Leonidas Guibas

Qiuhong Anna Wei, Sijie Ding, Jeong Joon Park et al. · stanford

Humans universally dislike the task of cleaning up a messy room. If machines were to help us with this task, they must understand human criteria for regular arrangements, such as several types of symmetry, co-linearity or co-circularity, spacing uniformity in linear or circular patterns, and further inter-object relationships that relate to style and functionality. Previous approaches for this task relied on human input to explicitly specify goal state, or synthesized scenes from scratch -- but such methods do not address the rearrangement of existing messy scenes without providing a goal state. In this paper, we present LEGO-Net, a data-driven transformer-based iterative method for LEarning reGular rearrangement of Objects in messy rooms. LEGO-Net is partly inspired by diffusion models -- it starts with an initial messy state and iteratively ''de-noises'' the position and orientation of objects to a regular state while reducing distance traveled. Given randomly perturbed object positions and orientations in an existing dataset of professionally-arranged scenes, our method is trained to recover a regular re-arrangement. Results demonstrate that our method is able to reliably rearrange room scenes and outperform other methods. We additionally propose a metric for evaluating regularity in room arrangements using number-theoretic machinery.

Sherwin Bahmani, Ivan Skorokhodov, Victor Rong et al. · stanford

Recent breakthroughs in text-to-4D generation rely on pre-trained text-to-image and text-to-video models to generate dynamic 3D scenes. However, current text-to-4D methods face a three-way tradeoff between the quality of scene appearance, 3D structure, and motion. For example, text-to-image models and their 3D-aware variants are trained on internet-scale image datasets and can be used to produce scenes with realistic appearance and 3D structure -- but no motion. Text-to-video models are trained on relatively smaller video datasets and can produce scenes with motion, but poorer appearance and 3D structure. While these models have complementary strengths, they also have opposing weaknesses, making it difficult to combine them in a way that alleviates this three-way tradeoff. Here, we introduce hybrid score distillation sampling, an alternating optimization procedure that blends supervision signals from multiple pre-trained diffusion models and incorporates benefits of each for high-fidelity text-to-4D generation. Using hybrid SDS, we demonstrate synthesis of 4D scenes with compelling appearance, 3D structure, and motion.

Arjun R. Akula, Brendan Driscoll, Pradyumna Narayana et al. · deepmind, ibm-research

Creativity is an indispensable part of human cognition and also an inherent part of how we make sense of the world. Metaphorical abstraction is fundamental in communicating creative ideas through nuanced relationships between abstract concepts such as feelings. While computer vision benchmarks and approaches predominantly focus on understanding and generating literal interpretations of images, metaphorical comprehension of images remains relatively unexplored. Towards this goal, we introduce MetaCLUE, a set of vision tasks on visual metaphor. We also collect high-quality and rich metaphor annotations (abstract objects, concepts, relationships along with their corresponding object boxes) as there do not exist any datasets that facilitate the evaluation of these tasks. We perform a comprehensive analysis of state-of-the-art models in vision and language based on our annotations, highlighting strengths and weaknesses of current approaches in visual metaphor Classification, Localization, Understanding (retrieval, question answering, captioning) and gEneration (text-to-image synthesis) tasks. We hope this work provides a concrete step towards developing AI systems with human-like creative capabilities.

Boyang Deng, Richard Tucker, Zhengqi Li et al. · deepmind

We present a method for generating Streetscapes-long sequences of views through an on-the-fly synthesized city-scale scene. Our generation is conditioned by language input (e.g., city name, weather), as well as an underlying map/layout hosting the desired trajectory. Compared to recent models for video generation or 3D view synthesis, our method can scale to much longer-range camera trajectories, spanning several city blocks, while maintaining visual quality and consistency. To achieve this goal, we build on recent work on video diffusion, used within an autoregressive framework that can easily scale to long sequences. In particular, we introduce a new temporal imputation method that prevents our autoregressive approach from drifting from the distribution of realistic city imagery. We train our Streetscapes system on a compelling source of data-posed imagery from Google Street View, along with contextual map data-which allows users to generate city views conditioned on any desired city layout, with controllable camera poses. Please see more results at our project page at https://boyangdeng.com/streetscapes.

Bingfan Zhu, Yanchao Yang, Xulong Wang et al.

We propose VDN-NeRF, a method to train neural radiance fields (NeRFs) for better geometry under non-Lambertian surface and dynamic lighting conditions that cause significant variation in the radiance of a point when viewed from different angles. Instead of explicitly modeling the underlying factors that result in the view-dependent phenomenon, which could be complex yet not inclusive, we develop a simple and effective technique that normalizes the view-dependence by distilling invariant information already encoded in the learned NeRFs. We then jointly train NeRFs for view synthesis with view-dependence normalization to attain quality geometry. Our experiments show that even though shape-radiance ambiguity is inevitable, the proposed normalization can minimize its effect on geometry, which essentially aligns the optimal capacity needed for explaining view-dependent variations. Our method applies to various baselines and significantly improves geometry without changing the volume rendering pipeline, even if the data is captured under a moving light source. Code is available at: https://github.com/BoifZ/VDN-NeRF.

Abhijit Kundu, Kyle Genova, Xiaoqi Yin et al.

We present Panoptic Neural Fields (PNF), an object-aware neural scene representation that decomposes a scene into a set of objects (things) and background (stuff). Each object is represented by an oriented 3D bounding box and a multi-layer perceptron (MLP) that takes position, direction, and time and outputs density and radiance. The background stuff is represented by a similar MLP that additionally outputs semantic labels. Each object MLPs are instance-specific and thus can be smaller and faster than previous object-aware approaches, while still leveraging category-specific priors incorporated via meta-learned initialization. Our model builds a panoptic radiance field representation of any scene from just color images. We use off-the-shelf algorithms to predict camera poses, object tracks, and 2D image semantic segmentations. Then we jointly optimize the MLP weights and bounding box parameters using analysis-by-synthesis with self-supervision from color images and pseudo-supervision from predicted semantic segmentations. During experiments with real-world dynamic scenes, we find that our model can be used effectively for several tasks like novel view synthesis, 2D panoptic segmentation, 3D scene editing, and multiview depth prediction.

Yajie Bao, Yang Li, Shao-Lun Huang et al.

Task transfer learning is a popular technique in image processing applications that uses pre-trained models to reduce the supervision cost of related tasks. An important question is to determine task transferability, i.e. given a common input domain, estimating to what extent representations learned from a source task can help in learning a target task. Typically, transferability is either measured experimentally or inferred through task relatedness, which is often defined without a clear operational meaning. In this paper, we present a novel metric, H-score, an easily-computable evaluation function that estimates the performance of transferred representations from one task to another in classification problems using statistical and information theoretic principles. Experiments on real image data show that our metric is not only consistent with the empirical transferability measurement, but also useful to practitioners in applications such as source model selection and task transfer curriculum learning.

