Guangrun Wang

Changlin Li, Bohan Zhuang, Guangrun Wang et al.

Recent advances in vision Transformers (ViTs) have come with a voracious appetite for computing power, high-lighting the urgent need to develop efficient training methods for ViTs. Progressive learning, a training scheme where the model capacity grows progressively during training, has started showing its ability in efficient training. In this paper, we take a practical step towards efficient training of ViTs by customizing and automating progressive learning. First, we develop a strong manual baseline for progressive learning of ViTs, by introducing momentum growth (MoGrow) to bridge the gap brought by model growth. Then, we propose automated progressive learning (AutoProg), an efficient training scheme that aims to achieve lossless acceleration by automatically increasing the training overload on-the-fly; this is achieved by adaptively deciding whether, where and how much should the model grow during progressive learning. Specifically, we first relax the optimization of the growth schedule to sub-network architecture optimization problem, then propose one-shot estimation of the sub-network performance via an elastic supernet. The searching overhead is reduced to minimal by recycling the parameters of the supernet. Extensive experiments of efficient training on ImageNet with two representative ViT models, DeiT and VOLO, demonstrate that AutoProg can accelerate ViTs training by up to 85.1% with no performance drop. Code: https://github.com/changlin31/AutoProg

Tao Tang, Changlin Li, Guangrun Wang et al.

Automatic data augmentation (AutoAugment) strategies are indispensable in supervised data-efficient training protocols of vision transformers, and have led to state-of-the-art results in supervised learning. Despite the success, its development and application on self-supervised vision transformers have been hindered by several barriers, including the high search cost, the lack of supervision, and the unsuitable search space. In this work, we propose AutoView, a self-regularized adversarial AutoAugment method, to learn views for self-supervised vision transformers, by addressing the above barriers. First, we reduce the search cost of AutoView to nearly zero by learning views and network parameters simultaneously in a single forward-backward step, minimizing and maximizing the mutual information among different augmented views, respectively. Then, to avoid information collapse caused by the lack of label supervision, we propose a self-regularized loss term to guarantee the information propagation. Additionally, we present a curated augmentation policy search space for self-supervised learning, by modifying the generally used search space designed for supervised learning. On ImageNet, our AutoView achieves remarkable improvement over RandAug baseline (+10.2% k-NN accuracy), and consistently outperforms sota manually tuned view policy by a clear margin (up to +1.3% k-NN accuracy). Extensive experiments show that AutoView pretraining also benefits downstream tasks (+1.2% mAcc on ADE20K Semantic Segmentation and +2.8% mAP on revisited Oxford Image Retrieval benchmark) and improves model robustness (+2.3% Top-1 Acc on ImageNet-A and +1.0% AUPR on ImageNet-O). Code and models will be available at https://github.com/Trent-tangtao/AutoView.

Xipeng Chen, Guangrun Wang, Dizhong Zhu et al.

3D Garment modeling is a critical and challenging topic in the area of computer vision and graphics, with increasing attention focused on garment representation learning, garment reconstruction, and controllable garment manipulation, whereas existing methods were constrained to model garments under specific categories or with relatively simple topologies. In this paper, we propose a novel Neural Sewing Machine (NSM), a learning-based framework for structure-preserving 3D garment modeling, which is capable of learning representations for garments with diverse shapes and topologies and is successfully applied to 3D garment reconstruction and controllable manipulation. To model generic garments, we first obtain sewing pattern embedding via a unified sewing pattern encoding module, as the sewing pattern can accurately describe the intrinsic structure and the topology of the 3D garment. Then we use a 3D garment decoder to decode the sewing pattern embedding into a 3D garment using the UV-position maps with masks. To preserve the intrinsic structure of the predicted 3D garment, we introduce an inner-panel structure-preserving loss, an inter-panel structure-preserving loss, and a surface-normal loss in the learning process of our framework. We evaluate NSM on the public 3D garment dataset with sewing patterns with diverse garment shapes and categories. Extensive experiments demonstrate that the proposed NSM is capable of representing 3D garments under diverse garment shapes and topologies, realistically reconstructing 3D garments from 2D images with the preserved structure, and accurately manipulating the 3D garment categories, shapes, and topologies, outperforming the state-of-the-art methods by a clear margin.

Zijian He, Peixin Chen, Guangrun Wang et al.

Video virtual try-on aims to generate realistic sequences that maintain garment identity and adapt to a person's pose and body shape in source videos. Traditional image-based methods, relying on warping and blending, struggle with complex human movements and occlusions, limiting their effectiveness in video try-on applications. Moreover, video-based models require extensive, high-quality data and substantial computational resources. To tackle these issues, we reconceptualize video try-on as a process of generating videos conditioned on garment descriptions and human motion. Our solution, WildVidFit, employs image-based controlled diffusion models for a streamlined, one-stage approach. This model, conditioned on specific garments and individuals, is trained on still images rather than videos. It leverages diffusion guidance from pre-trained models including a video masked autoencoder for segment smoothness improvement and a self-supervised model for feature alignment of adjacent frame in the latent space. This integration markedly boosts the model's ability to maintain temporal coherence, enabling more effective video try-on within an image-based framework. Our experiments on the VITON-HD and DressCode datasets, along with tests on the VVT and TikTok datasets, demonstrate WildVidFit's capability to generate fluid and coherent videos. The project page website is at wildvidfit-project.github.io.

Tang Tao, Longfei Gao, Guangrun Wang et al.

We introduce a new task, novel view synthesis for LiDAR sensors. While traditional model-based LiDAR simulators with style-transfer neural networks can be applied to render novel views, they fall short of producing accurate and realistic LiDAR patterns because the renderers rely on explicit 3D reconstruction and exploit game engines, that ignore important attributes of LiDAR points. We address this challenge by formulating, to the best of our knowledge, the first differentiable end-to-end LiDAR rendering framework, LiDAR-NeRF, leveraging a neural radiance field (NeRF) to facilitate the joint learning of geometry and the attributes of 3D points. However, simply employing NeRF cannot achieve satisfactory results, as it only focuses on learning individual pixels while ignoring local information, especially at low texture areas, resulting in poor geometry. To this end, we have taken steps to address this issue by introducing a structural regularization method to preserve local structural details. To evaluate the effectiveness of our approach, we establish an object-centric multi-view LiDAR dataset, dubbed NeRF-MVL. It contains observations of objects from 9 categories seen from 360-degree viewpoints captured with multiple LiDAR sensors. Our extensive experiments on the scene-level KITTI-360 dataset, and on our object-level NeRF-MVL show that our LiDAR-NeRF surpasses the model-based algorithms significantly.

Pengzhen Ren, Changlin Li, Hang Xu et al.

Recently, great success has been made in learning visual representations from text supervision, facilitating the emergence of text-supervised semantic segmentation. However, existing works focus on pixel grouping and cross-modal semantic alignment, while ignoring the correspondence among multiple augmented views of the same image. To overcome such limitation, we propose multi-\textbf{View} \textbf{Co}nsistent learning (ViewCo) for text-supervised semantic segmentation. Specifically, we first propose text-to-views consistency modeling to learn correspondence for multiple views of the same input image. Additionally, we propose cross-view segmentation consistency modeling to address the ambiguity issue of text supervision by contrasting the segment features of Siamese visual encoders. The text-to-views consistency benefits the dense assignment of the visual features by encouraging different crops to align with the same text, while the cross-view segmentation consistency modeling provides additional self-supervision, overcoming the limitation of ambiguous text supervision for segmentation masks. Trained with large-scale image-text data, our model can directly segment objects of arbitrary categories in a zero-shot manner. Extensive experiments show that ViewCo outperforms state-of-the-art methods on average by up to 2.9\%, 1.6\%, and 2.4\% mIoU on PASCAL VOC2012, PASCAL Context, and COCO, respectively.

