Daily Papers

Vision-language models encode continuous geometry that their text pathway fails to express: a 6,000-parameter linear probe extracts hand joint angles at 6.1 degrees MAE from frozen features, while the best text output achieves only 20.0 degrees -- a 3.3x bottleneck. LoRA fine-tuning (r=16, 2,000 images) narrows this gap to 6.5 degrees, providing evidence for a pathway-training deficit rather than a representational one. Training objective determines accuracy more than architecture: five encoders spanning self-supervised, contrastive, and hybrid paradigms converge to statistically equivalent accuracy (R^2 approximately 0.55, TOST-equivalent at delta=0.03) despite sharing as little as CKA=0.41 representational similarity -- functional convergence without representational convergence. Autoregressive generation damages geometric fidelity, but the damage originates in the generation process, not in language alignment: Qwen2.5-VL's LLM layers actually improve probe accuracy over its raw vision encoder. Layer-wise analysis reveals a universal mid-network accuracy peak across all architectures, with attention heads in layers 18-22 carrying disproportionate geometric signal. These findings enable a single frozen backbone to function as a multi-task geometric sensor through lightweight probes, without fine-tuning or text generation.

Parameter-efficient adaptation of pretrained vision models is commonly performed through linear probes, prompts, low-rank updates, or lightweight residual modules. While effective, these methods usually treat adaptation as a discrete Euclidean perturbation of frozen representations, without explicitly modeling the geometry of the task-induced feature displacement. We propose MC-RFM, a mixed-curvature Riemannian flow-matching framework for few-shot adaptation of frozen visual backbones. The key idea is to represent adapted features on a product manifold combining a hyperbolic factor, which captures hierarchy-sensitive semantic structure, and a Euclidean factor, which preserves locally discriminative visual variation. Adaptation is formulated as a task-conditioned continuous transport from frozen features to support-set prototypes, trained with a flow-matching objective and coupled to a hybrid prototype-linear classifier. The method is lightweight, backbone-agnostic, and operates entirely on cached frozen features. Across seven visual recognition benchmarks, five frozen backbones, and 1/4/16-shot regimes, MC-RFM is the best-performing method in a majority of evaluated settings, with the strongest gains on Transformer backbones and fine-grained datasets. Ablations show that the mixed-curvature head, task conditioning, adaptive branch gating, prototype shrinkage, and discriminative supervision each contribute to performance. These results suggest that few-shot adaptation benefits not only from deciding which parameters to update, but also from modeling how representations should move through a geometry matched to the structure of the downstream task.

This paper evaluates DINOv3, a recent large-scale self-supervised vision backbone, for visuomotor diffusion policy learning in robotic manipulation. We investigate whether a purely self-supervised encoder can match or surpass conventional supervised ImageNet-pretrained backbones (e.g., ResNet-18) under three regimes: training from scratch, frozen, and finetuned. Across four benchmark tasks (Push-T, Lift, Can, Square) using a unified FiLM-conditioned diffusion policy, we find that (i) finetuned DINOv3 matches or exceeds ResNet-18 on several tasks, (ii) frozen DINOv3 remains competitive, indicating strong transferable priors, and (iii) self-supervised features improve sample efficiency and robustness. These results support self-supervised large visual models as effective, generalizable perceptual front-ends for action diffusion policies, motivating further exploration of scalable label-free pretraining in robotic manipulation. Compared to using ResNet18 as a backbone, our approach with DINOv3 achieves up to a 10% absolute increase in test-time success rates on challenging tasks such as Can, and on-the-par performance in tasks like Lift, PushT, and Square.

Federated learning (FL) emerges as a decentralized learning framework which trains models from multiple distributed clients without sharing their data to preserve privacy. Recently, large-scale pre-trained models (e.g., Vision Transformer) have shown a strong capability of deriving robust representations. However, the data heterogeneity among clients, the limited computation resources, and the communication bandwidth restrict the deployment of large-scale models in FL frameworks. To leverage robust representations from large-scale models while enabling efficient model personalization for heterogeneous clients, we propose a novel personalized FL framework of client-specific Prompt Generation (pFedPG), which learns to deploy a personalized prompt generator at the server for producing client-specific visual prompts that efficiently adapts frozen backbones to local data distributions. Our proposed framework jointly optimizes the stages of personalized prompt adaptation locally and personalized prompt generation globally. The former aims to train visual prompts that adapt foundation models to each client, while the latter observes local optimization directions to generate personalized prompts for all clients. Through extensive experiments on benchmark datasets, we show that our pFedPG is favorable against state-of-the-art personalized FL methods under various types of data heterogeneity, allowing computation and communication efficient model personalization.

