Increasingly expensive training of ever larger models such as Vision Transfomers motivate reusing the vast library of already trained state-of-the-art networks. However, their latency, high computational costs and memory demands pose significant challenges for deployment, especially on resource-constrained hardware. While structured pruning methods can reduce these factors, they often require costly retraining, sometimes for up to hundreds of epochs, or even training from scratch to recover the lost accuracy resulting from the structural modifications. Maintaining the provided performance of trained models after structured pruning and thereby avoiding extensive retraining remains a challenge. To solve this, we introduce Variance-Based Pruning, a simple and structured one-shot pruning technique for efficiently compressing networks, with minimal finetuning. Our approach first gathers activation statistics, which are then used to select neurons for pruning. Simultaneously the mean activations are integrated back into the model to preserve a high degree of performance. On ImageNet-1k recognition tasks, we demonstrate that directly after pruning DeiT-Base retains over 70% of its original performance and requires only 10 epochs of fine-tuning to regain 99% of the original accuracy while simultaneously reducing MACs by 35% and model size by 36%, thus speeding up the model by 1.44 times.
论文检索
输入标题、作者或关键词，从 2,701 篇学术成果中精准定位
Baking Gaussian Splatting into Diffusion Denoiser for Fast and Scalable Single-stage Image-to-3D Generation and Reconstruction
PDF ↗Existing feedforward image-to-3D methods mainly rely on 2D multi-view diffusion models that cannot guarantee 3D consistency. These methods easily collapse when changing the prompt view direction and mainly handle object-centric cases. In this paper, we propose a novel single-stage 3D diffusion model, DiffusionGS, for object generation and scene reconstruction from a single view. DiffusionGS directly outputs 3D Gaussian point clouds at each timestep to enforce view consistency and allow the model to generate robustly given prompt views of any directions, beyond object-centric inputs. Plus, to improve the capability and generality of DiffusionGS, we scale up 3D training data by developing a scene-object mixed training strategy. Experiments show that DiffusionGS yields improvements of 2.20 dB/23.25 and 1.34 dB/19.16 in PSNR/FID for objects and scenes than the state-of-the-art methods, without using 2D diffusion prior and depth estimator. Plus, our method enjoys over 5x faster speed ( 6s on an A100 GPU). The project page is at https://caiyuanhao1998.github.io/project/DiffusionGS/
Towards More Diverse and Challenging Pre-training for Point Cloud Learning: Self-Supervised Cross Reconstruction with Decoupled Views
PDF ↗Point cloud learning, especially in a self-supervised way without manual labels, has gained growing attention in both vision and learning communities due to its potential utility in a wide range of applications. Most existing generative approaches for point cloud self-supervised learning focus on recovering masked points from visible ones within a single view. Recognizing that a two-view pre-training paradigm inherently introduces greater diversity and variance, it may thus enable more challenging and informative pre-training. Inspired by this, we explore the potential of two-view learning in this domain. In this paper, we propose Point-PQAE, a cross-reconstruction generative paradigm that first generates two decoupled point clouds/views and then reconstructs one from the other. To achieve this goal, we develop a crop mechanism for point cloud view generation for the first time and further propose a novel positional encoding to represent the 3D relative position between the two decoupled views. The cross-reconstruction significantly increases the difficulty of pre-training compared to self-reconstruction, which enables our method to surpass previous single-modal self-reconstruction methods in 3D self-supervised learning. Specifically, it outperforms the self-reconstruction baseline (Point-MAE) by 6.5%, 7.0%, and 6.7% in three variants of ScanObjectNN with the Mlp-Linear evaluation protocol. The code is available at https://github.com/aHapBean/Point-PQAE.