Guangrun Wang, Philip H. S. Torr

Classifiers and generators have long been separated. We break down this separation and showcase that conventional neural network classifiers can generate high-quality images of a large number of categories, being comparable to the state-of-the-art generative models (e.g., DDPMs and GANs). We achieve this by computing the partial derivative of the classification loss function with respect to the input to optimize the input to produce an image. Since it is widely known that directly optimizing the inputs is similar to targeted adversarial attacks incapable of generating human-meaningful images, we propose a mask-based stochastic reconstruction module to make the gradients semantic-aware to synthesize plausible images. We further propose a progressive-resolution technique to guarantee fidelity, which produces photorealistic images. Furthermore, we introduce a distance metric loss and a non-trivial distribution loss to ensure classification neural networks can synthesize diverse and high-fidelity images. Using traditional neural network classifiers, we can generate good-quality images of 256$\times$256 resolution on ImageNet. Intriguingly, our method is also applicable to text-to-image generation by regarding image-text foundation models as generalized classifiers. Proving that classifiers have learned the data distribution and are ready for image generation has far-reaching implications, for classifiers are much easier to train than generative models like DDPMs and GANs. We don't even need to train classification models because tons of public ones are available for download. Also, this holds great potential for the interpretability and robustness of classifiers. Project page is at \url{https://classifier-as-generator.github.io/}.

Binbin Yang, Yi Luo, Ziliang Chen et al.

Thanks to the rapid development of diffusion models, unprecedented progress has been witnessed in image synthesis. Prior works mostly rely on pre-trained linguistic models, but a text is often too abstract to properly specify all the spatial properties of an image, e.g., the layout configuration of a scene, leading to the sub-optimal results of complex scene generation. In this paper, we achieve accurate complex scene generation by proposing a semantically controllable Layout-AWare diffusion model, termed LAW-Diffusion. Distinct from the previous Layout-to-Image generation (L2I) methods that only explore category-aware relationships, LAW-Diffusion introduces a spatial dependency parser to encode the location-aware semantic coherence across objects as a layout embedding and produces a scene with perceptually harmonious object styles and contextual relations. To be specific, we delicately instantiate each object's regional semantics as an object region map and leverage a location-aware cross-object attention module to capture the spatial dependencies among those disentangled representations. We further propose an adaptive guidance schedule for our layout guidance to mitigate the trade-off between the regional semantic alignment and the texture fidelity of generated objects. Moreover, LAW-Diffusion allows for instance reconfiguration while maintaining the other regions in a synthesized image by introducing a layout-aware latent grafting mechanism to recompose its local regional semantics. To better verify the plausibility of generated scenes, we propose a new evaluation metric for the L2I task, dubbed Scene Relation Score (SRS) to measure how the images preserve the rational and harmonious relations among contextual objects. Comprehensive experiments demonstrate that our LAW-Diffusion yields the state-of-the-art generative performance, especially with coherent object relations.

Pengzhen Ren, Changlin Li, Guangrun Wang et al.

Recently, a surge of interest in visual transformers is to reduce the computational cost by limiting the calculation of self-attention to a local window. Most current work uses a fixed single-scale window for modeling by default, ignoring the impact of window size on model performance. However, this may limit the modeling potential of these window-based models for multi-scale information. In this paper, we propose a novel method, named Dynamic Window Vision Transformer (DW-ViT). The dynamic window strategy proposed by DW-ViT goes beyond the model that employs a fixed single window setting. To the best of our knowledge, we are the first to use dynamic multi-scale windows to explore the upper limit of the effect of window settings on model performance. In DW-ViT, multi-scale information is obtained by assigning windows of different sizes to different head groups of window multi-head self-attention. Then, the information is dynamically fused by assigning different weights to the multi-scale window branches. We conducted a detailed performance evaluation on three datasets, ImageNet-1K, ADE20K, and COCO. Compared with related state-of-the-art (SoTA) methods, DW-ViT obtains the best performance. Specifically, compared with the current SoTA Swin Transformers \cite{liu2021swin}, DW-ViT has achieved consistent and substantial improvements on all three datasets with similar parameters and computational costs. In addition, DW-ViT exhibits good scalability and can be easily inserted into any window-based visual transformers.

Guangyi Chen, Xiao Liu, Guangrun Wang et al.

Video-language pre-trained models have shown remarkable success in guiding video question-answering (VideoQA) tasks. However, due to the length of video sequences, training large-scale video-based models incurs considerably higher costs than training image-based ones. This motivates us to leverage the knowledge from image-based pretraining, despite the obvious gaps between image and video domains. To bridge these gaps, in this paper, we propose Tem-Adapter, which enables the learning of temporal dynamics and complex semantics by a visual Temporal Aligner and a textual Semantic Aligner. Unlike conventional pretrained knowledge adaptation methods that only concentrate on the downstream task objective, the Temporal Aligner introduces an extra language-guided autoregressive task aimed at facilitating the learning of temporal dependencies, with the objective of predicting future states based on historical clues and language guidance that describes event progression. Besides, to reduce the semantic gap and adapt the textual representation for better event description, we introduce a Semantic Aligner that first designs a template to fuse question and answer pairs as event descriptions and then learns a Transformer decoder with the whole video sequence as guidance for refinement. We evaluate Tem-Adapter and different pre-train transferring methods on two VideoQA benchmarks, and the significant performance improvement demonstrates the effectiveness of our method.

Kaixin Cai, Pengzhen Ren, Yi Zhu et al.

Recently, semantic segmentation models trained with image-level text supervision have shown promising results in challenging open-world scenarios. However, these models still face difficulties in learning fine-grained semantic alignment at the pixel level and predicting accurate object masks. To address this issue, we propose MixReorg, a novel and straightforward pre-training paradigm for semantic segmentation that enhances a model's ability to reorganize patches mixed across images, exploring both local visual relevance and global semantic coherence. Our approach involves generating fine-grained patch-text pairs data by mixing image patches while preserving the correspondence between patches and text. The model is then trained to minimize the segmentation loss of the mixed images and the two contrastive losses of the original and restored features. With MixReorg as a mask learner, conventional text-supervised semantic segmentation models can achieve highly generalizable pixel-semantic alignment ability, which is crucial for open-world segmentation. After training with large-scale image-text data, MixReorg models can be applied directly to segment visual objects of arbitrary categories, without the need for further fine-tuning. Our proposed framework demonstrates strong performance on popular zero-shot semantic segmentation benchmarks, outperforming GroupViT by significant margins of 5.0%, 6.2%, 2.5%, and 3.4% mIoU on PASCAL VOC2012, PASCAL Context, MS COCO, and ADE20K, respectively.

Zijian Song, Qichang Li, Sihan Qin et al.

The scarcity of large-scale robotic data has motivated the repurposing of foundation models from other modalities for policy learning. In this work, we introduce PhysGen (Learning Physics from Pretrained Video Generation Models), a scalable continuous and sequential world interaction framework that leverages autoregressive video generation to solve robotic manipulation tasks. By treating the pretrained video model as a proxy for a physics simulator, PhysGen models the dynamic interplay between the external environment and robot actions. We introduce a multimodal continuous representation that unifies video and action into shared physical tokens, bridging the gap between discrete video generation and continuous robotic control. This approach enables the seamless transfer of implicit physical knowledge-such as object permanence and dynamics-from video pretraining to downstream manipulation.To ensure efficient convergence, we incorporate causal masking, inverse kinematics, Lookahead Multi-Token Prediction (L-MTP), and key-value (KV) caching. Experimental results on the Libero and ManiSkill benchmarks demonstrate that PhysGen consistently outperforms robust baselines, surpassing OpenVLA and WorldVLA by margins of 13.8% and 8.8%, respectively. Notably, in real-world scenarios, PhysGen matches the performance of large-scale action-pretrained models like $π_0$ without requiring prior action-specific pretraining, demonstrating superior capability in physically complex tasks such as grasping transparent objects. These findings validate the potential of extracting physical intuition from pretrained video generators to facilitate generalizable robotic manipulation.

