论文检索

General Machine Learning · Probabilistic Methods

Tommy Rochussen, Vincent Fortuin

One of the core facets of Bayesianism is in the updating of prior beliefs in light of new evidence$\textemdash$so how can we maintain a Bayesian approach if we have no prior beliefs in the first place? This is one of the central challenges in the field of Bayesian deep learning, where it is not clear how to represent beliefs about a prediction task by prior distributions over model parameters. Bridging the fields of Bayesian deep learning and probabilistic meta-learning, we introduce a way to $\textit{learn}$ a weights prior from a collection of datasets by introducing a way to perform per-dataset amortised variational inference. The model we develop can be viewed as a neural process whose latent variable is the set of weights of a BNN and whose decoder is the neural network parameterised by a sample of the latent variable itself. This unique model allows us to study the behaviour of Bayesian neural networks under well-specified priors, use Bayesian neural networks as flexible generative models, and perform desirable but previously elusive feats in neural processes such as within-task minibatching or meta-learning under extreme data-starvation.

Zongrui Li, Xinhua Ma, Minghui HU, Yunqing Zhao, Yingchen Yu, Qian Zheng, Chang Liu, Jiang Xudong, Song Bai

Monocular normal estimation aims to estimate the normal map from a single RGB image of an object under arbitrary lights. Existing methods rely on deep models to directly predict normal maps. However, they often suffer from 3D misalignment: while the estimated normal maps may appear to have a correct appearance, the reconstructed surfaces often fail to align with the 3D geometry. We argue that this misalignment stems from the current paradigm: the model struggles to distinguish and estimate varying geometry represented in normal maps, as the differences in underlying geometry are reflected only through relatively subtle color variations. To address this issue, we propose a new paradigm that reformulates normal estimation as shading sequence estimation, where shading sequences are more sensitive to various geometry information. By learning to infer the shading sequence of an object, the model can better capture underlying 3D geometry and thereby produce more accurate normal predictions. Building on this paradigm, we present RoSE, a method that leverages image-to-video generative models to predict shading sequences, which are then converted into normal maps by solving a simple ordinary least-squares problem. To enhance robustness and better handle complex objects, RoSE is trained on a synthetic dataset, MultiShade, with diverse shapes, materials, and light conditions. Experiments demonstrate that RoSE achieves state-of-the-art performance on both synthetic and real-world benchmark datasets for object-based monocular normal estimation.

Computer Vision · Image and Video Generation

Zongrui Li, Xinhua Ma, Minghui HU, Yunqing Zhao, Yingchen Yu, Qian Zheng, Chang Liu, Jiang Xudong, Song Bai

Monocular normal estimation aims to estimate the normal map from a single RGB image of an object under arbitrary lights. Existing methods rely on deep models to directly predict normal maps. However, they often suffer from 3D misalignment: while the estimated normal maps may appear to have a correct appearance, the reconstructed surfaces often fail to align with the 3D geometry. We argue that this misalignment stems from the current paradigm: the model struggles to distinguish and estimate varying geometry represented in normal maps, as the differences in underlying geometry are reflected only through relatively subtle color variations. To address this issue, we propose a new paradigm that reformulates normal estimation as shading sequence estimation, where shading sequences are more sensitive to various geometry information. By learning to infer the shading sequence of an object, the model can better capture underlying 3D geometry and thereby produce more accurate normal predictions. Building on this paradigm, we present RoSE, a method that leverages image-to-video generative models to predict shading sequences, which are then converted into normal maps by solving a simple ordinary least-squares problem. To enhance robustness and better handle complex objects, RoSE is trained on a synthetic dataset, MultiShade, with diverse shapes, materials, and light conditions. Experiments demonstrate that RoSE achieves state-of-the-art performance on both synthetic and real-world benchmark datasets for object-based monocular normal estimation.