Jihyeon Je, Jiayi Liu, Guandao Yang et al. · stanford

Symmetries are ubiquitous across all kinds of objects, whether in nature or in man-made creations. While these symmetries may seem intuitive to the human eye, detecting them with a machine is nontrivial due to the vast search space. Classical geometry-based methods work by aggregating "votes" for each symmetry but struggle with noise. In contrast, learning-based methods may be more robust to noise, but often overlook partial symmetries due to the scarcity of annotated data. In this work, we address this challenge by proposing a novel symmetry detection method that marries classical symmetry detection techniques with recent advances in generative modeling. Specifically, we apply Langevin dynamics to a redefined symmetry space to enhance robustness against noise. We provide empirical results on a variety of shapes that suggest our method is not only robust to noise, but can also identify both partial and global symmetries. Moreover, we demonstrate the utility of our detected symmetries in various downstream tasks, such as compression and symmetrization of noisy shapes.

Congyue Deng, Chiyu "Max'' Jiang, Charles R. Qi et al.

2D-to-3D reconstruction is an ill-posed problem, yet humans are good at solving this problem due to their prior knowledge of the 3D world developed over years. Driven by this observation, we propose NeRDi, a single-view NeRF synthesis framework with general image priors from 2D diffusion models. Formulating single-view reconstruction as an image-conditioned 3D generation problem, we optimize the NeRF representations by minimizing a diffusion loss on its arbitrary view renderings with a pretrained image diffusion model under the input-view constraint. We leverage off-the-shelf vision-language models and introduce a two-section language guidance as conditioning inputs to the diffusion model. This is essentially helpful for improving multiview content coherence as it narrows down the general image prior conditioned on the semantic and visual features of the single-view input image. Additionally, we introduce a geometric loss based on estimated depth maps to regularize the underlying 3D geometry of the NeRF. Experimental results on the DTU MVS dataset show that our method can synthesize novel views with higher quality even compared to existing methods trained on this dataset. We also demonstrate our generalizability in zero-shot NeRF synthesis for in-the-wild images.

Rundong Luo, Haoran Geng, Congyue Deng et al. · berkeley

Interactable objects are ubiquitous in our daily lives. Recent advances in 3D generative models make it possible to automate the modeling of these objects, benefiting a range of applications from 3D printing to the creation of robot simulation environments. However, while significant progress has been made in modeling 3D shapes and appearances, modeling object physics, particularly for interactable objects, remains challenging due to the physical constraints imposed by inter-part motions. In this paper, we tackle the problem of physically plausible part completion for interactable objects, aiming to generate 3D parts that not only fit precisely into the object but also allow smooth part motions. To this end, we propose a diffusion-based part generation model that utilizes geometric conditioning through classifier-free guidance and formulates physical constraints as a set of stability and mobility losses to guide the sampling process. Additionally, we demonstrate the generation of dependent parts, paving the way toward sequential part generation for objects with complex part-whole hierarchies. Experimentally, we introduce a new metric for measuring physical plausibility based on motion success rates. Our model outperforms existing baselines over shape and physical metrics, especially those that do not adequately model physical constraints. We also demonstrate our applications in 3D printing, robot manipulation, and sequential part generation, showing our strength in realistic tasks with the demand for high physical plausibility.

Yuxuan Kuang, Junjie Ye, Haoran Geng et al. · berkeley, pku

This work proposes a retrieve-and-transfer framework for zero-shot robotic manipulation, dubbed RAM, featuring generalizability across various objects, environments, and embodiments. Unlike existing approaches that learn manipulation from expensive in-domain demonstrations, RAM capitalizes on a retrieval-based affordance transfer paradigm to acquire versatile manipulation capabilities from abundant out-of-domain data. First, RAM extracts unified affordance at scale from diverse sources of demonstrations including robotic data, human-object interaction (HOI) data, and custom data to construct a comprehensive affordance memory. Then given a language instruction, RAM hierarchically retrieves the most similar demonstration from the affordance memory and transfers such out-of-domain 2D affordance to in-domain 3D executable affordance in a zero-shot and embodiment-agnostic manner. Extensive simulation and real-world evaluations demonstrate that our RAM consistently outperforms existing works in diverse daily tasks. Additionally, RAM shows significant potential for downstream applications such as automatic and efficient data collection, one-shot visual imitation, and LLM/VLM-integrated long-horizon manipulation. For more details, please check our website at https://yxkryptonite.github.io/RAM/.

Xiaoshuai Zhang, Zhicheng Wang, Howard Zhou et al. · cmu, deepmind

To advance the state of the art in the creation of 3D foundation models, this paper introduces the ConDense framework for 3D pre-training utilizing existing pre-trained 2D networks and large-scale multi-view datasets. We propose a novel 2D-3D joint training scheme to extract co-embedded 2D and 3D features in an end-to-end pipeline, where 2D-3D feature consistency is enforced through a volume rendering NeRF-like ray marching process. Using dense per pixel features we are able to 1) directly distill the learned priors from 2D models to 3D models and create useful 3D backbones, 2) extract more consistent and less noisy 2D features, 3) formulate a consistent embedding space where 2D, 3D, and other modalities of data (e.g., natural language prompts) can be jointly queried. Furthermore, besides dense features, ConDense can be trained to extract sparse features (e.g., key points), also with 2D-3D consistency -- condensing 3D NeRF representations into compact sets of decorated key points. We demonstrate that our pre-trained model provides good initialization for various 3D tasks including 3D classification and segmentation, outperforming other 3D pre-training methods by a significant margin. It also enables, by exploiting our sparse features, additional useful downstream tasks, such as matching 2D images to 3D scenes, detecting duplicate 3D scenes, and querying a repository of 3D scenes through natural language -- all quite efficiently and without any per-scene fine-tuning.

Xiaoshuai Zhang, Abhijit Kundu, Thomas Funkhouser et al.

We address efficient and structure-aware 3D scene representation from images. Nerflets are our key contribution -- a set of local neural radiance fields that together represent a scene. Each nerflet maintains its own spatial position, orientation, and extent, within which it contributes to panoptic, density, and radiance reconstructions. By leveraging only photometric and inferred panoptic image supervision, we can directly and jointly optimize the parameters of a set of nerflets so as to form a decomposed representation of the scene, where each object instance is represented by a group of nerflets. During experiments with indoor and outdoor environments, we find that nerflets: (1) fit and approximate the scene more efficiently than traditional global NeRFs, (2) allow the extraction of panoptic and photometric renderings from arbitrary views, and (3) enable tasks rare for NeRFs, such as 3D panoptic segmentation and interactive editing.