Zhijian Huang, Tao Tang, Shaoxiang Chen et al.

Data-driven approaches for autonomous driving (AD) have been widely adopted in the past decade but are confronted with dataset bias and uninterpretability. Inspired by the knowledge-driven nature of human driving, recent approaches explore the potential of large language models (LLMs) to improve understanding and decision-making in traffic scenarios. They find that the pretrain-finetune paradigm of LLMs on downstream data with the Chain-of-Thought (CoT) reasoning process can enhance explainability and scene understanding. However, such a popular strategy proves to suffer from the notorious problems of misalignment between the crafted CoTs against the consequent decision-making, which remains untouched by previous LLM-based AD methods. To address this problem, we motivate an end-to-end decision-making model based on multimodality-augmented LLM, which simultaneously executes CoT reasoning and carries out planning results. Furthermore, we propose a reasoning-decision alignment constraint between the paired CoTs and planning results, imposing the correspondence between reasoning and decision-making. Moreover, we redesign the CoTs to enable the model to comprehend complex scenarios and enhance decision-making performance. We dub our proposed large language planners with reasoning-decision alignment as RDA-Driver. Experimental evaluations on the nuScenes and DriveLM-nuScenes benchmarks demonstrate the effectiveness of our RDA-Driver in enhancing the performance of end-to-end AD systems. Specifically, our RDA-Driver achieves state-of-the-art planning performance on the nuScenes dataset with 0.80 L2 error and 0.32 collision rate, and also achieves leading results on challenging DriveLM-nuScenes benchmarks with 0.82 L2 error and 0.38 collision rate.

Guangcong Wang, Guangrun Wang, Wenqi Liang et al.

We present a weight similarity measure method that can quantify the weight similarity of non-convex neural networks. To understand the weight similarity of different trained models, we propose to extract the feature representation from the weights of neural networks. We first normalize the weights of neural networks by introducing a chain normalization rule, which is used for weight representation learning and weight similarity measure. We extend the traditional hypothesis-testing method to a hypothesis-training-testing statistical inference method to validate the hypothesis on the weight similarity of neural networks. With the chain normalization rule and the new statistical inference, we study the weight similarity measure on Multi-Layer Perceptron (MLP), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN), and find that the weights of an identical neural network optimized with the Stochastic Gradient Descent (SGD) algorithm converge to a similar local solution in a metric space. The weight similarity measure provides more insight into the local solutions of neural networks. Experiments on several datasets consistently validate the hypothesis of weight similarity measure.

Xiao Liu, Guangyi Chen, Yansong Tang et al.

Composing simple elements into complex concepts is crucial yet challenging, especially for 3D action generation. Existing methods largely rely on extensive neural language annotations to discern composable latent semantics, a process that is often costly and labor-intensive. In this study, we introduce a novel framework to generate compositional actions without reliance on language auxiliaries. Our approach consists of three main components: Action Coupling, Conditional Action Generation, and Decoupling Refinement. Action Coupling utilizes an energy model to extract the attention masks of each sub-action, subsequently integrating two actions using these attentions to generate pseudo-training examples. Then, we employ a conditional generative model, CVAE, to learn a latent space, facilitating the diverse generation. Finally, we propose Decoupling Refinement, which leverages a self-supervised pre-trained model MAE to ensure semantic consistency between the sub-actions and compositional actions. This refinement process involves rendering generated 3D actions into 2D space, decoupling these images into two sub-segments, using the MAE model to restore the complete image from sub-segments, and constraining the recovered images to match images rendered from raw sub-actions. Due to the lack of existing datasets containing both sub-actions and compositional actions, we created two new datasets, named HumanAct-C and UESTC-C, and present a corresponding evaluation metric. Both qualitative and quantitative assessments are conducted to show our efficacy.

Zhihao Zhan, Yuhao Chen, Jiaying Zhou et al.

Vision-Language-Action (VLA) models have demonstrated impressive capabilities in generalized robotic control; however, they remain notoriously brittle to linguistic perturbations. We identify a critical ``modality collapse'' phenomenon where strong visual priors overwhelm sparse linguistic signals, causing agents to overfit to specific instruction phrasings while ignoring the underlying semantic intent. To address this, we propose \textbf{Residual Semantic Steering (RSS)}, a probabilistic framework that disentangles physical affordance from semantic execution. RSS introduces two theoretical innovations: (1) \textbf{Monte Carlo Syntactic Integration}, which approximates the true semantic posterior via dense, LLM-driven distributional expansion, and (2) \textbf{Residual Affordance Steering}, a dual-stream decoding mechanism that explicitly isolates the causal influence of language by subtracting the visual affordance prior. Theoretical analysis suggests that RSS effectively maximizes the mutual information between action and intent while suppressing visual distractors. Empirical results across diverse manipulation benchmarks demonstrate that RSS achieves state-of-the-art robustness, maintaining performance even under adversarial linguistic perturbations.

Tianshui Chen, Jianman Lin, Zhijing Yang et al.

Speech-preserving facial expression manipulation (SPFEM) aims to modify facial emotions while meticulously maintaining the mouth animation associated with spoken content. Current works depend on inaccessible paired training samples for the person, where two aligned frames exhibit the same speech content yet differ in emotional expression, limiting the SPFEM applications in real-world scenarios. In this work, we discover that speakers who convey the same content with different emotions exhibit highly correlated local facial animations in both spatial and temporal spaces, providing valuable supervision for SPFEM. To capitalize on this insight, we propose a novel spatial-temporal coherent correlation learning (STCCL) algorithm, which models the aforementioned correlations as explicit metrics and integrates the metrics to supervise manipulating facial expression and meanwhile better preserving the facial animation of spoken content. To this end, it first learns a spatial coherent correlation metric, ensuring that the visual correlations of adjacent local regions within an image linked to a specific emotion closely resemble those of corresponding regions in an image linked to a different emotion. Simultaneously, it develops a temporal coherent correlation metric, ensuring that the visual correlations of specific regions across adjacent image frames associated with one emotion are similar to those in the corresponding regions of frames associated with another emotion. Recognizing that visual correlations are not uniform across all regions, we have also crafted a correlation-aware adaptive strategy that prioritizes regions that present greater challenges. During SPFEM model training, we construct the spatial-temporal coherent correlation metric between corresponding local regions of the input and output image frames as an additional loss to supervise the generation process.

Guangrun Wang, Changlin Li, Liuchun Yuan et al.

Neural Architecture Search (NAS), aiming at automatically designing neural architectures by machines, has been considered a key step toward automatic machine learning. One notable NAS branch is the weight-sharing NAS, which significantly improves search efficiency and allows NAS algorithms to run on ordinary computers. Despite receiving high expectations, this category of methods suffers from low search effectiveness. By employing a generalization boundedness tool, we demonstrate that the devil behind this drawback is the untrustworthy architecture rating with the oversized search space of the possible architectures. Addressing this problem, we modularize a large search space into blocks with small search spaces and develop a family of models with the distilling neural architecture (DNA) techniques. These proposed models, namely a DNA family, are capable of resolving multiple dilemmas of the weight-sharing NAS, such as scalability, efficiency, and multi-modal compatibility. Our proposed DNA models can rate all architecture candidates, as opposed to previous works that can only access a subsearch space using heuristic algorithms. Moreover, under a certain computational complexity constraint, our method can seek architectures with different depths and widths. Extensive experimental evaluations show that our models achieve state-of-the-art top-1 accuracy of 78.9% and 83.6% on ImageNet for a mobile convolutional network and a small vision transformer, respectively. Additionally, we provide in-depth empirical analysis and insights into neural architecture ratings. Codes available: \url{https://github.com/changlin31/DNA}.

Linsheng Chen, Guangrun Wang, Liuchun Yuan et al.

