论文检索

Kangan Qian, Jinyu Miao, Xinyu Jiao, Ziang Luo, Zheng Fu, Yining Shi, Yunlong Wang, Kun Jiang, Diange Yang

Reliable spatial and motion perception is essential for safe autonomous navigation. Recently, class-agnostic motion prediction on bird's-eye view (BEV) cell grids derived from LiDAR point clouds has gained significant attention. However, existing frameworks typically perform cell classification and motion prediction on a per-pixel basis, neglecting important motion field priors such as rigidity constraints, temporal consistency, and future interactions between agents. These limitations lead to degraded performance, particularly in sparse and distant regions. To address these challenges, we introduce PriorMotion, an innovative generative framework designed for class-agnostic motion prediction that integrates essential motion priors by modeling them as distributions within a structured latent space. Specifically, our method captures structured motion priors using raster-vector representations and employs a variational autoencoder with distinct dynamic and static components to learn future motion distributions in the latent space. Experiments on the nuScenes dataset demonstrate that PriorMotion outperforms state-of-the-art methods across both traditional metrics and our newly proposed evaluation criteria. Notably, we achieve improvements of approximately 15.24% in accuracy for fast-moving objects, an 3.59% increase in generalization, a reduction of 0.0163 in motion stability, and a 31.52% reduction in prediction errors in distant regions. Further validation on FMCW LiDAR sensors confirms the robustness of our approach.

Lingxiao Li, Kaixuan Fan, Boqing Gong, Xiangyu Yue

Few-shot image generation aims to generate diverse and high-quality images for an unseen class given only a few examples in that class. A key challenge in this task is balancing category consistency and image diversity, which often compete with each other. Moreover, existing methods offer limited control over the attributes of newly generated images. In this work, we propose Hyperbolic Diffusion Autoencoders (HypDAE), a novel approach that operates in hyperbolic space to capture hierarchical relationships among images from seen categories. By leveraging pre-trained foundation models, HypDAE generates diverse new images for unseen categories with exceptional quality by varying stochastic subcodes or semantic codes. Most importantly, the hyperbolic representation introduces an additional degree of control over semantic diversity through the adjustment of radii within the hyperbolic disk. Extensive experiments and visualizations demonstrate that HypDAE significantly outperforms prior methods by achieving a better balance between preserving category-relevant features and promoting image diversity with limited data. Furthermore, HypDAE offers a highly controllable and interpretable generation process.

Xiaolei Wang, Xiaoyang Wang, Huihui Bai, Eng Gee Lim, Jimin Xiao

Existing distillation-based and reconstruction-based methods have a critical limitation: Autoencoder-based frameworks trained exclusively on normal samples unexpectedly well reconstruct abnormal features, leadingto degraded detection performance. We identify this phenomenon as 'anomaly leakage' (AL): the decoder optimized by reconstruction loss tends to directly copy the encoded input, regardless of whether the input is a normal or abnormal feature. To address this issue, we propose a novel framework that explicitly decouples encoded features into normal and abnormal components through a special invertible mapping in a prior latent space. Next, we remove abnormal components and leverage normal remainders for feature reconstruction. Compared to previous methods, the invertible structure can eliminate anomalous information point-to-point without damaging the information of neighboring patches, improving reconstruction. In this process, effective synthetic abnormal features are essential for training the decoupling process. Therefore, we propose applying adversarial training to find suitable perturbations to simulate feature-level anomalies. Extensive experimental evaluations on benchmark datasets, including MVTec AD, VisA, and Real-IAD, demonstrate that our method achieves competitive performance compared to state-of-the-art approaches. Code is available at: DecAD.

Ce Wang, Zhenyu Hu, Wanjie Sun, Zhenzhong Chen

Image rescaling aims to learn the optimal low-resolution (LR) image that can be accurately reconstructed to its original high-resolution (HR) counterpart, providing an efficient image processing and storage method for ultra-high definition media. However, extreme downscaling factors pose significant challenges to the upscaling process due to its highly ill-posed nature, causing existing image rescaling methods to struggle in generating semantically correct structures and perceptual friendly textures. In this work, we propose a novel framework called Timestep-Aware Diffusion Model (TADM) for extreme image rescaling, which performs rescaling operations in the latent space of a pre-trained autoencoder and effectively leverages powerful natural image priors learned by a pre-trained text-to-image diffusion model. Specifically, TADM adopts a pseudo-invertible module to establish the bidirectional mapping between the latent features of the HR image and the target-sized LR image. Then, the rescaled latent features are enhanced by a pre-trained diffusion model to generate more faithful details. Considering the spatially non-uniform degradation caused by the rescaling operation, we propose a novel time-step alignment strategy, which can adaptively allocate the generative capacity of the diffusion model based on the quality of the reconstructed latent features. Extensive experiments demonstrate the superiority of TADM over previous methods in both quantitative and qualitative evaluations. The code will be available at: https://github.com/wwangcece/TADM.