Deep Learning · Attention Mechanisms

Yongqi An, Chang Lu, Kuan Zhu, Tao Yu, Chaoyang Zhao, Hong Wu, Ming Tang, Jinqiao Wang

Large language models (LLMs) face growing challenges in efficient generative inference due to the increasing memory demands of Key-Value (KV) caches, especially for long sequences. Existing eviction methods typically retain KV pairs with high attention weights but overlook the impact of attention redistribution caused by token removal, as well as the spatial-temporal dynamics in KV selection. In this paper, we propose ReST-KV, a robust KV eviction method that combines layer-wise output **Re**construction and **S**patial-**T**emporal smoothing to provide a more comprehensive perspective for the KV cache eviction task. Specifically, ReST-KV formulates KV cache eviction as an optimization problem that minimizes output discrepancies through efficient layer-wise reconstruction. By directly modeling how each token’s removal affects the model output, our method naturally captures attention redistribution effects, going beyond simplistic reliance on raw attention weights. To further enhance robustness, we design exponential moving average smoothing to handle temporal variations and an adaptive window-based mechanism to capture spatial patterns. Our method, ReST-KV, significantly advances performance on long-context benchmarks. It surpasses state-of-the-art baselines by 2.58\% on LongBench and 15.2\% on RULER. Additionally, ReST-KV consistently outperforms existing methods on Needle-in-a-Haystack and InfiniteBench, all while achieving a remarkable 10.61$\times$ reduction in decoding latency at 128k context length. The code is included in the supplementary material and is designed for easy reproduction.

Deep Learning · Generative Models and Autoencoders

François Bertholom, Khalid Oublal

Denoising Diffusion Probabilistic Models (DDPMs) and Hierarchical Variational Autoencoders (HVAEs) are typically studied as distinct paradigms for high-dimensional generative modeling. In this work, we bridge this gap by establishing a formal equivalence between DDPMs and HVAEs in the limit of infinite depth with a fixed, Markovian inference process. We argue that this architectural isomorphism is not merely a mathematical curiosity but the structural key to understanding the superior generalization capabilities of diffusion models. By viewing the forward diffusion process as a fixed encoder, we elucidate how DDPMs circumvent the posterior collapse often observed in deep VAEs, effectively balancing the trade-off between structural guidance and texture synthesis. We support this theoretical unification with empirical analysis of the semantic phase transitions in latent space and demonstrate the invariance of the Variational Lower Bound under noise schedule reparameterizations, confirming the interpretation of diffusion as a continuous-time hierarchical variational framework.

Deep Learning · Algorithms

Patrick Blumenberg, Thomas Graave, Tim Fingscheidt

Large language models (LLMs) demand extensive memory capacity during both fine-tuning and inference. To enable memory-efficient fine-tuning, existing methods apply block-wise quantization techniques, such as NF4 and AF4, to the network weights. We show that these quantization techniques incur suboptimal quantization errors. Therefore, as a first novelty, we propose an optimization approach for block-wise quantization. Using this method, we design a family of quantizers named 4-bit block-wise optimal float (BOF4), which consistently reduces the quantization error compared to both baseline methods. We provide both a theoretical and a data-driven solution for the optimization process and prove their practical equivalence. Secondly, we propose a modification to the employed normalization method based on the signed absolute block maximum (BOF4-S), enabling further reduction of the quantization error and empirically achieving less degradation in language modeling performance. Thirdly, we explore additional variations of block-wise quantization methods applied to LLMs through an experimental study on the importance of accurately representing zero and large-amplitude weights on the one hand, and optimization towards various error metrics on the other hand. Lastly, we introduce a mixed-precision quantization strategy dubbed outlier-preserving quantization (OPQ) to address the distributional mismatch induced by outlier weights in block-wise quantization. By storing outlier weights in 16-bit precision (OPQ) while applying BOF4-S, we achieve top performance among 4-bit block-wise quantization techniques w.r.t. perplexity.