Neural Radiance Fields (NeRF) have garnered remarkable success in novel view synthesis. Nonetheless, the task of generating high-quality images for novel views persists as a critical challenge. While the existing efforts have exhibited commendable progress, capturing intricate details, enhancing textures, and achieving superior Peak Signal-to-Noise Ratio (PSNR) metrics warrant further focused attention and advancement. In this work, we propose NeRF-VPT, an innovative method for novel view synthesis to address these challenges. Our proposed NeRF-VPT employs a cascading view prompt tuning paradigm, wherein RGB information gained from preceding rendering outcomes serves as instructive visual prompts for subsequent rendering stages, with the aspiration that the prior knowledge embedded in the prompts can facilitate the gradual enhancement of rendered image quality. NeRF-VPT only requires sampling RGB data from previous stage renderings as priors at each training stage, without relying on extra guidance or complex techniques. Thus, our NeRF-VPT is plug-and-play and can be readily integrated into existing methods. By conducting comparative analyses of our NeRF-VPT against several NeRF-based approaches on demanding real-scene benchmarks, such as Realistic Synthetic 360, Real Forward-Facing, Replica dataset, and a user-captured dataset, we substantiate that our NeRF-VPT significantly elevates baseline performance and proficiently generates more high-quality novel view images than all the compared state-of-the-art methods. Furthermore, the cascading learning of NeRF-VPT introduces adaptability to scenarios with sparse inputs, resulting in a significant enhancement of accuracy for sparse-view novel view synthesis. The source code and dataset are available at \url{https://github.com/Freedomcls/NeRF-VPT}.

Zijian He, Renjie Liu, Yihao Wang et al.

World building with 3D scene representations is increasingly important for content creation, simulation, and interactive experiences, yet real workflows are inherently iterative: creators must repeatedly extend an existing scene under user control. Motivated by this research gap, we study 3D scene expansion in a user-centric workflow: starting from a real scene captured by multi-view images, we extend its coverage by inserting an additional view synthesized by a generative model. Unlike simple object editing or style transfer in a fixed scene, the inserted view is often 3D-misaligned with the original reconstruction, introducing geometry shifts, hallucinated content, or view-dependent artifacts that break global multi-view consistency. To address the challenge, we propose SceneExpander, which applies test-time adaptation to a parametric feed-forward 3D reconstruction model with two complementary distillation signals: anchor distillation stabilizes the original scene by distilling geometric cues from the captured views, while inserted-view self-distillation preserves observation-supported predictions yet adapts latent geometry and appearance to accommodate the misaligned inserted view. Experiments on ETH scenes and online data demonstrate improved expansion behavior and reconstruction quality under misalignment.

Zijian Song, Qichang Li, Jiawei Zhou et al.

At its core, robotic manipulation is a problem of vision-to-geometry mapping ($f(v) \rightarrow G$). Physical actions are fundamentally defined by geometric properties like 3D positions and spatial relationships. Consequently, we argue that the foundation for generalizable robotic control should be a vision-geometry backbone, rather than the widely adopted vision-language or video models. Conventional VLA and video-predictive models rely on backbones pretrained on large-scale 2D image-text or temporal pixel data. While effective, their representations are largely shaped by semantic concepts or 2D priors, which do not intrinsically align with the precise 3D geometric nature required for physical manipulation. Driven by this insight, we propose the Vision-Geometry-Action (VGA) model, which directly conditions action generation on pretrained native 3D representations. Specifically, VGA replaces conventional language or video backbones with a pretrained 3D world model, establishing a seamless vision-to-geometry mapping that translates visual inputs directly into physical actions. To further enhance geometric consistency, we introduce a Progressive Volumetric Modulation module and adopt a joint training strategy. Extensive experiments validate the effectiveness of our approach. In simulation benchmarks, VGA outperforms top-tier VLA baselines including $π_{0.5}$ and GeoVLA, demonstrating its superiority in precise manipulation. More importantly, VGA exhibits remarkable zero-shot generalization to unseen viewpoints in real-world deployments, consistently outperforming $π_{0.5}$. These results highlight that operating on native 3D representations-rather than translating through language or 2D video priors-is a highly promising direction for achieving generalizable physical intelligence.

Jiaying Zhou, Zhihao Zhan, Ruifeng Zhai et al.

Vision--Language--Action (VLA) policies have shown strong progress in mapping language instructions and visual observations to robotic actions, yet their reliability degrades in cluttered scenes with distractors. By analyzing failure cases, we find that many errors do not arise from infeasible motions, but from instance-level grounding failures: the policy often produces a plausible grasp trajectory that lands slightly off-target or even on the wrong object instance. To address this issue, we propose TAG (Target-Agnostic Guidance), a simple inference-time guidance mechanism that explicitly reduces distractor- and appearance-induced bias in VLA policies. Inspired by classifier-free guidance (CFG), TAG contrasts policy predictions under the original observation and an object-erased observation, and uses their difference as a residual steering signal that strengthens the influence of object evidence in the decision process. TAG does not require modifying the policy architecture and can be integrated with existing VLA policies with minimal training and inference changes. We evaluate TAG on standard manipulation benchmarks, including LIBERO, LIBERO-Plus, and VLABench, where it consistently improves robustness under clutter and reduces near-miss and wrong-object executions.

Weijia Liufu, Xiaoyu Guo, Ruiyi Chen et al.

Vision-Language-Action (VLA) models remain brittle in long-horizon, contact-rich manipulation because success-only imitation provides little supervision for execution drift, while failed rollouts are often discarded. We introduce RePO-VLA, a recovery-driven policy optimization framework that assigns distinct roles to success, recovery, and failure trajectories. RePO-VLA first applies Recovery-Aware Initialization (RAI), slicing recovery segments and resetting history so corrective actions depend on the current adverse state rather than the preceding failure. It then learns a Progress-Aware Semantic Value Function (PAS-VF), aligning spatiotemporal trajectory features with instructions and successful references. The resulting labels salvage useful failure prefixes via reliability decay, while low-value labels mark drift and terminal breakdowns, teaching differences among nominal, failed, and corrective actions. The data engine turns adverse states into planner-generated or human-collected corrective rollouts, teaching recovery to the success manifold. Value-Conditioned Refinement (VCR) trains the policy to prefer high-progress actions. At deployment, a fixed high value ($v=1.0$) biases actions toward the learned success manifold without online failure detectors or heuristic retries. We introduce FRBench, with standardized error injection and recovery-focused evaluation. Across simulated and real-world bimanual tasks, RePO-VLA improves robustness, raising adversarial success from 20% to 75% on average and up to 80% in scaled real-world trials.

Changlin Li, Guangrun Wang, Bing Wang et al.

Dynamic networks have shown their promising capability in reducing theoretical computation complexity by adapting their architectures to the input during inference. However, their practical runtime usually lags behind the theoretical acceleration due to inefficient sparsity. Here, we explore a hardware-efficient dynamic inference regime, named dynamic weight slicing, which adaptively slice a part of network parameters for inputs with diverse difficulty levels, while keeping parameters stored statically and contiguously in hardware to prevent the extra burden of sparse computation. Based on this scheme, we present dynamic slimmable network (DS-Net) and dynamic slice-able network (DS-Net++) by input-dependently adjusting filter numbers of CNNs and multiple dimensions in both CNNs and transformers, respectively. To ensure sub-network generality and routing fairness, we propose a disentangled two-stage optimization scheme with training techniques such as in-place bootstrapping (IB), multi-view consistency (MvCo) and sandwich gate sparsification (SGS) to train supernet and gate separately. Extensive experiments on 4 datasets and 3 different network architectures demonstrate our method consistently outperforms state-of-the-art static and dynamic model compression methods by a large margin (up to 6.6%). Typically, DS-Net++ achieves 2-4x computation reduction and 1.62x real-world acceleration over MobileNet, ResNet-50 and Vision Transformer, with minimal accuracy drops (0.1-0.3%) on ImageNet. Code release: https://github.com/changlin31/DS-Net

Chenhe Dong, Guangrun Wang, Hang Xu et al.