This paper presents Randomized AutoRegressive modeling (RAR) for visual generation, which sets a new state-of-the-art performance on the image generation task while maintaining full compatibility with language modeling frameworks. The proposed RAR is simple: during a standard autoregressive training process with a next-token prediction objective, the input sequence-typically ordered in raster form-is randomly permuted into different factorization orders with a probability r, where r starts at 1 and linearly decays to 0 over the course of training. This annealing training strategy enables the model to learn to maximize the expected likelihood over all factorization orders and thus effectively improve the model's capability of modeling bidirectional contexts. Importantly, RAR preserves the integrity of the autoregressive modeling framework, ensuring full compatibility with language modeling while significantly improving performance in image generation. On the ImageNet-256 benchmark, RAR achieves an FID score of 1.48, not only surpassing prior state-of-the-art autoregressive image generators but also outperforming leading diffusion-based and masked transformer-based methods. Code and models are available at https://github.com/bytedance/1d-tokenizer
Multi-Modal Multi-Task Unified Embedding Model (M3T-UEM): A Task-Adaptive Representation Learning Framework
PDF ↗We present Multi-Modal Multi-Task Unified Embedding Model (M3T-UEM), a framework that advances vision-language matching and retrieval by leveraging a large language model (LLM) backbone. While concurrent LLM-based approaches like VLM2VEC, MM-Embed, NV-Embed, and MM-GEM have demonstrated impressive capabilities in multi-modal and multi-task scenarios, our work introduces novel mechanisms for task-adaptive learning and embedding extraction that further enhance the potential of LLM-based retrieval systems. Our key technical contribution lies in the development of a task-aware contrastive learning framework with an automated Bayesian weighing mechanism. This approach provides a principled way to balance multiple tasks during training, departing from conventional contrastive learning strategies. We further enhance the framework through a multiple-token summarization strategy and an auxiliary language modeling objective, which together significantly improve retrieval performance.Comprehensive experiments on M-BEIR and ICinW benchmarks demonstrate M3T-UEM's effectiveness, showing competitive or superior performance compared to both traditional encoder-based methods and recent LLM-based approaches. Furthermore, we demonstrate particular strengths in handling compositional conceptual changes and multilingual scenarios owing to the incorporation of an LLM backbone where the method drastically outperforms CLIP in zero-shot settings, often by orders of magnitude.
Reconstructing atmospheric surface \text CO _2 is crucial for understanding climate dynamics and informing global mitigation strategies. Traditional inversion models achieve precise global \text CO _2 reconstruction but rely heavily on uncertain prior estimates of fluxes and emissions. Inspired by recent advances in data-driven weather forecasting, we explore whether data-driven models can reduce reliance on these priors. However, \text CO _2 reconstruction presents unique challenges, including complex spatio-temporal dynamics, periodic patterns and sparse observations. We propose \text CO _2-Net, a data-driven model that addresses these challenges without requiring extensive prior data. We formulate \text CO _2 reconstruction as solving a constrained advection-diffusion equation and derive three key components: physics-informed spatio-temporal factorization for capturing complex transport dynamics, wind-based embeddings for modeling periodic variations and a semi-supervised loss for integrating sparse \text CO _2 observations with dense meteorological data. \text CO _2-Net is designed in three sizes---small (S), base (B) and large (L)---to balance performance and efficiency. On CMIP6 reanalysis data, \text CO _2-Net (S) and (L) reduce RMSE by 11% and 71% , respectively, when compared to the best data-driven baseline. On real observations, \text CO _2-Net (L) achieves RMSE comparable to inversion models. The ablation study shows that the effectiveness of wind-based embedding and semi-supervised loss stems from their compatibility with our spatio-temporal factorization. Code is available at https://github.com/Leamonz/CORE.
We propose a neural inverse rendering approach that jointly reconstructs geometry, spatially varying reflectance, and lighting conditions from multi-view images captured under varying directional lighting. Unlike prior multi-view photometric stereo methods that require light calibration or intermediate cues such as per-view normal maps, our method jointly optimizes all scene parameters from raw images in a single stage. We represent both geometry and reflectance as neural implicit fields and apply shadow-aware volume rendering. A spatial network first predicts the signed distance and a reflectance latent code for each scene point. A reflectance network then estimates reflectance values conditioned on the latent code and angularly encoded surface normal, view, and light directions. The proposed method outperforms state-of-the-art normal-guided approaches in shape and lighting estimation accuracy, generalizes to view-unaligned multi-light images, and handles objects with challenging geometry and reflectance.
This work addresses motion-guided few-shot video object segmentation (FSVOS), which aims to segment dynamic objects in videos based on a few annotated examples with the same motion patterns. Existing FSVOS datasets and methods typically focus on object categories, which are static attributes that ignore the rich temporal dynamics in videos, limiting their application in scenarios requiring motion understanding. To fill this gap, we introduce MOVE, a large-scale dataset specifically designed for motion-guided FSVOS. Based on MOVE, we comprehensively evaluate 6 state-of-the-art methods from 3 different related tasks across 2 experimental settings. Our results reveal that current methods struggle to address motion-guided FSVOS, prompting us to analyze the associated challenges and propose a baseline method, Decoupled Motion Appearance Network (DMA). Experiments demonstrate that our approach achieves superior performance in few shot motion understanding, establishing a solid foundation for future research in this direction.