Shengkai Sun, Zefan Zhang, Jianfeng Dong, Zhiyong Cheng, Xiaojun Chang, Meng Wang

Recent advances in the masked autoencoder (MAE) paradigm have significantly propelled self-supervised skeleton-based action recognition. However, most existing approaches limit reconstruction targets to raw joint coordinates or their simple variants, resulting in computational redundancy and limited semantic representation. To address this, we propose a novel General Feature Prediction framework (GFP) for efficient mask skeleton modeling. Our key innovation is replacing conventional low-level reconstruction with high-level feature prediction that spans from local motion patterns to global semantic representations. Specifically, we introduce a collaborative learning framework where a lightweight target generation network dynamically produces diversified supervision signals across spatial-temporal hierarchies, avoiding reliance on pre-computed offline features. The framework incorporates constrained optimization to ensure feature diversity while preventing model collapse. Experiments on NTU RGB+D 60, NTU RGB+D 120 and PKU-MMD demonstrate the benefits of our approach: Computational efficiency (with 6.2x faster training than standard masked skeleton modeling methods) and superior representation quality, achieving state-of-the-art performance in various downstream tasks.

Kyle Sargent, Kyle Hsu, Justin Johnson, Li Fei-Fei, Jiajun Wu

Since the advent of popular visual generation frameworks like VQGAN and Latent Diffusion Models, state-of-the-art image generation systems have generally been two-stage systems that first tokenize or compress visual data into a lower-dimensional latent space before learning a generative model. Tokenizer training typically follows a standard recipe in which images are compressed and reconstructed subject to a combination of MSE, perceptual, and adversarial losses. Diffusion autoencoders have been proposed in prior work as a way to learn end-to-end perceptually-oriented image compression, but have not yet shown state-of-the-art performance on the competitive task of ImageNet-1K reconstruction. In this work, we propose FlowMo, a transformer-based diffusion autoencoder. FlowMo achieves a new state-of-the-art for image tokenization at multiple bitrates. We achieve this without using convolutions, adversarial losses, spatially-aligned 2D latent codes, or distilling from other tokenizers. Our key insight is that FlowMo training should be broken into a mode-matching pre-training stage and a mode-seeking post-training stage. We conduct extensive analysis and ablations, and we additionally train generative models atop the FlowMo tokenizer and verify the performance. We will release our code and model checkpoints upon acceptance.

Hyunjoon Lee, Joonkyu Min, Jaesik Park

3D Gaussian Splatting (3DGS) has begun incorporating rich information from 2D foundation models. However, most approaches rely on a bottom-up optimization process that treats raw 2D features as ground truth, incurring increased computational costs. We propose a top-down pipeline for constructing compact and fast 3D feature fields, namely, CF3. We first perform a fast weighted fusion of multi-view 2D features with pre-trained Gaussians. This approach enables training a per-Gaussian autoencoder directly on the lifted features, instead of training autoencoders in the 2D domain. As a result, the autoencoder better aligns with the inference-time feature distribution, achieving high compression fidelity and storage efficiency. Finally, we introduce an adaptive sparsification method that prunes and merges Gaussians, constructing an efficient representation while maintaining details. Our approach achieves a competitive 3D feature field while using as little as 5% of the Gaussians compared to Feature-3DGS.

Jiahui Geng, Qing Li

Unlearning methods for vision-language models (VLMs) have primarily adapted techniques from large language models (LLMs), relying on weight updates that demand extensive annotated forget sets. Moreover, these methods perform unlearning at a coarse granularity, often leading to excessive forgetting and reduced model utility. To address this issue, we introduce SAUCE, a novel method that leverages sparse autoencoders (SAEs) for fine-grained and selective concept unlearning in VLMs. Briefly, SAUCE first trains SAEs to capture high-dimensional, semantically rich sparse features. It then identifies the features most relevant to the target concept for unlearning. During inference, it selectively modifies these features to suppress specific concepts while preserving unrelated information. We evaluate SAUCE on two distinct VLMs, LLaVA-v1.5-7B and LLaMA-3.2-11B-Vision-Instruct, across two types of tasks: concrete concept unlearning (objects and sports scenes) and abstract concept unlearning (emotions, colors, and materials), encompassing a total of 60 concepts. Extensive experiments demonstrate that SAUCE outperforms state-of-the-art methods by 18.04% in unlearning quality while maintaining comparable model utility. Furthermore, we investigate SAUCE's robustness against widely used adversarial attacks, its transferability across models, and its scalability in handling multiple simultaneous unlearning requests. Our findings establish SAUCE as an effective and scalable solution for selective concept unlearning in VLMs.