Nilesh Kulkarni, Davis Rempe, Kyle Genova et al.

We address the problem of generating realistic 3D motions of humans interacting with objects in a scene. Our key idea is to create a neural interaction field attached to a specific object, which outputs the distance to the valid interaction manifold given a human pose as input. This interaction field guides the sampling of an object-conditioned human motion diffusion model, so as to encourage plausible contacts and affordance semantics. To support interactions with scarcely available data, we propose an automated synthetic data pipeline. For this, we seed a pre-trained motion model, which has priors for the basics of human movement, with interaction-specific anchor poses extracted from limited motion capture data. Using our guided diffusion model trained on generated synthetic data, we synthesize realistic motions for sitting and lifting with several objects, outperforming alternative approaches in terms of motion quality and successful action completion. We call our framework NIFTY: Neural Interaction Fields for Trajectory sYnthesis.

Sherwin Bahmani, Jeong Joon Park, Despoina Paschalidou et al.

In this work, we introduce CC3D, a conditional generative model that synthesizes complex 3D scenes conditioned on 2D semantic scene layouts, trained using single-view images. Different from most existing 3D GANs that limit their applicability to aligned single objects, we focus on generating complex scenes with multiple objects, by modeling the compositional nature of 3D scenes. By devising a 2D layout-based approach for 3D synthesis and implementing a new 3D field representation with a stronger geometric inductive bias, we have created a 3D GAN that is both efficient and of high quality, while allowing for a more controllable generation process. Our evaluations on synthetic 3D-FRONT and real-world KITTI-360 datasets demonstrate that our model generates scenes of improved visual and geometric quality in comparison to previous works.

Panos Achlioptas, Maks Ovsjanikov, Leonidas Guibas et al.

In this work, we explore the emotional reactions that real-world images tend to induce by using natural language as the medium to express the rationale behind an affective response to a given visual stimulus. To embark on this journey, we introduce and share with the research community a large-scale dataset that contains emotional reactions and free-form textual explanations for 85,007 publicly available images, analyzed by 6,283 annotators who were asked to indicate and explain how and why they felt in a particular way when observing a specific image, producing a total of 526,749 responses. Even though emotional reactions are subjective and sensitive to context (personal mood, social status, past experiences) - we show that there is significant common ground to capture potentially plausible emotional responses with a large support in the subject population. In light of this crucial observation, we ask the following questions: i) Can we develop multi-modal neural networks that provide reasonable affective responses to real-world visual data, explained with language? ii) Can we steer such methods towards producing explanations with varying degrees of pragmatic language or justifying different emotional reactions while adapting to the underlying visual stimulus? Finally, iii) How can we evaluate the performance of such methods for this novel task? With this work, we take the first steps in addressing all of these questions, thus paving the way for richer, more human-centric, and emotionally-aware image analysis systems. Our introduced dataset and all developed methods are available on https://affective-explanations.org

Mikaela Angelina Uy, Ricardo Martin-Brualla, Leonidas Guibas et al.

Neural radiance fields (NeRFs) have enabled high fidelity 3D reconstruction from multiple 2D input views. However, a well-known drawback of NeRFs is the less-than-ideal performance under a small number of views, due to insufficient constraints enforced by volumetric rendering. To address this issue, we introduce SCADE, a novel technique that improves NeRF reconstruction quality on sparse, unconstrained input views for in-the-wild indoor scenes. To constrain NeRF reconstruction, we leverage geometric priors in the form of per-view depth estimates produced with state-of-the-art monocular depth estimation models, which can generalize across scenes. A key challenge is that monocular depth estimation is an ill-posed problem, with inherent ambiguities. To handle this issue, we propose a new method that learns to predict, for each view, a continuous, multimodal distribution of depth estimates using conditional Implicit Maximum Likelihood Estimation (cIMLE). In order to disambiguate exploiting multiple views, we introduce an original space carving loss that guides the NeRF representation to fuse multiple hypothesized depth maps from each view and distill from them a common geometry that is consistent with all views. Experiments show that our approach enables higher fidelity novel view synthesis from sparse views. Our project page can be found at https://scade-spacecarving-nerfs.github.io .

Sherwin Bahmani, Jeong Joon Park, Despoina Paschalidou et al.

Generative models have emerged as an essential building block for many image synthesis and editing tasks. Recent advances in this field have also enabled high-quality 3D or video content to be generated that exhibits either multi-view or temporal consistency. With our work, we explore 4D generative adversarial networks (GANs) that learn unconditional generation of 3D-aware videos. By combining neural implicit representations with time-aware discriminator, we develop a GAN framework that synthesizes 3D video supervised only with monocular videos. We show that our method learns a rich embedding of decomposable 3D structures and motions that enables new visual effects of spatio-temporal renderings while producing imagery with quality comparable to that of existing 3D or video GANs.

Konstantinos Tertikas, Despoina Paschalidou, Boxiao Pan et al.

Impressive progress in generative models and implicit representations gave rise to methods that can generate 3D shapes of high quality. However, being able to locally control and edit shapes is another essential property that can unlock several content creation applications. Local control can be achieved with part-aware models, but existing methods require 3D supervision and cannot produce textures. In this work, we devise PartNeRF, a novel part-aware generative model for editable 3D shape synthesis that does not require any explicit 3D supervision. Our model generates objects as a set of locally defined NeRFs, augmented with an affine transformation. This enables several editing operations such as applying transformations on parts, mixing parts from different objects etc. To ensure distinct, manipulable parts we enforce a hard assignment of rays to parts that makes sure that the color of each ray is only determined by a single NeRF. As a result, altering one part does not affect the appearance of the others. Evaluations on various ShapeNet categories demonstrate the ability of our model to generate editable 3D objects of improved fidelity, compared to previous part-based generative approaches that require 3D supervision or models relying on NeRFs.

Yang You, Mikaela Angelina Uy, Jiaqi Han et al.

Reverse engineering 3D computer-aided design (CAD) models from images is an important task for many downstream applications including interactive editing, manufacturing, architecture, robotics, etc. The difficulty of the task lies in vast representational disparities between the CAD output and the image input. CAD models are precise, programmatic constructs that involve sequential operations combining discrete command structure with continuous attributes, making it challenging to learn and optimize in an end-to-end fashion. Concurrently, input images introduce inherent challenges such as photometric variability and sensor noise, complicating the reverse engineering process. In this work, we introduce a novel approach that conditionally factorizes the task into two sub-problems. First, we leverage vision-language foundation models (VLMs), a finetuned Llama3.2, to predict the global discrete base structure with semantic information. Second, we propose TrAssembler that, conditioned on the discrete structure with semantics, predicts the continuous attribute values. To support the training of our TrAssembler, we further constructed an annotated CAD dataset of common objects from ShapeNet. Putting all together, our approach and data demonstrate significant first steps towards CAD-ifying images in the wild. Code and data can be found in https://github.com/qq456cvb/Img2CAD.