General Machine Learning · Probabilistic Methods

Kaiyu Li, Yiming Yang, Xiaoyuan Cheng, Yi He, Zhuo Sun

Control variates are variance reduction techniques for Monte Carlo estimators. They play a critical role in improving Monte Carlo estimators in scientific and machine learning applications that involve computationally expensive integrals. We introduce \emph{multilevel control functionals} (MLCFs), a novel and widely applicable extension of control variates that combines non-parametric Stein-based control variates with multi-fidelity methods. We show that when the integrand and the density are smooth, and when the dimensionality is not very high, MLCFs enjoy a faster convergence rate. We provide both theoretical analysis and empirical assessments on differential equation examples, including Bayesian inference for ecological models, to demonstrate the effectiveness of our proposed approach. Furthermore, we extend MLCFs for variational inference, and demonstrate improved performance empirically through Bayesian neural network examples.

Vladimir Kondratyev, Alexander Fishkov, Mahmoud Hegazy, Nikita Kotelevskii, Rémi Flamary, Maxim Panov, Eric Moulines

We propose a framework for conditional vector quantile regression (CVQR) that combines neural optimal transport with amortized optimization, and apply it to multivariate conformal prediction. Classical quantile regression does not extend naturally to multivariate responses, while existing approaches often ignore the geometry of joint distributions. Our method parameterizes the conditional vector quantile function as the gradient of a convex potential implemented by an input-convex neural network, ensuring monotonicity and uniform ranks. To reduce the cost of solving high-dimensional variational problems, we introduce amortized optimization of the dual potentials, yielding efficient training and faster inference. We then exploit the induced multivariate ranks for conformal prediction, constructing distribution-free predictive regions with finite-sample validity. Unlike coordinatewise methods, our approach adapts to the geometry of the conditional distribution, producing tighter and more informative regions. Experiments on benchmark datasets show improved coverage–efficiency trade-offs compared to baselines, highlighting the benefits of integrating neural optimal transport with conformal prediction.

Computer Vision · Image and Video Generation

Jianhao Yuan, Fabio Pizzati, Francesco Pinto, Lars Kunze, Ivan Laptev, Paul Newman, Philip Torr, Daniele De Martini

Intuitive physics understanding in video diffusion models plays an essential role in building general-purpose physically plausible world simulators, yet accurately evaluating such capacity remains a challenging task due to the difficulty in disentangling physics correctness from visual appearance in generation. To the end, we introduce LikePhys, a training-free method that evaluates intuitive physics in video diffusion models by distinguishing physically valid and impossible videos using the denoising objective as an ELBO-based likelihood surrogate on a curated dataset of valid-invalid pairs. By testing on our constructed benchmark of twelve scenarios spanning over four physics domains, we show that our evaluation metric, Plausibility Preference Error (PPE), demonstrates strong alignment with human preference, outperforming state-of-the-art evaluator baselines. We then systematically benchmark intuitive physics understanding in current video diffusion models. Our study further analyses how model design and inference settings affect intuitive physics understanding and highlights domain-specific capacity variations across physical laws. Empirical results show that, despite current models struggling with complex and chaotic dynamics, there is a clear trend of improvement in physics understanding as model capacity and inference settings scale up.

Zhen Han, Mattias Teye, Derek Yadgaroff, Judith Bütepage

The training of high-quality, robust machine learning models for speech-driven 3D facial animation requires a large, diverse dataset of high-quality audio-animation pairs. To overcome the lack of such a dataset, recent work has introduced large pre-trained speech encoders that are robust to variations in the input audio and, therefore, enable the facial animation model to generalize across speakers, audio quality, and languages. However, the resulting facial animation models are prohibitively large and lend themselves only to offline inference on a dedicated machine. In this work, we explore on-device, real-time facial animation models in the context of game development. We overcome the lack of large datasets by using hybrid knowledge distillation with pseudo-labeling. Given a large audio dataset, we employ a high-performing teacher model to train very small student models. In contrast to the pre-trained speech encoders, our student models only consist of convolutional and fully-connected layers, removing the need for attention context or recurrent updates. In our experiments, we demonstrate that we can reduce the memory footprint to up to 3.4 MB and required future audio context to up to 81 ms while maintaining high-quality animations. This paves the way for on-device inference, an important step towards realistic, model-driven digital characters.