Baoyou Chen, Ce Liu, Weihao Yuan, Zilong Dong, Siyu Zhu

Video face restoration faces a critical challenge in maintaining temporal consistency while recovering fine facial details from degraded inputs. This paper presents a novel approach that extends Vector-Quantized Variational Autoencoders (VQ-VAEs), pretrained on static high-quality portraits, into a video restoration framework through variational latent space modeling. Our key innovation lies in reformulating discrete codebook representations as Dirichlet-distributed continuous variables, enabling probabilistic transitions between facial features across frames. A spatio-temporal Transformer architecture jointly models inter-frame dependencies and predicts latent distributions, while a Laplacian-constrained reconstruction loss combined with perceptual (LPIPS) regularization enhances both pixel accuracy and visual quality. Comprehensive evaluations on blind face restoration, video inpainting, and facial colorization tasks demonstrate state-of-the-art performance. This work establishes an effective paradigm for adapting intensive image priors, pretrained on high-quality images, to video restoration while addressing the critical challenge of flicker artifacts. The source code has been open-sourced and is available at https://github.com/fudan-generative-vision/DicFace.

Uzay Gökay, Federico Spurio, Dominik R. Bach, Juergen Gall

Current state-of-the-art methods for skeleton-based temporal action segmentation are predominantly supervised and require annotated data, which is expensive to collect. In contrast, existing unsupervised temporal action segmentation methods have focused primarily on video data, while skeleton sequences remain underexplored, despite their relevance to real-world applications, robustness, and privacy-preserving nature. In this paper, we propose a novel approach for unsupervised skeleton-based temporal action segmentation. Our method utilizes a sequence-to-sequence temporal autoencoder that keeps the information of the different joints disentangled in the embedding space. Latent skeleton sequences are then divided into non-overlapping patches and quantized to obtain distinctive skeleton motion words, driving the discovery of semantically meaningful action clusters. We thoroughly evaluate the proposed approach on three widely used skeleton-based datasets, namely HuGaDB, LARa, and BABEL. The results demonstrate that our model outperforms the current state-of-the-art unsupervised temporal action segmentation methods.

Tian-Xing Xu, Xiangjun Gao, Wenbo Hu, Xiaoyu Li, Song-Hai Zhang, Ying Shan

Despite remarkable advancements in video depth estimation, existing methods fall short in geometric fidelity due to their affine-invariant predictions, restricting their applicability in reconstruction and other metrically grounded downstream tasks. We propose a novel point map Variational Autoencoder (VAE) for encoding and decoding unbounded point maps. Notably, its latent space is agnostic to video latent distributions of video diffusion models, allowing us to leverage generation priors to model the distribution of point map sequences conditioned on the input videos. Thus, we can recover high-fidelity point map sequences with temporal coherence from open-world videos, facilitating accurate 3D/4D reconstruction, camera parameter estimation, and other depth-based applications. Extensive evaluations on diverse datasets demonstrate that our method achieves state-of-the-art 3D accuracy, temporal consistency, and generalization capability.

Goker Erdogan, Nikhil Parthasarathy, Catalin Ionescu, Drew A. Hudson, Alexander Lerchner, Andrew Zisserman, Mehdi S. M. Sajjadi, Joao Carreira

We introduce LayerLock, a simple yet effective approach for self-supervised visual representation learning, that gradually transitions throughout training from predicting shallow features to deeper ones through progressive layer freezing. First, we make the observation that during training of video masked-autoencoding (MAE) models, ViT layers converge in the order of their depth: shallower layers converge early, deeper layers converge late. We then show that this observation can be exploited to accelerate standard MAE by progressively freezing the model according to an explicit schedule, throughout training. Furthermore, this same schedule can be used in a simple and scalable approach to latent prediction that does not suffer from "representation collapse". We apply our proposed approach, LayerLock, to both pixel and latent prediction approaches with large models of up to 4B parameters and show improvements on both semantic (action classification) and low level (depth estimation) vision tasks.