Zhen Wang, Shijie Zhou, Jeong Joon Park et al.

This work introduces alternating latent topologies (ALTO) for high-fidelity reconstruction of implicit 3D surfaces from noisy point clouds. Previous work identifies that the spatial arrangement of latent encodings is important to recover detail. One school of thought is to encode a latent vector for each point (point latents). Another school of thought is to project point latents into a grid (grid latents) which could be a voxel grid or triplane grid. Each school of thought has tradeoffs. Grid latents are coarse and lose high-frequency detail. In contrast, point latents preserve detail. However, point latents are more difficult to decode into a surface, and quality and runtime suffer. In this paper, we propose ALTO to sequentially alternate between geometric representations, before converging to an easy-to-decode latent. We find that this preserves spatial expressiveness and makes decoding lightweight. We validate ALTO on implicit 3D recovery and observe not only a performance improvement over the state-of-the-art, but a runtime improvement of 3-10$\times$. Project website at https://visual.ee.ucla.edu/alto.htm/.

Bokui Shen, Xinchen Yan, Charles R. Qi et al.

Modeling the 3D world from sensor data for simulation is a scalable way of developing testing and validation environments for robotic learning problems such as autonomous driving. However, manually creating or re-creating real-world-like environments is difficult, expensive, and not scalable. Recent generative model techniques have shown promising progress to address such challenges by learning 3D assets using only plentiful 2D images -- but still suffer limitations as they leverage either human-curated image datasets or renderings from manually-created synthetic 3D environments. In this paper, we introduce GINA-3D, a generative model that uses real-world driving data from camera and LiDAR sensors to create realistic 3D implicit neural assets of diverse vehicles and pedestrians. Compared to the existing image datasets, the real-world driving setting poses new challenges due to occlusions, lighting-variations and long-tail distributions. GINA-3D tackles these challenges by decoupling representation learning and generative modeling into two stages with a learned tri-plane latent structure, inspired by recent advances in generative modeling of images. To evaluate our approach, we construct a large-scale object-centric dataset containing over 1.2M images of vehicles and pedestrians from the Waymo Open Dataset, and a new set of 80K images of long-tail instances such as construction equipment, garbage trucks, and cable cars. We compare our model with existing approaches and demonstrate that it achieves state-of-the-art performance in quality and diversity for both generated images and geometries.

Jiahui Lei, Congyue Deng, Karl Schmeckpeper et al.

We introduce Equivariant Neural Field Expectation Maximization (EFEM), a simple, effective, and robust geometric algorithm that can segment objects in 3D scenes without annotations or training on scenes. We achieve such unsupervised segmentation by exploiting single object shape priors. We make two novel steps in that direction. First, we introduce equivariant shape representations to this problem to eliminate the complexity induced by the variation in object configuration. Second, we propose a novel EM algorithm that can iteratively refine segmentation masks using the equivariant shape prior. We collect a novel real dataset Chairs and Mugs that contains various object configurations and novel scenes in order to verify the effectiveness and robustness of our method. Experimental results demonstrate that our method achieves consistent and robust performance across different scenes where the (weakly) supervised methods may fail. Code and data available at https://www.cis.upenn.edu/~leijh/projects/efem

Yichen Li, Kaichun Mo, Yueqi Duan et al.

Shape assembly composes complex shapes geometries by arranging simple part geometries and has wide applications in autonomous robotic assembly and CAD modeling. Existing works focus on geometry reasoning and neglect the actual physical assembly process of matching and fitting joints, which are the contact surfaces connecting different parts. In this paper, we consider contacting joints for the task of multi-part assembly. A successful joint-optimized assembly needs to satisfy the bilateral objectives of shape structure and joint alignment. We propose a hierarchical graph learning approach composed of two levels of graph representation learning. The part graph takes part geometries as input to build the desired shape structure. The joint-level graph uses part joints information and focuses on matching and aligning joints. The two kinds of information are combined to achieve the bilateral objectives. Extensive experiments demonstrate that our method outperforms previous methods, achieving better shape structure and higher joint alignment accuracy.

Connor Z. Lin, Niloy J. Mitra, Gordon Wetzstein et al.

Neural representations are popular for representing shapes, as they can be learned form sensor data and used for data cleanup, model completion, shape editing, and shape synthesis. Current neural representations can be categorized as either overfitting to a single object instance, or representing a collection of objects. However, neither allows accurate editing of neural scene representations: on the one hand, methods that overfit objects achieve highly accurate reconstructions, but do not generalize to unseen object configurations and thus cannot support editing; on the other hand, methods that represent a family of objects with variations do generalize but produce only approximate reconstructions. We propose NEUFORM to combine the advantages of both overfitted and generalizable representations by adaptively using the one most appropriate for each shape region: the overfitted representation where reliable data is available, and the generalizable representation everywhere else. We achieve this with a carefully designed architecture and an approach that blends the network weights of the two representations, avoiding seams and other artifacts. We demonstrate edits that successfully reconfigure parts of human-designed shapes, such as chairs, tables, and lamps, while preserving semantic integrity and the accuracy of an overfitted shape representation. We compare with two state-of-the-art competitors and demonstrate clear improvements in terms of plausibility and fidelity of the resultant edits.

Kangrui Wang, Linjie Li, Zhengyuan Yang et al.

Can VLMs predict how each camera move changes the view, and plan many such moves ahead? We call this capability view planning, requiring (1)understanding how a single action transforms the view, and (2)composing many such transformations across multi-turn plans to identify a target view. We probe both abilities in our proposed ViewSuite, a 3D point-cloud environment on real ScanNet scenes. Across 13 frontier VLMs, a critical planning gap emerges: they possess basic view-action knowledge but fail to compose it across multi-turn plans, with the gap widening as viewpoint distance grows. To close this gap, we propose an iterative framework that alternates self-exploration with view graph distillation. The key insight is that all exploration trajectories, regardless of their outcome, collectively form a view graph that compactly captures how viewpoints connect across a scene. Distilling this graph into diverse supervised tasks reshapes the policy distribution and overcomes the sparse rewards that stall pure RL. This improves Qwen2.5-VL-7B from 2.5% to 47.8% on interactive view planning, surpassing GPT-5.4 Pro (18.5%) and Gemini 3.1 Pro (21.4%). Self-exploration emerges as a promising path toward VLMs that can actively reason and plan in 3D space.