Roi Benita, Michael Finkelson, Tavi Halperin, Gleb Sterkin, Yossi Adi

Foley is a key element in video production, refers to the process of adding an audio signal to a silent video while ensuring semantic and temporal alignment. In recent years, the rise of personalized content creation and advancements in automatic video-to-audio models have increased the demand for greater user control in the process. One possible approach is to incorporate text to guide audio generation. While supported by existing methods, challenges remain in ensuring compatibility between modalities, particularly when the text introduces additional information or contradicts the sounds naturally inferred from the visuals. In this work, we introduce CAFA (Controllable Automatic Foley Artist) a video-and-text-to-audio model that generates semantically and temporally aligned audio for a given video, guided by text input. CAFA is built upon a text-to-audio model and integrates video information through a modality adapter mechanism. By incorporating text, users can refine semantic details and introduce creative variations, guiding the audio synthesis beyond the expected video contextual cues. Experiments show that besides its superior quality in terms of semantic alignment and audio-visual synchronization the proposed method enable high textual controllability as demonstrated in subjective and objective evaluations.

Junfu Tan, Peiguang Jing, Yu Zhu, Yu Liu

Open-set fine-grained recognition (OSFGR) is the core exploration of building open-world intelligent systems. The challenge lies in the gradual semantic drift during the transition from coarse-grained to fine-grained categories. However, although existing methods leverage hierarchical representations to assist progressive reasoning, they neglect semantic consistency across hierarchies. To address this, we propose a multimodal progressive bidirectional reasoning framework: (1) In forward reasoning, the model progressively refines visual features to capture hierarchical structural representations, while (2) in backward reasoning, variational inference integrates multimodal information to constrain consistency in category-aware latent spaces. This mechanism mitigates semantic drift through bidirectional information flow and cross-hierarchical feature consistency constraints. Extensive experiments on the iNat2021-OSR dataset demonstrate that our proposed method achieves superior performance.

Ivan Sabolić, Matej Grcić, Siniša Šegvić

We propose VIBE, a model-agnostic framework that trains classifiers resilient to backdoor attacks. The key concept behind our approach is to treat malicious inputs and corrupted labels from the training dataset as observed random variables, while the actual clean labels are latent. VIBE then recovers the corresponding latent clean label posterior through variational inference. The resulting training procedure follows the expectation-maximization (EM) algorithm. The E-step infers the clean pseudolabels by solving an entropy-regularized optimal transport problem, while the M-step updates the classifier parameters via gradient descent. Being modular, VIBE can seamlessly integrate with recent advancements in self-supervised representation learning, which enhance its ability to resist backdoor attacks. We experimentally validate the method effectiveness against contemporary backdoor attacks on standard datasets, a large-scale setup with 1k classes, and a dataset poisoned with multiple attacks. VIBE consistently outperforms previous defenses across all tested scenarios.

Guanxing Lu, Baoxiong Jia, Puhao Li, Yixin Chen, Ziwei Wang, Yansong Tang, Siyuan Huang

Training robot policies within a learned world model is trending due to the inefficiency of real-world interactions. The established image-based world models and policies have shown prior success, but lack robust geometric information that requires consistent spatial and physical understanding of the three-dimensional world, even pre-trained on internet-scale video sources. To this end, we propose a novel branch of world model named Gaussian World Model (GWM) for robotic manipulation, which reconstructs the future state by inferring the propagation of Gaussian primitives under the effect of robot actions. At its core is a latent Diffusion Transformer (DiT) combined with a 3D variational autoencoder, enabling fine-grained scene-level future state reconstruction with Gaussian Splatting. GWM can not only enhance the visual representation for imitation learning agent by self-supervised future prediction training, but can serve as a neural simulator that supports model-based reinforcement learning. Both simulated and real-world experiments depict that GWM can precisely predict future scenes conditioned on diverse robot actions, and can be further utilized to train policies that outperform the state-of-the-art by impressive margins, showcasing the initial data scaling potential of 3D world model.