Yuxuan Wang, Xuanyu Yi, Haohan Weng, Qingshan Xu, Xiaokang Wei, Xianghui Yang, Chunchao Guo, Long Chen, Hanwang Zhang

Triangle meshes are fundamental to 3D applications. Current automatic mesh generation methods typically rely on intermediate representations that lack the continuous surface quality inherent to meshes. Converting these representations into meshes produces dense, suboptimal outputs. Although recent autoregressive approaches demonstrate promise in directly modeling mesh vertices and faces, they are constrained by the limitation in face count, scalability, and structural fidelity.To address these challenges, we propose Nautilus, a locality-aware autoencoder for artist-like mesh generation that leverages the local properties of manifold meshes to achieve structural fidelity and efficient representation. Our approach introduces a novel tokenization algorithm that preserves face proximity relationships and compresses sequence length through locally shared vertices and edges, enabling the generation of meshes with an unprecedented scale of up to 5,000 faces. Furthermore, we develop a Dual-stream Point Conditioner that captures fine-grained geometry, ensuring global consistency and local structural fidelity. Our experiments demonstrate that Nautilus significantly outperforms existing methods in generation quality.

Yu Cheng, Fajie Yuan

Recent advances in Latent Video Diffusion Models (LVDMs) have revolutionized video generation by leveraging Video Variational Autoencoders (Video VAEs) to compress intricate video data into a compact latent space. However, as LVDM training scales, the computational overhead of Video VAEs becomes a critical bottleneck, particularly for encoding high-resolution videos. To address this, we propose LeanVAE, a novel and ultra-efficient Video VAE framework that introduce two key innovations: (1) a lightweight architecture based on a Neighborhood-Aware Feedforward (NAF) module and non-overlapping patch operations, drastically reducing computational cost, and (2) the integration of wavelet transforms and compressed sensing techniques to enhance reconstruction quality. Extensive experiments validate LeanVAE's superiority in video reconstruction and generation, particularly in enhancing efficiency over existing Video VAEs. Our model offers up to 50x fewer FLOPs and 44x faster inference speed while maintaining competitive reconstruction quality, providing insights for scalable, efficient video generation. Our models and code are available at https://github.com/westlake-repl/LeanVAE.

Quankai Gao, Iliyan Georgiev, Tuanfeng Y. Wang, Krishna Kumar Singh, Ulrich Neumann, Jae Shin Yoon

3D generation has made significant progress, however, it still largely remains at the object-level. Feedforward 3D scene-level generation has been rarely explored due to the lack of models capable of scaling-up latent representation learning on 3D scene-level data. Unlike object-level generative models, which are trained on well-labeled 3D data in a bounded canonical space, scene-level generations with 3D scenes represented by 3D Gaussian Splatting (3DGS) are unbounded and exhibit scale inconsistency across different scenes, making unified latent representation learning for generative purposes extremely challenging. In this paper, we introduce Can3Tok, the first 3D scene-level variational autoencoder (VAE) capable of encoding a large number of Gaussian primitives into a low-dimensional latent embedding, which effectively captures both semantic and spatial information of the inputs. Beyond model design, we propose a general pipeline for 3D scene data processing to address scale inconsistency issue. We validate our method on the recent scene-level 3D dataset DL3DV-10K, where we found that only Can3Tok successfully generalizes to novel 3D scenes, while compared methods fail to converge on even a few hundred scene inputs during training and exhibit zero generalization ability during inference. Finally, we demonstrate image-to-3DGS and text-to-3DGS generation as our applications to demonstrate it's ability to faciliate downstream generation tasks. Code will be released.

Xihua Wang, Xin Cheng, Yuyue Wang, Ruihua Song, Yunfeng Wang

Video-to-audio (V2A) generation aims to synthesize temporally aligned, realistic sounds for silent videos, a critical capability for immersive multimedia applications. Current V2A methods, predominantly based on diffusion or flow models, rely on suboptimal noise-to-audio paradigms that entangle cross-modal mappings with stochastic priors, resulting in inefficient training and convoluted transport paths. We propose VAFlow, a novel flow-based framework that directly models the video-to-audio transformation, eliminating reliance on noise priors. To address modality discrepancies, we employ an alignment variational autoencoder that compresses heterogeneous video features into audio-aligned latent spaces while preserving spatiotemporal semantics. By retaining cross-attention mechanisms between video features and flow blocks, our architecture enables classifier-free guidance within video source-driven generation. Without external data or complex training tricks, VAFlow achieves state-of-the-art performance on VGGSound benchmark, surpassing even text-augmented models in audio fidelity, diversity, and distribution alignment. This work establishes a new paradigm for V2A generation with a direct and effective video-to-audio transformation via flow matching.