Photorealistic reconstruction of street scenes is essential for developing real-world simulators in autonomous driving. While recent methods based on 3D/4D Gaussian Splatting (GS) have demonstrated promising results, they still encounter challenges in complex street scenes due to the unpredictable motion of dynamic objects. Current methods typically decompose street scenes into static and dynamic objects, learning the Gaussians in either a supervised manner (e.g., w/ 3D bounding-box) or a self-supervised manner (e.g., w/o 3D bounding-box). However, these approaches do not effectively model the motions of dynamic objects (e.g., the motion speed of pedestrians is clearly different from that of vehicles), resulting in suboptimal scene decomposition. To address this, we propose Explicit Motion Decomposition (EMD), which models the motions of dynamic objects by introducing learnable motion embeddings to the Gaussians, enhancing the decomposition in street scenes. The proposed plug-and-play EMD module compensates for the lack of motion modeling in self-supervised street Gaussian splatting methods. We also introduce tailored training strategies to extend EMD to supervised approaches. Comprehensive experiments demonstrate the effectiveness of our method, achieving state-of-the-art novel view synthesis performance in self-supervised settings. The code is available at: https://qingpowuwu.github.io/emd
Whole-body biometric recognition is a challenging multi-modal task that integrates various biometric modalities, including face, gait, and body. This integration is essential for overcoming the limitations of unimodal systems. Traditionally, whole-body recognition involves deploying different models to process multiple modalities, achieving the final outcome by score-fusion (e.g., weighted averaging similarity matrices from each model). However, these conventional methods may overlook the variations in score distributions of individual modalities, making it challenging to improve final performance. In this work, we present Quality-guided Mixture of score-fusion Experts (QME), a novel framework designed for improving whole-body biometric recognition performance through a learnable score-fusion strategy using a Mixture of Experts (MoE). We introduce a novel pseudo quality loss for quality estimation with a modality-specific Quality Estimator (QE), and a score triplet loss to improve the metric performance. Extensive experiments on multiple whole-body biometric datasets demonstrate the effectiveness of our proposed approach, achieving state-of-the-art results across various metrics compared to baseline methods. Our method is effective for multi-modal and multi-model, addressing key challenges such as model misalignment in the similarity score domain and variability in data quality. Code is available at the \href https://github.com/jiezhu23/QME_ICCV25 Project Link
MixA-Q: Revisiting Activation Sparsity for Vision Transformers from a Mixed-Precision Quantization Perspective
PDF ↗In this paper, we propose MixA-Q, a mixed-precision activation quantization framework that leverages intra-layer activation sparsity (a concept widely explored in activation pruning methods) for efficient inference of quantized window-based vision transformers. For a given uniform-bit quantization configuration, MixA-Q separates the batched window computations within Swin blocks and assigns a lower bit width to the activations of less important windows, improving the trade-off between model performance and efficiency. We introduce a Two-Branch Swin Block that processes activations separately in high- and low-bit precision, enabling seamless integration of our method with most quantization-aware training (QAT) and post-training quantization (PTQ) methods, or with simple modifications. Our experimental evaluations over the COCO dataset demonstrate that MixA-Q achieves a training-free 1.35x computational speedup without accuracy loss in PTQ configuration. With QAT, MixA-Q achieves a lossless 1.25x speedup and a 1.53x speedup with only a 1% mAP drop by incorporating activation pruning. Notably, by reducing the quantization error in important regions, our sparsity-aware quantization adaptation improves the mAP of the quantized W4A4 model (with both weights and activations in 4-bit precision) by 0.7%, reducing quantization degradation by 24%.
The remarkable performance of large multimodal models (LMMs) has attracted significant interest from the image segmentation community.To align with the next-token-prediction paradigm, current LMM-driven segmentation methods either use object boundary points to represent masks or introduce special segmentation tokens, whose hidden states are decoded by a segmentation model requiring the original image as input.However, these approaches often suffer from inadequate mask representation and complex architectures, limiting the potential of LMMs.In this work, we propose the Hierarchical Mask Tokenizer (HiMTok), which represents segmentation masks with up to 32 tokens and eliminates the need for the original image during mask de-tokenization.HiMTok allows for compact and coarse-to-fine mask representations, aligning well with the LLM next-token-prediction paradigm and facilitating the direct acquisition of segmentation capabilities.We develop a 3-stage training recipe for progressive learning of segmentation and visual capabilities, featuring a hierarchical mask loss for effective coarse-to-fine learning.Additionally, we enable bidirectional information flow, allowing conversion between bounding boxes and mask tokens to fully leverage multi-task training potential.Extensive experiments demonstrate that our method achieves state-of-the-art performance across various segmentation tasks,while also enhancing visual grounding and maintaining overall visual understanding.