Chenghu Du, Shengwu Xiong, Yi Rong

Current virtual try-on methods primarily enhance performance through network optimization, employing strategies such as coarse-to-fine structures and ReferenceNet to inject clothing information. However, these methods are often constrained by the limited quantity and diversity of training samples, which ultimately restricts their potential for further improvement. To address this challenge, we propose a unified, mask-free virtual try-on framework. Our approach leverages the inherent strengths of latent diffusion models to enhance the capability of each pipeline component to accurately model the target distribution, thereby achieving superior performance. Specifically, we introduce a text-driven pseudo-input preparation strategy that significantly increases the diversity of clothing regions within the generated person pseudo-samples. This encourages the generator to focus on variations in these areas and improves the model's generalization capability. Within the generator itself, we develop a gated manipulation mechanism to prevent weight forgetting and reduce training costs. Furthermore, we incorporate a texture-aware injection module to explicitly integrate human-perceptible clothing texture information into the generation process. During inference, we propose a refining conditional inference strategy that mitigates the randomness introduced by Gaussian noise. This effectively preserves identity information and fine clothing details in the final results. Extensive experiments demonstrate that our method outperforms existing state-of-the-art virtual try-on approaches.

Guilian Chen, Huisi Wu, Jing Qin

Automated segmentation of polyps from colonoscopy videos is of great clinical significance as it can assist clinicians in making accurate diagnoses and precise interventions. However, video polyp segmentation (VPS) is challenging due to ambiguous polyp boundaries, as well as variations in polyp scale, contrast, and position across consecutive frames. Moreover, to meet clinical requirements, the inference must operate in real-time to enable intraoperative tracking and guidance. In this paper, we propose a novel and efficient segmentation network, STDDNet, which integrates a spatial-aligned temporal modeling strategy and a discriminative dynamic representation learning mechanism, to comprehensively address these challenges by harnessing the advantages of Mamba. Specifically, a spatial-aligned temporal dependency propagation (STDP) module is developed to model temporal consistency from the consecutive frames based on a bidirectional scanning Mamba block. Furthermore, we design a discriminative dynamic feature extraction (DDFE) module to explore frame-wise dynamic information from the structural feature generated by the Mamba block. Such dynamic features can effectively deal with the variations across colonoscopy frames, providing more details for refined segmentation. We extensively evaluate STDDNet on two benchmark datasets, SUN-SEG and CVC-ClinicDB, demonstrating superior segmentation performance compared to state-of-the-art methods while maintaining real-time inference. Codes are available at https://github.com/C-GLGLGL/STDDNet.

Yi Liu, Shengqian Li, Zuzeng Lin, Feng Wang, Si Liu

The current conditional autoregressive image generation methods have shown promising results, yet their potential remains largely unexplored in the practical unsupervised image translation domain, which operates without explicit cross-domain correspondences.A critical limitation stems from the discrete quantization inherent in traditional Vector Quantization-based frameworks, which disrupts gradient flow between the Variational Autoencoder decoder and causal Transformer, impeding end-to-end optimization during adversarial training in image space.To tackle this issue, we propose using Softmax Relaxed Quantization, a novel approach that reformulates codebook selection as a continuous probability mixing process via Softmax, thereby preserving gradient propagation. Building upon this differentiable foundation, we introduce CycleVAR, which reformulates image-to-image translation as image-conditional visual autoregressivegeneration by injecting multi-scale source image tokens as contextual prompts, analogous to prefix-based conditioning in language models.CycleVAR exploits two modes to generate the target image tokens, including (1) serial multi-step generation enabling iterative refinement across scales and (2) parallel one-step generation synthesizing all resolution outputs in a single forward pass.Experimental findings indicate that the parallel one-step generation mode attains superior translation quality with quicker inference speed than the serial multi-step mode in unsupervised scenarios.Furthermore, both quantitativeand qualitative results indicate that CycleVAR surpasses previous state-of-the-artunsupervised image translation models, e.g., CycleGAN-Turbo.