Pre-trained language models have shown remarkable results on various NLP tasks. Nevertheless, due to their bulky size and slow inference speed, it is hard to deploy them on edge devices. In this paper, we have a critical insight that improving the feed-forward network (FFN) in BERT has a higher gain than improving the multi-head attention (MHA) since the computational cost of FFN is 2$\sim$3 times larger than MHA. Hence, to compact BERT, we are devoted to designing efficient FFN as opposed to previous works that pay attention to MHA. Since FFN comprises a multilayer perceptron (MLP) that is essential in BERT optimization, we further design a thorough search space towards an advanced MLP and perform a coarse-to-fine mechanism to search for an efficient BERT architecture. Moreover, to accelerate searching and enhance model transferability, we employ a novel warm-up knowledge distillation strategy at each search stage. Extensive experiments show our searched EfficientBERT is 6.9$\times$ smaller and 4.4$\times$ faster than BERT$\rm_{BASE}$, and has competitive performances on GLUE and SQuAD Benchmarks. Concretely, EfficientBERT attains a 77.7 average score on GLUE \emph{test}, 0.7 higher than MobileBERT$\rm_{TINY}$, and achieves an 85.3/74.5 F1 score on SQuAD v1.1/v2.0 \emph{dev}, 3.2/2.7 higher than TinyBERT$_4$ even without data augmentation. The code is released at https://github.com/cheneydon/efficient-bert.

Jiefeng Peng, Jiqi Zhang, Changlin Li et al.

Recently proposed neural architecture search (NAS) methods co-train billions of architectures in a supernet and estimate their potential accuracy using the network weights detached from the supernet. However, the ranking correlation between the architectures' predicted accuracy and their actual capability is incorrect, which causes the existing NAS methods' dilemma. We attribute this ranking correlation problem to the supernet training consistency shift, including feature shift and parameter shift. Feature shift is identified as dynamic input distributions of a hidden layer due to random path sampling. The input distribution dynamic affects the loss descent and finally affects architecture ranking. Parameter shift is identified as contradictory parameter updates for a shared layer lay in different paths in different training steps. The rapidly-changing parameter could not preserve architecture ranking. We address these two shifts simultaneously using a nontrivial supernet-Pi model, called Pi-NAS. Specifically, we employ a supernet-Pi model that contains cross-path learning to reduce the feature consistency shift between different paths. Meanwhile, we adopt a novel nontrivial mean teacher containing negative samples to overcome parameter shift and model collision. Furthermore, our Pi-NAS runs in an unsupervised manner, which can search for more transferable architectures. Extensive experiments on ImageNet and a wide range of downstream tasks (e.g., COCO 2017, ADE20K, and Cityscapes) demonstrate the effectiveness and universality of our Pi-NAS compared to supervised NAS. See Codes: https://github.com/Ernie1/Pi-NAS.

Guangrun Wang, Keze Wang, Guangcong Wang et al.

Self-supervised learning (especially contrastive learning) has attracted great interest due to its huge potential in learning discriminative representations in an unsupervised manner. Despite the acknowledged successes, existing contrastive learning methods suffer from very low learning efficiency, e.g., taking about ten times more training epochs than supervised learning for comparable recognition accuracy. In this paper, we reveal two contradictory phenomena in contrastive learning that we call under-clustering and over-clustering problems, which are major obstacles to learning efficiency. Under-clustering means that the model cannot efficiently learn to discover the dissimilarity between inter-class samples when the negative sample pairs for contrastive learning are insufficient to differentiate all the actual object classes. Over-clustering implies that the model cannot efficiently learn features from excessive negative sample pairs, forcing the model to over-cluster samples of the same actual classes into different clusters. To simultaneously overcome these two problems, we propose a novel self-supervised learning framework using a truncated triplet loss. Precisely, we employ a triplet loss tending to maximize the relative distance between the positive pair and negative pairs to address the under-clustering problem; and we construct the negative pair by selecting a negative sample deputy from all negative samples to avoid the over-clustering problem, guaranteed by the Bernoulli Distribution model. We extensively evaluate our framework in several large-scale benchmarks (e.g., ImageNet, SYSU-30k, and COCO). The results demonstrate our model's superiority (e.g., the learning efficiency) over the latest state-of-the-art methods by a clear margin. Codes available at: https://github.com/wanggrun/triplet .

Changlin Li, Guangrun Wang, Bing Wang et al.

Current dynamic networks and dynamic pruning methods have shown their promising capability in reducing theoretical computation complexity. However, dynamic sparse patterns on convolutional filters fail to achieve actual acceleration in real-world implementation, due to the extra burden of indexing, weight-copying, or zero-masking. Here, we explore a dynamic network slimming regime, named Dynamic Slimmable Network (DS-Net), which aims to achieve good hardware-efficiency via dynamically adjusting filter numbers of networks at test time with respect to different inputs, while keeping filters stored statically and contiguously in hardware to prevent the extra burden. Our DS-Net is empowered with the ability of dynamic inference by the proposed double-headed dynamic gate that comprises an attention head and a slimming head to predictively adjust network width with negligible extra computation cost. To ensure generality of each candidate architecture and the fairness of gate, we propose a disentangled two-stage training scheme inspired by one-shot NAS. In the first stage, a novel training technique for weight-sharing networks named In-place Ensemble Bootstrapping is proposed to improve the supernet training efficacy. In the second stage, Sandwich Gate Sparsification is proposed to assist the gate training by identifying easy and hard samples in an online way. Extensive experiments demonstrate our DS-Net consistently outperforms its static counterparts as well as state-of-the-art static and dynamic model compression methods by a large margin (up to 5.9%). Typically, DS-Net achieves 2-4x computation reduction and 1.62x real-world acceleration over ResNet-50 and MobileNet with minimal accuracy drops on ImageNet. Code release: https://github.com/changlin31/DS-Net .

Changlin Li, Tao Tang, Guangrun Wang et al.

A myriad of recent breakthroughs in hand-crafted neural architectures for visual recognition have highlighted the urgent need to explore hybrid architectures consisting of diversified building blocks. Meanwhile, neural architecture search methods are surging with an expectation to reduce human efforts. However, whether NAS methods can efficiently and effectively handle diversified search spaces with disparate candidates (e.g. CNNs and transformers) is still an open question. In this work, we present Block-wisely Self-supervised Neural Architecture Search (BossNAS), an unsupervised NAS method that addresses the problem of inaccurate architecture rating caused by large weight-sharing space and biased supervision in previous methods. More specifically, we factorize the search space into blocks and utilize a novel self-supervised training scheme, named ensemble bootstrapping, to train each block separately before searching them as a whole towards the population center. Additionally, we present HyTra search space, a fabric-like hybrid CNN-transformer search space with searchable down-sampling positions. On this challenging search space, our searched model, BossNet-T, achieves up to 82.5% accuracy on ImageNet, surpassing EfficientNet by 2.4% with comparable compute time. Moreover, our method achieves superior architecture rating accuracy with 0.78 and 0.76 Spearman correlation on the canonical MBConv search space with ImageNet and on NATS-Bench size search space with CIFAR-100, respectively, surpassing state-of-the-art NAS methods. Code: https://github.com/changlin31/BossNAS

Bailin Li, Bowen Wu, Jiang Su et al.

Finding out the computational redundant part of a trained Deep Neural Network (DNN) is the key question that pruning algorithms target on. Many algorithms try to predict model performance of the pruned sub-nets by introducing various evaluation methods. But they are either inaccurate or very complicated for general application. In this work, we present a pruning method called EagleEye, in which a simple yet efficient evaluation component based on adaptive batch normalization is applied to unveil a strong correlation between different pruned DNN structures and their final settled accuracy. This strong correlation allows us to fast spot the pruned candidates with highest potential accuracy without actually fine-tuning them. This module is also general to plug-in and improve some existing pruning algorithms. EagleEye achieves better pruning performance than all of the studied pruning algorithms in our experiments. Concretely, to prune MobileNet V1 and ResNet-50, EagleEye outperforms all compared methods by up to 3.8%. Even in the more challenging experiments of pruning the compact model of MobileNet V1, EagleEye achieves the highest accuracy of 70.9% with an overall 50% operations (FLOPs) pruned. All accuracy results are Top-1 ImageNet classification accuracy. Source code and models are accessible to open-source community https://github.com/anonymous47823493/EagleEye .