Jingyun Yang, Congyue Deng, Jimmy Wu et al.

If a robot masters folding a kitchen towel, we would expect it to master folding a large beach towel. However, existing policy learning methods that rely on data augmentation still don't guarantee such generalization. Our insight is to add equivariance to both the visual object representation and policy architecture. We propose EquivAct which utilizes SIM(3)-equivariant network structures that guarantee generalization across all possible object translations, 3D rotations, and scales by construction. EquivAct is trained in two phases. We first pre-train a SIM(3)-equivariant visual representation on simulated scene point clouds. Then, we learn a SIM(3)-equivariant visuomotor policy using a small amount of source task demonstrations. We show that the learned policy directly transfers to objects that substantially differ from demonstrations in scale, position, and orientation. We evaluate our method in three manipulation tasks involving deformable and articulated objects, going beyond typical rigid object manipulation tasks considered in prior work. We conduct experiments both in simulation and in reality. For real robot experiments, our method uses 20 human demonstrations of a tabletop task and transfers zero-shot to a mobile manipulation task in a much larger setup. Experiments confirm that our contrastive pre-training procedure and equivariant architecture offer significant improvements over prior work. Project website: https://equivact.github.io

Minjung Son, Jeong Joon Park, Leonidas Guibas et al.

Generative models have shown great promise in synthesizing photorealistic 3D objects, but they require large amounts of training data. We introduce SinGRAF, a 3D-aware generative model that is trained with a few input images of a single scene. Once trained, SinGRAF generates different realizations of this 3D scene that preserve the appearance of the input while varying scene layout. For this purpose, we build on recent progress in 3D GAN architectures and introduce a novel progressive-scale patch discrimination approach during training. With several experiments, we demonstrate that the results produced by SinGRAF outperform the closest related works in both quality and diversity by a large margin.

Xiaomeng Xu, Yanchao Yang, Kaichun Mo et al.

We propose a method that trains a neural radiance field (NeRF) to encode not only the appearance of the scene but also semantic correlations between scene points, regions, or entities -- aiming to capture their mutual co-variation patterns. In contrast to the traditional first-order photometric reconstruction objective, our method explicitly regularizes the learning dynamics to align the Jacobians of highly-correlated entities, which proves to maximize the mutual information between them under random scene perturbations. By paying attention to this second-order information, we can shape a NeRF to express semantically meaningful synergies when the network weights are changed by a delta along the gradient of a single entity, region, or even a point. To demonstrate the merit of this mutual information modeling, we leverage the coordinated behavior of scene entities that emerges from our shaping to perform label propagation for semantic and instance segmentation. Our experiments show that a JacobiNeRF is more efficient in propagating annotations among 2D pixels and 3D points compared to NeRFs without mutual information shaping, especially in extremely sparse label regimes -- thus reducing annotation burden. The same machinery can further be used for entity selection or scene modifications.

Nikolaos Zygouras, Nikolaos Panagiotou, Yang Li et al.

Travel time estimation is a critical task, useful to many urban applications at the individual citizen and the stakeholder level. This paper presents a novel hybrid algorithm for travel time estimation that leverages historical and sparse real-time trajectory data. Given a path and a departure time we estimate the travel time taking into account the historical information, the real-time trajectory data and the correlations among different road segments. We detect similar road segments using historical trajectories, and use a latent representation to model the similarities. Our experimental evaluation demonstrates the effectiveness of our approach.

Ian Huang, Panos Achlioptas, Tianyi Zhang et al.

Natural language interaction is a promising direction for democratizing 3D shape design. However, existing methods for text-driven 3D shape editing face challenges in producing decoupled, local edits to 3D shapes. We address this problem by learning disentangled latent representations that ground language in 3D geometry. To this end, we propose a complementary tool set including a novel network architecture, a disentanglement loss, and a new editing procedure. Additionally, to measure edit locality, we define a new metric that we call part-wise edit precision. We show that our method outperforms existing SOTA methods by 20% in terms of edit locality, and up to 6.6% in terms of language reference resolution accuracy. Our work suggests that by solely disentangling language representations, downstream 3D shape editing can become more local to relevant parts, even if the model was never given explicit part-based supervision.

Ian Huang, Vrishab Krishna, Omoruyi Atekha et al.

What constitutes the "vibe" of a particular scene? What should one find in "a busy, dirty city street", "an idyllic countryside", or "a crime scene in an abandoned living room"? The translation from abstract scene descriptions to stylized scene elements cannot be done with any generality by extant systems trained on rigid and limited indoor datasets. In this paper, we propose to leverage the knowledge captured by foundation models to accomplish this translation. We present a system that can serve as a tool to generate stylized assets for 3D scenes described by a short phrase, without the need to enumerate the objects to be found within the scene or give instructions on their appearance. Additionally, it is robust to open-world concepts in a way that traditional methods trained on limited data are not, affording more creative freedom to the 3D artist. Our system demonstrates this using a foundation model "team" composed of a large language model, a vision-language model and several image diffusion models, which communicate using an interpretable and user-editable intermediate representation, thus allowing for more versatile and controllable stylized asset generation for 3D artists. We introduce novel metrics for this task, and show through human evaluations that in 91% of the cases, our system outputs are judged more faithful to the semantics of the input scene description than the baseline, thus highlighting the potential of this approach to radically accelerate the 3D content creation process for 3D artists.

Yang You, Bokui Shen, Congyue Deng et al. · berkeley

Deformable object manipulation stands as one of the most captivating yet formidable challenges in robotics. While previous techniques have predominantly relied on learning latent dynamics through demonstrations, typically represented as either particles or images, there exists a pertinent limitation: acquiring suitable demonstrations, especially for long-horizon tasks, can be elusive. Moreover, basing learning entirely on demonstrations can hamper the model's ability to generalize beyond the demonstrated tasks. In this work, we introduce a demonstration-free hierarchical planning approach capable of tackling intricate long-horizon tasks without necessitating any training. We employ large language models (LLMs) to articulate a high-level, stage-by-stage plan corresponding to a specified task. For every individual stage, the LLM provides both the tool's name and the Python code to craft intermediate subgoal point clouds. With the tool and subgoal for a particular stage at our disposal, we present a granular closed-loop model predictive control strategy. This leverages Differentiable Physics with Point-to-Point correspondence (DiffPhysics-P2P) loss in the earth mover distance (EMD) space, applied iteratively. Experimental findings affirm that our technique surpasses multiple benchmarks in dough manipulation, spanning both short and long horizons. Remarkably, our model demonstrates robust generalization capabilities to novel and previously unencountered complex tasks without any preliminary demonstrations. We further substantiate our approach with experimental trials on real-world robotic platforms. Our project page: https://qq456cvb.github.io/projects/donut.