Jungwoo Huh, Yeseung Park, Seongjean Kim, Jungsu Kim, Sanghoon Lee

Human motion estimation models typically assume a fixed number of input frames, making them sensitive to variations in frame rate and leading to inconsistent motion predictions across different temporal resolutions. This limitation arises because input frame rates inherently determine the temporal granularity of motion capture, causing discrepancies when models trained on a specific frame rate encounter different sampling frequencies. To address this challenge, we propose MBTI (Masked Blending Transformers with Implicit Positional Encoding), a frame rate-agnostic human motion estimation framework designed to maintain temporal consistency across varying input frame rates. Our approach leverages a masked autoencoder (MAE) architecture with masked token blending, which aligns input tokens with a predefined high-reference frame rate, ensuring a standardized temporal representation. Additionally, we introduce implicit positional encoding, which encodes absolute time information using neural implicit functions, enabling more natural motion reconstruction beyond discrete sequence indexing. By reconstructing motion at a high reference frame rate and optional downsampling, MBTI ensures both frame rate generalization and temporal consistency. To comprehensively evaluate MBTI, we introduce EMDB-FPS, an augmented benchmark designed to assess motion estimation robustness across multiple frame rates in both local and global motion estimation tasks. To further assess MBTI's robustness, we introduce the Motion Consistency across Frame rates (MCF), a novel metric to quantify the deviation of motion predictions across different input frame rates. Our results demonstrate that MBTI outperforms state-of-the-art methods in both motion accuracy and temporal consistency, achieving the most stable and consistent motion predictions across varying frame rates.

Tongyan Hua, Lutao Jiang, Ying-Cong Chen, Wufan Zhao

Recent advancements in generative models have enabled 3D urban scene generation from satellite imagery, unlocking promising applications in gaming, digital twins, and beyond. However, most existing methods rely heavily on neural rendering techniques, which hinder their ability to produce detailed 3D structures on a broader scale, largely due to the inherent structural ambiguity derived from relatively limited 2D observations. To address this challenge, we propose Sat2City, a novel framework that synergizes the representational capacity of sparse voxel grids with latent diffusion models, tailored specifically for our novel 3D city dataset. Our approach is enabled by three key components: (1) A cascaded latent diffusion framework that progressively recovers 3D city structures from satellite imagery, (2) a Re-Hash operation at its Variational Autoencoder (VAE) bottleneck to compute multi-scale feature grids for stable appearance optimization, and (3) an inverse sampling strategy enabling implicit supervision for smooth appearance transitioning. To overcome the challenge of collecting real-world city-scale 3D models with high-quality geometry and appearance, we introduce a dataset of synthesized large-scale 3D cities paired with satellite-view height maps. Validated on this dataset, our framework generates detailed 3D structures from a single satellite image, achieving superior fidelity compared to existing city generation models.

Dahye Kim, Xavier Thomas, Deepti Ghadiyaram

We study how rich visual semantic information is represented within various layers and denoising timesteps of different diffusion architectures. We uncover monosemantic interpretable features by leveraging k-sparse autoencoders (k-SAE). We substantiate our mechanistic interpretations via transfer learning using light-weight classifiers on off-the-shelf diffusion models' features. On 4 datasets, we demonstrate the effectiveness of diffusion features for representation learning. We provide an in-depth analysis of how different diffusion architectures, pre-training datasets, and language model conditioning impacts visual representation granularity, inductive biases, and transfer learning capabilities. Our work is a critical step towards deepening interpretability of black-box diffusion models. Code and visualizations available at: https://github.com/revelio-diffusion/revelio

Lanmiao Liu, Esam Ghaleb, Asli Ozyurek, Zerrin Yumak

Creating a virtual avatar with semantically coherent gestures that are aligned with speech is a challenging task. Existing gesture generation research mainly focused on generating rhythmic beat gestures, neglecting the semantic context of the gestures. In this paper, we propose a novel approach for semantic grounding in co-speech gesture generation that integrates semantic information at both fine-grained and global levels. Our approach starts with learning the motion prior through a vector-quantized variational autoencoder. Built on this model, a second-stage module is applied to automatically generate gestures from speech, text-based semantics and speaker identity that ensures consistency between the semantic relevance of generated gestures and co-occurring speech semantics through semantic coherence and relevance modules. Experimental results demonstrate that our approach enhances the realism and coherence of semantic gestures. Extensive experiments and user studies show that our method outperforms state-of-the-art approaches across two benchmarks in co-speech gesture generation in both objective and subjective metrics. The qualitative results of our model can be viewed at https://semgesture.github.io. Our code, dataset and pre-trained models will be shared upon acceptance.