Hongjun Wang, Guangrun Wang, Ya Li et al.

The success of DNNs has driven the extensive applications of person re-identification (ReID) into a new era. However, whether ReID inherits the vulnerability of DNNs remains unexplored. To examine the robustness of ReID systems is rather important because the insecurity of ReID systems may cause severe losses, e.g., the criminals may use the adversarial perturbations to cheat the CCTV systems. In this work, we examine the insecurity of current best-performing ReID models by proposing a learning-to-mis-rank formulation to perturb the ranking of the system output. As the cross-dataset transferability is crucial in the ReID domain, we also perform a back-box attack by developing a novel multi-stage network architecture that pyramids the features of different levels to extract general and transferable features for the adversarial perturbations. Our method can control the number of malicious pixels by using differentiable multi-shot sampling. To guarantee the inconspicuousness of the attack, we also propose a new perception loss to achieve better visual quality. Extensive experiments on four of the largest ReID benchmarks (i.e., Market1501 [45], CUHK03 [18], DukeMTMC [33], and MSMT17 [40]) not only show the effectiveness of our method, but also provides directions of the future improvement in the robustness of ReID systems. For example, the accuracy of one of the best-performing ReID systems drops sharply from 91.8% to 1.4% after being attacked by our method. Some attack results are shown in Fig. 1. The code is available at https://github.com/whj363636/Adversarial-attack-on-Person-ReID-With-Deep-Mis-Ranking.

Guangcong Wang, Jianhuang Lai, Wenqi Liang et al.

Most of the existing approaches focus on specific visual tasks while ignoring the relations between them. Estimating task relation sheds light on the learning of high-order semantic concepts, e.g., transfer learning. How to reveal the underlying relations between different visual tasks remains largely unexplored. In this paper, we propose a novel \textbf{L}earnable \textbf{P}arameter \textbf{S}imilarity (\textbf{LPS}) method that learns an effective metric to measure the similarity of second-order semantics hidden in trained models. LPS is achieved by using a second-order neural network to align high-dimensional model parameters and learning second-order similarity in an end-to-end way. In addition, we create a model set called ModelSet500 as a parameter similarity learning benchmark that contains 500 trained models. Extensive experiments on ModelSet500 validate the effectiveness of the proposed method. Code will be released at \url{https://github.com/Wanggcong/learnable-parameter-similarity}.

Guangrun Wang, Guangcong Wang, Xujie Zhang et al.

Person re-identification (Re-ID) benefits greatly from the accurate annotations of existing datasets (e.g., CUHK03 [1] and Market-1501 [2]), which are quite expensive because each image in these datasets has to be assigned with a proper label. In this work, we ease the annotation of Re-ID by replacing the accurate annotation with inaccurate annotation, i.e., we group the images into bags in terms of time and assign a bag-level label for each bag. This greatly reduces the annotation effort and leads to the creation of a large-scale Re-ID benchmark called SYSU-30$k$. The new benchmark contains $30k$ individuals, which is about $20$ times larger than CUHK03 ($1.3k$ individuals) and Market-1501 ($1.5k$ individuals), and $30$ times larger than ImageNet ($1k$ categories). It sums up to 29,606,918 images. Learning a Re-ID model with bag-level annotation is called the weakly supervised Re-ID problem. To solve this problem, we introduce a differentiable graphical model to capture the dependencies from all images in a bag and generate a reliable pseudo label for each person image. The pseudo label is further used to supervise the learning of the Re-ID model. When compared with the fully supervised Re-ID models, our method achieves state-of-the-art performance on SYSU-30$k$ and other datasets. The code, dataset, and pretrained model will be available at \url{https://github.com/wanggrun/SYSU-30k}.

Guangrun Wang, Keze Wang, Liang Lin

This paper presents a novel adaptively connected neural network (ACNet) to improve the traditional convolutional neural networks (CNNs) {in} two aspects. First, ACNet employs a flexible way to switch global and local inference in processing the internal feature representations by adaptively determining the connection status among the feature nodes (e.g., pixels of the feature maps) \footnote{In a computer vision domain, a node refers to a pixel of a feature map{, while} in {the} graph domain, a node denotes a graph node.}. We can show that existing CNNs, the classical multilayer perceptron (MLP), and the recently proposed non-local network (NLN) \cite{nonlocalnn17} are all special cases of ACNet. Second, ACNet is also capable of handling non-Euclidean data. Extensive experimental analyses on {a variety of benchmarks (i.e.,} ImageNet-1k classification, COCO 2017 detection and segmentation, CUHK03 person re-identification, CIFAR analysis, and Cora document categorization) demonstrate that {ACNet} cannot only achieve state-of-the-art performance but also overcome the limitation of the conventional MLP and CNN \footnote{Corresponding author: Liang Lin (linliang@ieee.org)}. The code is available at \url{https://github.com/wanggrun/Adaptively-Connected-Neural-Networks}.

Jing Wu, Jia-Wang Bian, Xinghui Li et al. · bytedance, oxford

We propose GaussCtrl, a text-driven method to edit a 3D scene reconstructed by the 3D Gaussian Splatting (3DGS). Our method first renders a collection of images by using the 3DGS and edits them by using a pre-trained 2D diffusion model (ControlNet) based on the input prompt, which is then used to optimise the 3D model. Our key contribution is multi-view consistent editing, which enables editing all images together instead of iteratively editing one image while updating the 3D model as in previous works. It leads to faster editing as well as higher visual quality. This is achieved by the two terms: (a) depth-conditioned editing that enforces geometric consistency across multi-view images by leveraging naturally consistent depth maps. (b) attention-based latent code alignment that unifies the appearance of edited images by conditioning their editing to several reference views through self and cross-view attention between images' latent representations. Experiments demonstrate that our method achieves faster editing and better visual results than previous state-of-the-art methods.

Weiqi Li, Quande Zhang, Ruifeng Zhai et al.

Vision-language-action (VLA) models achieve strong in-distribution performance but degrade sharply under novel camera viewpoints and visual perturbations. We show that this brittleness primarily arises from misalignment in Spatial Modeling, rather than Physical Modeling. To address this, we propose a one-shot adaptation framework that recalibrates visual representations through lightweight, learnable updates. Our first method, Feature Token Modulation (FTM), applies a global affine transformation to visual tokens and improves Libero viewpoint accuracy from 48.5% to 87.1% with only 4K parameters. Building on this, Feature Linear Adaptation (FLA) introduces low-rank updates to the ViT encoder, achieving 90.8% success with 4.7M parameters -- matching LoRA-scale finetuning at far lower cost. Together, these results reveal substantial untapped robustness in pretrained VLA models and demonstrate that targeted, minimal visual adaptation is sufficient to restore viewpoint generalization.

Weiqi Li, Zehao Zhang, Liang Lin et al.

Controllability is a fundamental requirement in video synthesis, where accurate alignment with conditioning signals is essential. Existing classifier-free guidance methods typically achieve conditioning indirectly by modeling the joint distribution of data and conditions, which often results in limited controllability over the specified conditions. Classifier-based guidance enforces conditions through an external classifier, but the model may exploit this mechanism to raise the classifier score without genuinely satisfying the intended condition, resulting in adversarial artifacts and limited effective controllability. In this paper, we propose Attention-Conditional Diffusion (ACD), a novel framework for direct conditional control in video diffusion models via attention supervision. By aligning the model's attention maps with external control signals, ACD achieves better controllability. To support this, we introduce a sparse 3D-aware object layout as an efficient conditioning signal, along with a dedicated Layout ControlNet and an automated annotation pipeline for scalable layout integration. Extensive experiments on benchmark video generation datasets demonstrate that ACD delivers superior alignment with conditioning inputs while preserving temporal coherence and visual fidelity, establishing an effective paradigm for conditional video synthesis.