Ivica Kičić, Pantelis R. Vlachas, Georgios Arampatzis et al.

Predictive simulations are essential for applications ranging from weather forecasting to material design. The veracity of these simulations hinges on their capacity to capture the effective system dynamics. Massively parallel simulations predict the systems dynamics by resolving all spatiotemporal scales, often at a cost that prevents experimentation. On the other hand, reduced order models are fast but often limited by the linearization of the system dynamics and the adopted heuristic closures. We propose a novel systematic framework that bridges large scale simulations and reduced order models to extract and forecast adaptively the effective dynamics (AdaLED) of multiscale systems. AdaLED employs an autoencoder to identify reduced-order representations of the system dynamics and an ensemble of probabilistic recurrent neural networks (RNNs) as the latent time-stepper. The framework alternates between the computational solver and the surrogate, accelerating learned dynamics while leaving yet-to-be-learned dynamics regimes to the original solver. AdaLED continuously adapts the surrogate to the new dynamics through online training. The transitions between the surrogate and the computational solver are determined by monitoring the prediction accuracy and uncertainty of the surrogate. The effectiveness of AdaLED is demonstrated on three different systems - a Van der Pol oscillator, a 2D reaction-diffusion equation, and a 2D Navier-Stokes flow past a cylinder for varying Reynolds numbers (400 up to 1200), showcasing its ability to learn effective dynamics online, detect unseen dynamics regimes, and provide net speed-ups. To the best of our knowledge, AdaLED is the first framework that couples a surrogate model with a computational solver to achieve online adaptive learning of effective dynamics. It constitutes a potent tool for applications requiring many expensive simulations.

William Yicheng Zhu, Keren Ye, Junjie Ke et al.

Recognizing and disentangling visual attributes from objects is a foundation to many computer vision applications. While large vision language representations like CLIP had largely resolved the task of zero-shot object recognition, zero-shot visual attribute recognition remains a challenge because CLIP's contrastively-learned vision-language representation cannot effectively capture object-attribute dependencies. In this paper, we target this weakness and propose a sentence generation-based retrieval formulation for attribute recognition that is novel in 1) explicitly modeling a to-be-measured and retrieved object-attribute relation as a conditional probability graph, which converts the recognition problem into a dependency-sensitive language-modeling problem, and 2) applying a large pretrained Vision-Language Model (VLM) on this reformulation and naturally distilling its knowledge of image-object-attribute relations to use towards attribute recognition. Specifically, for each attribute to be recognized on an image, we measure the visual-conditioned probability of generating a short sentence encoding the attribute's relation to objects on the image. Unlike contrastive retrieval, which measures likelihood by globally aligning elements of the sentence to the image, generative retrieval is sensitive to the order and dependency of objects and attributes in the sentence. We demonstrate through experiments that generative retrieval consistently outperforms contrastive retrieval on two visual reasoning datasets, Visual Attribute in the Wild (VAW), and our newly-proposed Visual Genome Attribute Ranking (VGARank).

Sidhika Balachandar, Adrien Poulenard, Congyue Deng et al.

Equivariant networks have been adopted in many 3-D learning areas. Here we identify a fundamental limitation of these networks: their ambiguity to symmetries. Equivariant networks cannot complete symmetry-dependent tasks like segmenting a left-right symmetric object into its left and right sides. We tackle this problem by adding components that resolve symmetry ambiguities while preserving rotational equivariance. We present OAVNN: Orientation Aware Vector Neuron Network, an extension of the Vector Neuron Network. OAVNN is a rotation equivariant network that is robust to planar symmetric inputs. Our network consists of three key components. 1) We introduce an algorithm to calculate symmetry detecting features. 2) We create a symmetry-sensitive orientation aware linear layer. 3) We construct an attention mechanism that relates directional information across points. We evaluate the network using left-right segmentation and find that the network quickly obtains accurate segmentations. We hope this work motivates investigations on the expressivity of equivariant networks on symmetric objects.

Wen-Hsuan Chu, Adam W. Harley, Pavel Tokmakov et al.

Object tracking is central to robot perception and scene understanding. Tracking-by-detection has long been a dominant paradigm for object tracking of specific object categories. Recently, large-scale pre-trained models have shown promising advances in detecting and segmenting objects and parts in 2D static images in the wild. This begs the question: can we re-purpose these large-scale pre-trained static image models for open-vocabulary video tracking? In this paper, we re-purpose an open-vocabulary detector, segmenter, and dense optical flow estimator, into a model that tracks and segments objects of any category in 2D videos. Our method predicts object and part tracks with associated language descriptions in monocular videos, rebuilding the pipeline of Tractor with modern large pre-trained models for static image detection and segmentation: we detect open-vocabulary object instances and propagate their boxes from frame to frame using a flow-based motion model, refine the propagated boxes with the box regression module of the visual detector, and prompt an open-world segmenter with the refined box to segment the objects. We decide the termination of an object track based on the objectness score of the propagated boxes, as well as forward-backward optical flow consistency. We re-identify objects across occlusions using deep feature matching. We show that our model achieves strong performance on multiple established video object segmentation and tracking benchmarks, and can produce reasonable tracks in manipulation data. In particular, our model outperforms previous state-of-the-art in UVO and BURST, benchmarks for open-world object tracking and segmentation, despite never being explicitly trained for tracking. We hope that our approach can serve as a simple and extensible framework for future research.

Qianxu Wang, Haotong Zhang, Congyue Deng et al.

Humans demonstrate remarkable skill in transferring manipulation abilities across objects of varying shapes, poses, and appearances, a capability rooted in their understanding of semantic correspondences between different instances. To equip robots with a similar high-level comprehension, we present SparseDFF, a novel DFF for 3D scenes utilizing large 2D vision models to extract semantic features from sparse RGBD images, a domain where research is limited despite its relevance to many tasks with fixed-camera setups. SparseDFF generates view-consistent 3D DFFs, enabling efficient one-shot learning of dexterous manipulations by mapping image features to a 3D point cloud. Central to SparseDFF is a feature refinement network, optimized with a contrastive loss between views and a point-pruning mechanism for feature continuity. This facilitates the minimization of feature discrepancies w.r.t. end-effector parameters, bridging demonstrations and target manipulations. Validated in real-world scenarios with a dexterous hand, SparseDFF proves effective in manipulating both rigid and deformable objects, demonstrating significant generalization capabilities across object and scene variations.

