consistently improves segmentation accuracy compared to the previous state-of-the-art methods tailored for specific FoVs. Notably, Deformable Mamba achieves a +2.5% performance improvement on the 360{\deg} Stanford2D3D dataset, and shows better results across FoVs from 60{\deg} to 360{\deg}.