Yuanfeng Xu, Yuhao Chen, Liang Lin et al.

The bifurcation of generative modeling into autoregressive approaches for discrete data (text) and diffusion approaches for continuous data (images) hinders the development of truly unified multimodal systems. While Masked Language Models (MLMs) offer efficient bidirectional context, they traditionally lack the generative fidelity of autoregressive models and the semantic continuity of diffusion models. Furthermore, extending masked generation to multimodal settings introduces severe alignment challenges and training instability. In this work, we propose \textbf{CoM-DAD} (\textbf{Co}upled \textbf{M}anifold \textbf{D}iscrete \textbf{A}bsorbing \textbf{D}iffusion), a novel probabilistic framework that reformulates multimodal generation as a hierarchical dual-process. CoM-DAD decouples high-level semantic planning from low-level token synthesis. First, we model the semantic manifold via a continuous latent diffusion process; second, we treat token generation as a discrete absorbing diffusion process, regulated by a \textbf{Variable-Rate Noise Schedule}, conditioned on these evolving semantic priors. Crucially, we introduce a \textbf{Stochastic Mixed-Modal Transport} strategy that aligns disparate modalities without requiring heavy contrastive dual-encoders. Our method demonstrates superior stability over standard masked modeling, establishing a new paradigm for scalable, unified text-image generation.

Hongyu Chen, Liang Lin, Guangrun Wang

Standard Chain-of-Thought (CoT) prompting empowers Large Language Models (LLMs) with reasoning capabilities, yet its reliance on linear natural language is inherently insufficient for effective world modeling in embodied tasks. While text offers flexibility, it fails to explicitly represent the state-space, object hierarchies, and causal dependencies required for robust robotic planning. To address these limitations, we propose Object-Oriented World Modeling (OOWM), a novel framework that structures embodied reasoning through the lens of software engineering formalisms. We redefine the world model not as a latent vector space, but as an explicit symbolic tuple $W = \langle S, T \rangle$: a State Abstraction ($G_\text{state}$) instantiating the environmental state $S$, coupled with a Control Policy ($G_\text{control}$) representing the transition logic $T: S \times A \rightarrow S'$. OOWM leverages the Unified Modeling Language (UML) to materialize this definition: it employs Class Diagrams to ground visual perception into rigorous object hierarchies, and Activity Diagrams to operationalize planning into executable control flows. Furthermore, we introduce a three-stage training pipeline combining Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO). Crucially, this method utilizes outcome-based rewards from the final plan to implicitly optimize the underlying object-oriented reasoning structure, enabling effective learning even with sparse annotations. Extensive evaluations on the MRoom-30k benchmark demonstrate that OOWM significantly outperforms unstructured textual baselines in planning coherence, execution success, and structural fidelity, establishing a new paradigm for structured embodied reasoning.

Tao Tang, Guangrun Wang, Yixing Lao et al.

Neural implicit fields have been a de facto standard in novel view synthesis. Recently, there exist some methods exploring fusing multiple modalities within a single field, aiming to share implicit features from different modalities to enhance reconstruction performance. However, these modalities often exhibit misaligned behaviors: optimizing for one modality, such as LiDAR, can adversely affect another, like camera performance, and vice versa. In this work, we conduct comprehensive analyses on the multimodal implicit field of LiDAR-camera joint synthesis, revealing the underlying issue lies in the misalignment of different sensors. Furthermore, we introduce AlignMiF, a geometrically aligned multimodal implicit field with two proposed modules: Geometry-Aware Alignment (GAA) and Shared Geometry Initialization (SGI). These modules effectively align the coarse geometry across different modalities, significantly enhancing the fusion process between LiDAR and camera data. Through extensive experiments across various datasets and scenes, we demonstrate the effectiveness of our approach in facilitating better interaction between LiDAR and camera modalities within a unified neural field. Specifically, our proposed AlignMiF, achieves remarkable improvement over recent implicit fusion methods (+2.01 and +3.11 image PSNR on the KITTI-360 and Waymo datasets) and consistently surpasses single modality performance (13.8% and 14.2% reduction in LiDAR Chamfer Distance on the respective datasets).

Zijian He, Yuwei Ning, Yipeng Qin et al.

Virtual Try-On (VTON) is a transformative technology in e-commerce and fashion design, enabling realistic digital visualization of clothing on individuals. In this work, we propose VTON 360, a novel 3D VTON method that addresses the open challenge of achieving high-fidelity VTON that supports any-view rendering. Specifically, we leverage the equivalence between a 3D model and its rendered multi-view 2D images, and reformulate 3D VTON as an extension of 2D VTON that ensures 3D consistent results across multiple views. To achieve this, we extend 2D VTON models to include multi-view garments and clothing-agnostic human body images as input, and propose several novel techniques to enhance them, including: i) a pseudo-3D pose representation using normal maps derived from the SMPL-X 3D human model, ii) a multi-view spatial attention mechanism that models the correlations between features from different viewing angles, and iii) a multi-view CLIP embedding that enhances the garment CLIP features used in 2D VTON with camera information. Extensive experiments on large-scale real datasets and clothing images from e-commerce platforms demonstrate the effectiveness of our approach. Project page: https://scnuhealthy.github.io/VTON360.

Kalyan Ramakrishnan, Lars L. Schaaf, Chen Lin et al.

We show that neural networks can be optimized to represent minimum energy paths as continuous functions, offering a flexible alternative to discrete path-search methods such as Nudged Elastic Band (NEB). Our approach parameterizes reaction paths with a network trained on a loss function that discards tangential energy gradients and enables instant estimation of the transition state. We first validate the method on two-dimensional potentials and then demonstrate its advantages over NEB on challenging atomistic systems where (i) poor initial guesses yield unphysical paths, (ii) multiple competing paths exist, or (iii) the reaction follows a complex multi-step mechanism. Results highlight the versatility of the method: for instance, a simple adjustment to the sampling strategy during optimization can help escape local-minimum solutions. Finally, in a low-dimensional setting, we demonstrate that a single neural network can learn from existing paths and generalize to unseen systems, showing promise for a universal reaction path representation.

Zijian Song, Sihan Qin, Tianshui Chen et al.

The scarcity of manipulation data has motivated the use of pretrained large models from other modalities in robotics. In this work, we build upon autoregressive video generation models to propose a Physical Autoregressive Model (PAR), where physical tokens combine frames and actions to represent the joint evolution of the robot and its environment. PAR leverages the world knowledge embedded in video pretraining to understand physical dynamics without requiring action pretraining, enabling accurate video prediction and consistent action trajectories. It also adopts a DiT-based de-tokenizer to model frames and actions as continuous tokens, mitigating quantization errors and facilitating mutual enhancement. Furthermore, we incorporate a causal mask with inverse kinematics, parallel training, and the KV-cache mechanism to further improve performance and efficiency. Experiments on the ManiSkill benchmark show that PAR achieves a 100\% success rate on the PushCube task, matches the performance of action-pretrained baselines on other tasks, and accurately predicts future videos with tightly aligned action trajectories. These findings underscore a promising direction for robotic manipulation by transferring world knowledge from autoregressive video pretraining. The project page is here: https://hcplab-sysu.github.io/PhysicalAutoregressiveModel/

Zijian Song, Xiaoxin Lin, Qiuming Huang et al.