Yumeng He, Xiaoying Wang, Peihao Li et al.

Large-scale controllable 3D assets are critical for computer graphics, embodied AI, robotics, and interactive content creation, yet creating diverse 3D assets remains challenging due to the high cost of manual modeling and rigging. Shape deformation offers a natural way to generate variations from existing meshes, but existing data-driven methods often rely on sparse user inputs, while parametric editing frameworks require manually designed control structures and category-specific configurations. Inspired by natural creatures, where a central spine governs global shape and cross-sectional ribs control local variation, we introduce Fishbone, a unified rib-spine representation for general shapes that supports controllable parametric mesh deformation, reduced-space dynamics, and animation. Given an input mesh, Fishbone computes a geodesic scalar field with an adaptive heat method, extracts iso-contours as cross-sectional ribs, constructs a smooth geometry-aware spine through rib centers, and associates surface vertices with nearby rib and spine structures using Gaussian-weighted skinning. The resulting representation enables real-time and predictable deformation: ribs control local profiles such as thickness, orientation, and cross-sectional variation, while the spine controls global bending, twisting, and stretching. The same structure also supports reduced-space simulation and keyframe animation. We further construct Fishbone-136K by augmenting Hunyuan3D with rib-spine structures, and demonstrate applications in controllable 3D generation, deformation-based data augmentation for robot learning, interactive mesh editing, and agentic generation. Experiments demonstrate the effectiveness, efficiency, and versatility of the proposed framework.

Tyler Zhu, Tengda Han, Leonidas Guibas et al.

The alignment of representations from different modalities has recently been shown to provide insights on the structural similarities and downstream capabilities of different encoders across diverse data types. While significant progress has been made in aligning images with text, the temporal nature of video data remains largely unexplored in this context. In this work, we conduct the first comprehensive study of video-text representation alignment, probing the capabilities of modern video and language encoders. Our findings reveal several key insights. First, we demonstrate that cross-modal alignment highly depends on the richness of both visual (static images vs. multi-frame videos) and text (single caption vs. a collection) data provided at test time, especially when using state-of-the-art video encoders. We propose parametric test-time scaling laws that capture this behavior and show remarkable predictive power against empirical observations. Secondly, we investigate the correlation between semantic alignment and performance on both semantic and non-semantic downstream tasks, providing initial evidence that strong alignment against text encoders may be linked to general-purpose video representation and understanding. Finally, we correlate temporal reasoning with cross-modal alignment providing a challenging test-bed for vision and language models. Overall, our work introduces video-text alignment as an informative zero-shot way to probe the representation power of different encoders for spatio-temporal data. Project page can be found at https://video-prh.github.io/

Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis et al.

Recent advances in large-scale pretraining have yielded visual foundation models with strong capabilities. Not only can recent models generalize to arbitrary images for their training task, their intermediate representations are useful for other visual tasks such as detection and segmentation. Given that such models can classify, delineate, and localize objects in 2D, we ask whether they also represent their 3D structure? In this work, we analyze the 3D awareness of visual foundation models. We posit that 3D awareness implies that representations (1) encode the 3D structure of the scene and (2) consistently represent the surface across views. We conduct a series of experiments using task-specific probes and zero-shot inference procedures on frozen features. Our experiments reveal several limitations of the current models. Our code and analysis can be found at https://github.com/mbanani/probe3d.

Mikaela Angelina Uy, Kiyohiro Nakayama, Guandao Yang et al.

Neural radiance fields (NeRF) rely on volume rendering to synthesize novel views. Volume rendering requires evaluating an integral along each ray, which is numerically approximated with a finite sum that corresponds to the exact integral along the ray under piecewise constant volume density. As a consequence, the rendered result is unstable w.r.t. the choice of samples along the ray, a phenomenon that we dub quadrature instability. We propose a mathematically principled solution by reformulating the sample-based rendering equation so that it corresponds to the exact integral under piecewise linear volume density. This simultaneously resolves multiple issues: conflicts between samples along different rays, imprecise hierarchical sampling, and non-differentiability of quantiles of ray termination distances w.r.t. model parameters. We demonstrate several benefits over the classical sample-based rendering equation, such as sharper textures, better geometric reconstruction, and stronger depth supervision. Our proposed formulation can be also be used as a drop-in replacement to the volume rendering equation of existing NeRF-based methods. Our project page can be found at pl-nerf.github.io.

Arijit Ray, Ahmed Abdelkader, Chengzhi Mao et al.

Reasoning goes beyond language; the real world requires reasoning about space, time, affordances, and much more that words alone cannot convey. Existing multimodal models exploring the potential of reasoning with images are brittle and do not scale. They rely on calling specialist tools, costly generation of images, or handcrafted reasoning data to switch between text and image thoughts. Instead, we offer a simpler alternative -- Mull-Tokens -- modality-agnostic latent tokens pre-trained to hold intermediate information in either image or text modalities to let the model think free-form towards the correct answer. We investigate best practices to train Mull-Tokens inspired by latent reasoning frameworks. We first train Mull-Tokens using supervision from interleaved text-image traces, and then fine-tune without any supervision by only using the final answers. Across four challenging spatial reasoning benchmarks involving tasks such as solving puzzles and taking different perspectives, we demonstrate that Mull-Tokens improve upon several baselines utilizing text-only reasoning or interleaved image-text reasoning, achieving a +3% average improvement and up to +16% on a puzzle solving reasoning-heavy split compared to our strongest baseline. Adding to conversations around challenges in grounding textual and visual reasoning, Mull-Tokens offers a simple solution to abstractly think in multiple modalities.

Songlin Li, Despoina Paschalidou, Leonidas Guibas

The increased demand for tools that automate the 3D content creation process led to tremendous progress in deep generative models that can generate diverse 3D objects of high fidelity. In this paper, we present PASTA, an autoregressive transformer architecture for generating high quality 3D shapes. PASTA comprises two main components: An autoregressive transformer that generates objects as a sequence of cuboidal primitives and a blending network, implemented with a transformer decoder that composes the sequences of cuboids and synthesizes high quality meshes for each object. Our model is trained in two stages: First we train our autoregressive generative model using only annotated cuboidal parts as supervision and next, we train our blending network using explicit 3D supervision, in the form of watertight meshes. Evaluations on various ShapeNet objects showcase the ability of our model to perform shape generation from diverse inputs \eg from scratch, from a partial object, from text and images, as well size-guided generation, by explicitly conditioning on a bounding box that defines the object's boundaries. Moreover, as our model considers the underlying part-based structure of a 3D object, we are able to select a specific part and produce shapes with meaningful variations of this part. As evidenced by our experiments, our model generates 3D shapes that are both more realistic and diverse than existing part-based and non part-based methods, while at the same time is simpler to implement and train.