Minghao Wen, Shengjie Wu, Kangkan Wang, Dong Liang

3D Gaussian Splatting based 3D editing has demonstrated impressive performance in recent years. However, the multi-view editing often exhibits significant local inconsistency, especially in areas of non-rigid deformation, which lead to local artifacts, texture blurring, or semantic variations in edited 3D scenes. We also found that the existing editing methods, which rely entirely on text prompts make the editing process a "one-shot deal", making it difficult for users to control the editing degree flexibly. In response to these challenges, we present InterGSEdit, a novel framework for high-quality 3DGS editing via interactively selecting key views with users' preferences. We propose a CLIP-based Semantic Consistency Selection (CSCS) strategy to adaptively screen a group of semantically consistent reference views for each user-selected key view. Then, the cross-attention maps derived from the reference views are used in a weighted Gaussian Splatting unprojectionto construct the 3D Geometry-Consistent Attention Prior (GAP^ 3D ). We project GAP^ 3D to obtain 3D-constrained attention, which are fused with 2D cross-attention via Attention Fusion Network (AFN). AFN employs an adaptive attention strategy that prioritizes 3D-constrained attention for geometric consistency during early inference, and gradually prioritizes 2D cross-attention maps in diffusion for fine-grained features during the later inference. Extensive experiments demonstrate that InterGSEdit achieves state-of-the-art performance, delivering consistent, high-fidelity 3DGS editing with improved user experience.

Zhaoyang Wu, Fang Liu, Licheng Jiao, Shuo Li, Lingling Li, Xu Liu, Puhua Chen, Wenping Ma

Vision-language models like CLIP have demonstrated strong zero-shot generalization, making them valuable for various downstream tasks through prompt learning. However, existing test-time prompt tuning methods, such as entropy minimization, treat both text and visual prompts as fixed learnable parameters, limiting their adaptability to unseen domains. In contrast, we propose Hierarchical Variational Test-Time Prompt Generation, a novel approach where both text and visual prompts are dynamically generated via a HyperTransformer at inference time. This enables the model to produce data-specific prompts for each modality, significantly improving generalization. To further address template sensitivity and distribution shifts, we introduce variational prompt generation, leveraging variational inference to mitigate biases introduced by different prompt templates and data augmentations. Additionally, our hierarchical variational prompt generation conditions prompts at each layer on those from previous layers, allowing the model to capture deeper contextual dependencies and refine prompt interactions for robust adaptation. Extensive experiments on domain generalization benchmarks demonstrate that our method significantly outperforms existing prompt-learning techniques, achieving state-of-the-art zero-shot accuracy while maintaining efficiency.

Francisco Caetano, Christiaan Viviers, Luis A. Zavala-Mondragón, Peter H.N. De With, Fons van der Sommen

Out-of-distribution (OOD) detection holds significant importance across many applications. While semantic and domain-shift OOD problems are well-studied, this work focuses on covariate shifts - subtle variations in the data distribution that can degrade machine learning performance. We hypothesize that detecting these subtle shifts can improve our understanding of in-distribution boundaries, ultimately improving OOD detection. In adversarial discriminators trained with Batch Normalization (BN), real and adversarial samples form distinct domains with unique batch statistics - a property we exploit for OOD detection. We introduce DisCoPatch, an unsupervised Adversarial Variational Autoencoder (VAE) framework that harnesses this mechanism. During inference, batches consist of patches from the same image, ensuring a consistent data distribution that allows the model to rely on batch statistics. DisCoPatch uses the VAE's suboptimal outputs (generated and reconstructed) as negative samples to train the discriminator, thereby improving its ability to delineate the boundary between in-distribution samples and covariate shifts. By tightening this boundary, DisCoPatch achieves state-of-the-art results in public OOD detection benchmarks. The proposed model not only excels in detecting covariate shifts, achieving 95.5% AUROC on ImageNet-1K(-C), but also outperforms all prior methods on public Near-OOD (95.0%) benchmarks. With a compact model size of 25MB, it achieves high OOD detection performance at notably lower latency than existing methods, making it an efficient and practical solution for real-world OOD detection applications.