Large Language Models (LLMs) have undergone rapid progress, largely attributed to reinforcement learning on complex reasoning tasks. In contrast, while spatial intelligence is fundamental for Vision-Language Models (VLMs) in real-world interaction, the systematic study of their complex spatial reasoning remains underexplored. To bridge this gap, we introduce SIRI-Bench, a benchmark designed to evaluate VLMs' structural spatial intelligence through spatial-grounded reasoning tasks. SIRI-Bench comprises 9,000 video-question-answer triplets, where each problem is embedded in a realistic 3D scene. The benchmark is carefully designed so that solving each problem requires both spatial comprehension and structural reasoning. To facilitate large-scale data synthesis, we develop an Automatic Scene Creation Engine that employs collaborative LLM agents to translate abstract mathematical problems into faithful 3D scenes. Experimental results reveal that state-of-the-art VLMs struggle significantly on SIRI-Bench, underscoring the challenge of structural spatial reasoning. We hope that our study will bring researchers' attention to spatially grounded reasoning and advance VLMs in visual problem-solving.

Yuanfeng Xu, Yuhao Chen, Zhongzhan Huang et al.

Recent video editing advancements rely on accurate pose sequences to animate subjects. However, these efforts are not suitable for cross-species animation due to pose misalignment between species (for example, the poses of a cat differs greatly from that of a pig due to differences in body structure). In this paper, we present AnimateZoo, a zero-shot diffusion-based video generator to address this challenging cross-species animation issue, aiming to accurately produce animal animations while preserving the background. The key technique used in our AnimateZoo is subject alignment, which includes two steps. First, we improve appearance feature extraction by integrating a Laplacian detail booster and a prompt-tuning identity extractor. These components are specifically designed to capture essential appearance information, including identity and fine details. Second, we align shape features and address conflicts from differing subjects by introducing a scale-information remover. This ensures accurate cross-species animation. Moreover, we introduce two high-quality animal video datasets featuring a wide variety of species. Trained on these extensive datasets, our model is capable of generating videos characterized by accurate movements, consistent appearance, and high-fidelity frames, without the need for the pre-inference fine-tuning that prior arts required. Extensive experiments showcase the outstanding performance of our method in cross-species action following tasks, demonstrating exceptional shape adaptation capability. The project page is available at https://justinxu0.github.io/AnimateZoo/.

Yuhao Chen, Shubin Chen, Liang Lin et al.

Language-driven image segmentation is a fundamental task in vision-language understanding, requiring models to segment regions of an image corresponding to natural language expressions. Traditional methods approach this as a discriminative problem, assigning each pixel to foreground or background based on semantic alignment. Recently, diffusion models have been introduced to this domain, but existing approaches remain image-centric: they either (i) use image diffusion models as visual feature extractors, (ii) synthesize segmentation data via image generation to train discriminative models, or (iii) perform diffusion inversion to extract attention cues from pre-trained image diffusion models-thereby treating segmentation as an auxiliary process. In this paper, we propose GS (Generative Segmentation), a novel framework that formulates segmentation itself as a generative task via label diffusion. Instead of generating images conditioned on label maps and text, GS reverses the generative process: it directly generates segmentation masks from noise, conditioned on both the input image and the accompanying language description. This paradigm makes label generation the primary modeling target, enabling end-to-end training with explicit control over spatial and semantic fidelity. To demonstrate the effectiveness of our approach, we evaluate GS on Panoptic Narrative Grounding (PNG), a representative and challenging benchmark for multimodal segmentation that requires panoptic-level reasoning guided by narrative captions. Experimental results show that GS significantly outperforms existing discriminative and diffusion-based methods, setting a new state-of-the-art for language-driven segmentation.

Zhihao Zhan, Jiaying Zhou, Likui Zhang et al.

Vision-Language-Action (VLA) models offer a unified framework for robotic manipulation by integrating visual perception, language understanding, and control generation. Yet existing VLA models still struggle to generalize across diverse tasks, scenes, and camera viewpoints, and often produce coarse or unstable actions. We introduce E0, a continuized discrete diffusion framework that formulates action generation as iterative denoising over quantized action tokens. Compared with continuous diffusion policies, E0 offers two key advantages: (1) discrete action tokens align naturally with the symbolic structure of pretrained VLM/VLA backbones, enabling stronger semantic conditioning; and 2. discrete diffusion matches the true quantized nature of real-world robot control-whose hardware constraints (e.g., encoder resolution, control frequency, actuation latency) inherently discretize continuous signals-and therefore benefits from a Bayes-optimal denoiser that models the correct discrete action distribution, leading to stronger generalization. Compared with discrete autoregressive and mask-based discrete diffusion models, E0 supports a significantly larger and finer-grained action vocabulary and avoids the distributional mismatch introduced by masking-based corruptions-yielding more accurate fine-grained action control. We further introduce a spherical viewpoint perturbation augmentation method to improve robustness to camera shifts without additional data. Experiments on LIBERO, VLABench, and ManiSkill show that E0 achieves state-of-the-art performance across 14 diverse environments, outperforming strong baselines by 10.7% on average. Real-world evaluation on a Franka arm confirms that E0 delivers precise, robust, and transferable manipulation, establishing discrete diffusion as a promising direction for generalizable VLA policy learning.

Zijian Song, Xiaoxin Lin, Tao Pu et al.

Recent progress in robotics and embodied AI is largely driven by Large Multimodal Models (LMMs). However, a key challenge remains underexplored: how can we advance LMMs to discover tasks that assist humans in open-future scenarios, where human intentions are highly concurrent and dynamic. In this work, we formalize the problem of Human-centric Open-future Task Discovery (HOTD), focusing particularly on identifying tasks that reduce human effort across plausible futures. To facilitate this study, we propose HOTD-Bench, which features over 2K real-world videos, a semi-automated annotation pipeline, and a simulation-based protocol tailored for open-set future evaluation. Additionally, we propose the Collaborative Multi-Agent Search Tree (CMAST) framework, which decomposes complex reasoning through a multi-agent system and structures the reasoning process through a scalable search tree module. In our experiments, CMAST achieves the best performance on the HOTD-Bench, significantly surpassing existing LMMs. It also integrates well with existing LMMs, consistently improving performance.

Xiao Li, Jiaqi Zhang, Shuxiang Zhang et al.

We propose an Authentic Discrete Diffusion (ADD) framework that fundamentally redefines prior pseudo-discrete approaches by preserving core diffusion characteristics directly in the one-hot space through a suite of coordinated mechanisms. Unlike conventional "pseudo" discrete diffusion (PDD) methods, ADD reformulates the diffusion input by directly using float-encoded one-hot class data, without relying on diffusing in the continuous latent spaces or masking policies. At its core, a timestep-conditioned cross-entropy loss is introduced between the diffusion model's outputs and the original one-hot labels. This synergistic design establishes a bridge between discriminative and generative learning. Our experiments demonstrate that ADD not only achieves superior performance on classification tasks compared to the baseline, but also exhibits excellent text generation capabilities on Image captioning. Extensive ablations validate the measurable gains of each component.

Hongyu Chen, Guangrun Wang

Chain-of-Thought (CoT) prompting improves reasoning in large language models (LLMs), but its reliance on unstructured text limits interpretability and executability in embodied tasks. Prior work has explored structured CoTs using scene or logic graphs, yet these remain fundamentally limited: they model only low-order relations, lack constructs like inheritance or behavioral abstraction, and provide no standardized semantics for sequential or conditional planning. We propose UML-CoT, a structured reasoning and planning framework that leverages Unified Modeling Language (UML) to generate symbolic CoTs and executable action plans. UML class diagrams capture compositional object semantics, while activity diagrams model procedural control flow. Our three-stage training pipeline combines supervised fine-tuning with Group Relative Policy Optimization (GRPO), including reward learning from answer-only data. We evaluate UML-CoT on MRoom-30k, a new benchmark of cluttered room-cleaning scenarios. UML-CoT outperforms unstructured CoTs in interpretability, planning coherence, and execution success, highlighting UML as a more expressive and actionable structured reasoning formalism.