Deep Adaptive Unfolded Network via Spatial Morphology Stripping and Spectral Filtration for Pan-sharpening
PDF ↗In the field of pan-sharpening, existing deep methods are hindered in deepening cross-modal complementarity in the intermediate feature, and lack effective strategies to harness the network entirety for optimal solutions, exhibiting limited feasibility and interpretability due to their black-box designs. Besides, validating pan-sharpening performance in high-level semantic tasks is intractable for the absence of datasets. To tackle these issues, we propose a deep adaptive unfolded network via spatial morphology stripping and spectral filtration for pan-sharpening, which is conceptualized as a linear inverse problem regularized by spatial and spectral priors. Specifically, we incorporate phase-oriented constraints into the spatial prior to facilitate the extraction of modal-invariant spatial morphology by intrinsic decomposition and leverage a physics-driven spectral filtration attention mechanism aligned with the spectral prior to mine the inherent spectral correlation. After transparently unfolding the model into a multi-stage network, an adaptive stage-exiting mechanism is designed to capitalize on fusion diversity by aggregating optimal image patches across candidate stages. To systematically complete the assessment, we construct the first panoptic segmentation dataset as a semantic-level benchmark for pan-sharpening performance validation. Extensive experiments are conducted to verify the merits of our method with state-of-the-arts. Code is available at https://github.com/Baixuzx7/DAPNet.
Decomposing physically-based materials from images into their constituent properties remains challenging, particularly when maintaining both computational efficiency and physical consistency. While recent diffusion-based approaches have shown promise, they face substantial computational overhead due to multiple denoising steps and separate models for different material properties. We present SuperMat, a single-step framework that achieves high-quality material decomposition with one-step inference. This enables end-to-end training with perceptual and re-render losses while decomposing albedo, metallic, and roughness maps at millisecond-scale speeds. We further extend our framework to 3D objects through a UV refinement network, enabling consistent material estimation across viewpoints while maintaining efficiency. Experiments demonstrate that SuperMat achieves state-of-the-art PBR material decomposition quality while reducing inference time from seconds to milliseconds per image, and completes PBR material estimation for 3D objects in approximately 3 seconds. The project page is at https://hyj542682306.github.io/SuperMat/.
Mamba, a selective state-space model, has recently seen widespread application across various visual tasks due to its exceptional ability to capture long-range dependencies. While promising results have been demonstrated in image classification, its potential in multi-label image classification remains underexplored. To bridge this gap, we propose a novel Mamba-based decoder, which utilizes the intrinsic attention of Mamba to aggregate visual information from image features into label embeddings, yielding label-specific visual representations. Building upon this, a MambaML framework is developed for multi-label image classification, which models the self-correlations of image features and label embeddings with bi-directional Mamba, as well as their cross-correlations with the Mamba-based decoder, allowing visual spatial relationships, label semantic dependencies, and cross-modal associations to be explored in a unified system. In this way, robust label-specific visual representations are acquired, facilitating the training of binary classifiers towards accurate label recognition. Experiments on public benchmarks suggest that our MambaML achieves performance comparable to state-of-the-art methods in multi-label image classification, while requiring fewer parameters and computational overhead.
GenieBlue: Integrating both Linguistic and Multimodal Capabilities for Large Language Models on Mobile Devices
PDF ↗Recent advancements in Multimodal Large Language Models (MLLMs) have enabled their deployment on mobile devices. However, challenges persist in maintaining strong language capabilities and ensuring hardware compatibility, both of which are crucial for user experience and practical deployment efficiency. In our deployment process, we observe that existing MLLMs often face performance degradation on pure language tasks, and the current NPU platforms on smartphones do not support the MoE architecture, which is commonly used to preserve pure language capabilities during multimodal training. To address these issues, we systematically analyze methods to maintain pure language capabilities during the training of MLLMs, focusing on both training data and model architecture aspects. Based on these analyses, we propose GenieBlue, an efficient MLLM structural design that integrates both linguistic and multimodal capabilities for LLMs on mobile devices. GenieBlue freezes the original LLM parameters during MLLM training to maintain pure language capabilities. It acquires multimodal capabilities by duplicating specific transformer blocks for full fine-tuning and integrating lightweight LoRA modules. This approach preserves language capabilities while achieving comparable multimodal performance through extensive training. Deployed on smartphone NPUs, GenieBlue demonstrates efficiency and practicality for applications on mobile devices.