Zilu Li, Guandao Yang, Qingqing Zhao et al.

This paper presents a method to leverage arbitrary neural network architecture for control variates. Control variates are crucial in reducing the variance of Monte Carlo integration, but they hinge on finding a function that both correlates with the integrand and has a known analytical integral. Traditional approaches rely on heuristics to choose this function, which might not be expressive enough to correlate well with the integrand. Recent research alleviates this issue by modeling the integrands with a learnable parametric model, such as a neural network. However, the challenge remains in creating an expressive parametric model with a known analytical integral. This paper proposes a novel approach to construct learnable parametric control variates functions from arbitrary neural network architectures. Instead of using a network to approximate the integrand directly, we employ the network to approximate the anti-derivative of the integrand. This allows us to use automatic differentiation to create a function whose integration can be constructed by the antiderivative network. We apply our method to solve partial differential equations using the Walk-on-sphere algorithm. Our results indicate that this approach is unbiased and uses various network architectures to achieve lower variance than other control variate methods.

Nikhil Raghuraman, Adam W. Harley, Leonidas Guibas

Current machine learning methods struggle to solve Bongard problems, which are a type of IQ test that requires deriving an abstract "concept" from a set of positive and negative "support" images, and then classifying whether or not a new query image depicts the key concept. On Bongard-HOI, a benchmark for natural-image Bongard problems, most existing methods have reached at best 69% accuracy (where chance is 50%). Low accuracy is often attributed to neural nets' lack of ability to find human-like symbolic rules. In this work, we point out that many existing methods are forfeiting accuracy due to a much simpler problem: they do not adapt image features given information contained in the support set as a whole, and rely instead on information extracted from individual supports. This is a critical issue, because the "key concept" in a typical Bongard problem can often only be distinguished using multiple positives and multiple negatives. We explore simple methods to incorporate this context and show substantial gains over prior works, leading to new state-of-the-art accuracy on Bongard-LOGO (75.3%) and Bongard-HOI (76.4%) compared to methods with equivalent vision backbone architectures and strong performance on the original Bongard problem set (60.8%).

Yijia Weng, Bowen Wen, Jonathan Tremblay et al.

We address the problem of building digital twins of unknown articulated objects from two RGBD scans of the object at different articulation states. We decompose the problem into two stages, each addressing distinct aspects. Our method first reconstructs object-level shape at each state, then recovers the underlying articulation model including part segmentation and joint articulations that associate the two states. By explicitly modeling point-level correspondences and exploiting cues from images, 3D reconstructions, and kinematics, our method yields more accurate and stable results compared to prior work. It also handles more than one movable part and does not rely on any object shape or structure priors. Project page: https://github.com/NVlabs/DigitalTwinArt

Alexandros Delitzas, Chenyangguang Zhang, Alexey Gavryushin et al.

We present FunRec, a method for reconstructing functional 3D digital twins of indoor scenes directly from egocentric RGB-D interaction videos. Unlike existing methods on articulated reconstruction, which rely on controlled setups, multi-state captures, or CAD priors, FunRec operates directly on in-the-wild human interaction sequences to recover interactable 3D scenes. It automatically discovers articulated parts, estimates their kinematic parameters, tracks their 3D motion, and reconstructs static and moving geometry in canonical space, yielding simulation-compatible meshes. Across new real and simulated benchmarks, FunRec surpasses prior work by a large margin, achieving up to +50 mIoU improvement in part segmentation, 5-10 times lower articulation and pose errors, and significantly higher reconstruction accuracy. We further demonstrate applications on URDF/USD export for simulation, hand-guided affordance mapping and robot-scene interaction.

Yining Hong, Jiageng Liu, Han Yin et al.

Spatial intelligence unfolds through a perception-action loop: agents act to acquire observations, and reason about how observations vary as a function of action. Rather than passively processing what is seen, they actively uncover what is unseen - occluded structure, dynamics, containment, and functionality that cannot be resolved from passive sensing alone. We move beyond prior formulations of spatial intelligence that assume oracle observations by recasting the observer as an actor. We introduce ESI-BENCH, a comprehensive benchmark for embodied spatial intelligence spanning 10 task categories and 29 subcategories built on OmniGibson, grounded in Spelke's core knowledge systems. Agents must decide what abilities to deploy - perception, locomotion, and manipulation - and how to sequence them to actively accumulate task-relevant evidence. We conduct extensive experiments on state-of-the-art MLLMs and find that active exploration substantially outperforms passive counterparts, with agents spontaneously discovering emergent spatial strategies without explicit instructions, while random multi-view often adds noise rather than signal despite consuming far more images. Most failures stem not from weak perception but from action blindness: poor action choices lead to poor observations, which in turn drive cascading errors. While explicit 3D grounding stabilizes reasoning on depth-sensitive tasks, imperfect 3D representation proves more harmful than 2D baselines by distorting spatial relations. Human studies further reveal that unlike humans who seek falsifying viewpoints and revise beliefs under contradiction, models commit prematurely with high confidence regardless of evidence quality, exposing a metacognitive gap that neither better perception nor more embodied interaction alone can close.

Jiawei Yang, Katie Z Luo, Jiefeng Li et al.

We study a crucial yet often overlooked issue inherent to Vision Transformers (ViTs): feature maps of these models exhibit grid-like artifacts, which hurt the performance of ViTs in downstream dense prediction tasks such as semantic segmentation, depth prediction, and object discovery. We trace this issue down to the positional embeddings at the input stage. To mitigate this, we propose a two-stage denoising approach, termed Denoising Vision Transformers (DVT). In the first stage, we separate the clean features from those contaminated by positional artifacts by enforcing cross-view feature consistency with neural fields on a per-image basis. This per-image optimization process extracts artifact-free features from raw ViT outputs, providing clean feature estimates for offline applications. In the second stage, we train a lightweight transformer block to predict clean features from raw ViT outputs, leveraging the derived estimates of the clean features as supervision. Our method, DVT, does not require re-training the existing pre-trained ViTs, and is immediately applicable to any Vision Transformer architecture. We evaluate our method on a variety of representative ViTs (DINO, DeiT-III, EVA02, CLIP, DINOv2, DINOv2-reg) and demonstrate that DVT consistently improves existing state-of-the-art general-purpose models in semantic and geometric tasks across multiple datasets. We hope our study will encourage a re-evaluation of ViT design, especially regarding the naive use of positional embeddings. Our code and checkpoints are publicly available.