Physical adversarial attack methods expose the vulnerabilities of deep neural networks and pose a significant threat to safety-critical scenarios such as autonomous driving. Camouflage-based physical attack is a more promising approach compared to the patch-based attack, offering stronger adversarial effectiveness in complex physical environments. However, most prior work relies on mesh priors of the target object and virtual environments constructed by simulators, which are time-consuming to obtain and inevitably differ from the real world. Moreover, due to the limitations of the backgrounds in training images, previous methods often fail to produce multi-view robust adversarial camouflage and tend to fall into sub-optimal solutions. Due to these reasons, prior work lacks adversarial effectiveness and robustness across diverse viewpoints and physical environments. We propose a physical attack framework based on 3D Gaussian Splatting (3DGS), named PGA, which provides rapid and precise reconstruction with few images, along with photo-realistic rendering capabilities. Our framework further enhances cross-view robustness and adversarial effectiveness by preventing mutual and self-occlusion among Gaussians and employing a min-max optimization approach that adjusts the imaging background of each viewpoint, helping the algorithm filter out non-robust adversarial features. Extensive experiments validate the effectiveness and superiority of PGA. Our code is available at: https://github.com/TRLou/PGA.
DiffVSR: Revealing an Effective Recipe for Taming Robust Video Super-Resolution Against Complex Degradations
PDF ↗Diffusion models have demonstrated exceptional capabilities in image restoration, yet their application to video super-resolution (VSR) faces significant challenges in balancing fidelity with temporal consistency. Our evaluation reveals a critical gap: existing approaches consistently fail on severely degraded videos--precisely where diffusion models' generative capabilities are most needed. We identify that existing diffusion-based VSR methods struggle primarily because they face an overwhelming learning burden: simultaneously modeling complex degradation distributions, content representations, and temporal relationships with limited high-quality training data. To address this fundamental challenge, we present DiffVSR, featuring a Progressive Learning Strategy (PLS) that systematically decomposes this learning burden through staged training, enabling superior performance on complex degradations. Our framework additionally incorporates an Interweaved Latent Transition (ILT) technique that maintains competitive temporal consistency without additional training overhead. Experiments demonstrate that our approach excels in scenarios where competing methods struggle, particularly on severely degraded videos. Our work reveals that addressing the learning strategy, rather than focusing solely on architectural complexity, is the critical path toward robust real-world video super-resolution with diffusion models.
Professional 3D asset creation often requires diverse sculpting brushes to add surface details and geometric structures.Despite recent progress in 3D generation, producing reusable sculpting brushes compatible with artists' workflows remains an open and challenging problem.These sculpting brushes are typically represented as vector displacement maps (VDMs), which existing models cannot easily generate compared to natural images.This paper presents Text2VDM, a novel framework for text-to-VDM brush generation through the deformation of a dense planar mesh guided by score distillation sampling (SDS).The original SDS loss is designed for generating full objects and struggles with generating desirable sub-object structures from scratch in brush generation.We refer to this issue as semantic coupling, which we address by introducing weighted blending of prompt tokens to SDS, resulting in a more accurate target distribution and semantic guidance.Experiments demonstrate that Text2VDM can generate diverse, high-quality VDM brushes for sculpting surface details and geometric structures.Our generated brushes can be seamlessly integrated into mainstream modeling software, enabling various applications such as mesh stylization and real-time interactive modeling.
Gait recognition is a computer vision task that identifies individuals based on their walking patterns. Its performance is commonly evaluated by ranking a gallery of candidates and measuring the identification accuracy at Rank-K. Existing models are typically single-staged, searching for the probe's nearest neighbors in a gallery, using a global feature representation. While these models can excel at retrieving the correct identity within the top-K predictions, they often struggle when hard negatives are among the top shortlist, leading to relatively low performance at the highest ranks (e.g., Rank-1). In this paper, we introduce CarGait, a Re-ranking (re-ordering the top-K list) method for gait recognition, leveraging the fine-grained correlations between pairs of gait sequences, through cross-attention between gait strips. This re-ranking scheme can be adapted to existing single-stage models to enhance their final results. We demonstrate the capabilities of CarGait by extensive experiments on three common gait datasets, Gait3D, GREW, and OU-MVLP, and seven different gait models, showing consistent gains in Rank-1,5 accuracy, while outperforming existing re-ranking approaches, and a strong baseline.