Vision-language models (VLMs) like CLIP (Contrastive Language-Image Pre-Training) have seen remarkable success in visual recognition, highlighting the increasing need to safeguard the intellectual property (IP) of well-trained models. Effective IP protection extends beyond ensuring authorized usage; it also necessitates restricting model deployment to authorized data domains, particularly when the model is fine-tuned for specific target domains. However, current IP protection methods often rely solely on the visual backbone, which may lack sufficient semantic richness. To bridge this gap, we introduce IP-CLIP, a lightweight IP protection strategy tailored to CLIP, employing a prompt-based learning approach. By leveraging the frozen visual backbone of CLIP, we extract both image style and content information, incorporating them into the learning of IP prompt. This strategy acts as a robust barrier, effectively preventing the unauthorized transfer of features from authorized domains to unauthorized ones. Additionally, we propose a style-enhancement branch that constructs feature banks for both authorized and unauthorized domains. This branch integrates self-enhanced and cross-domain features, further strengthening IP-CLIP's capability to block features from unauthorized domains. Finally, we present new three metrics designed to better balance the performance degradation of authorized and unauthorized domains. Comprehensive experiments in various scenarios demonstrate its promising potential for application in IP protection tasks for VLMs.

Large vision--language models (VLMs) often use a frozen vision backbone, whose image features are mapped into a large language model through a lightweight connector. While transformer-based encoders are the standard visual backbone, we ask whether state space model (SSM) vision backbones can be a strong alternative. We systematically evaluate SSM vision backbones for VLMs in a controlled setting. Under matched ImageNet-1K initialization, the SSM backbone achieves the strongest overall performance across both VQA and grounding/localization. We further adapt both SSM and ViT-family backbones with detection or segmentation training and find that dense-task tuning generally improves performance across families; after this adaptation, the SSM backbone remains competitive while operating at a substantially smaller model scale. We further observe that (i) higher ImageNet accuracy or larger backbones do not reliably translate into better VLM performance, and (ii) some visual backbones are unstable in localization. Based on these findings, we propose stabilization strategies that improve robustness for both backbone families and highlight SSM backbones as a strong alternative to transformer-based vision encoders in VLMs.

Open-vocabulary segmentation poses significant challenges, as it requires segmenting and recognizing objects across an open set of categories in unconstrained environments. Building on the success of powerful vision-language (ViL) foundation models, such as CLIP, recent efforts sought to harness their zero-short capabilities to recognize unseen categories. Despite notable performance improvements, these models still encounter the critical issue of generating precise mask proposals for unseen categories and scenarios, resulting in inferior segmentation performance eventually. To address this challenge, we introduce a novel approach, FrozenSeg, designed to integrate spatial knowledge from a localization foundation model (e.g., SAM) and semantic knowledge extracted from a ViL model (e.g., CLIP), in a synergistic framework. Taking the ViL model's visual encoder as the feature backbone, we inject the space-aware feature into the learnable queries and CLIP features within the transformer decoder. In addition, we devise a mask proposal ensemble strategy for further improving the recall rate and mask quality. To fully exploit pre-trained knowledge while minimizing training overhead, we freeze both foundation models, focusing optimization efforts solely on a lightweight transformer decoder for mask proposal generation-the performance bottleneck. Extensive experiments demonstrate that FrozenSeg advances state-of-the-art results across various segmentation benchmarks, trained exclusively on COCO panoptic data, and tested in a zero-shot manner. Code is available at https://github.com/chenxi52/FrozenSeg.

The goal of this paper is to provide a new perspective on audio-visual target speaker extraction (AV-TSE) by decoupling the separation and target selection. Conventional AV-TSE systems typically integrate audio and visual features deeply to re-learn the entire separation process, which can act as a fidelity ceiling due to the noisy nature of in-the-wild audio-visual datasets. To address this, we propose Plug-and-Steer, which assigns high-fidelity separation to a frozen audio-only backbone and limits the role of visual modality strictly to target selection. We introduce the Latent Steering Matrix (LSM), a minimalist linear transformation that re-routes latent features within the backbone to anchor the target speaker to a designated channel. Experiments across four representative architectures show that our method effectively preserves the acoustic priors of diverse backbones, achieving perceptual quality comparable to the original backbones. Audio samples are available at: https://plugandsteer.github.io

The deployment of automated pavement defect detection is often hindered by poor cross-domain generalization. Supervised detectors achieve strong in-domain accuracy but require costly re-annotation for new environments, while standard self-supervised methods capture generic features and remain vulnerable to domain shift. We propose \ours, a self-supervised framework that visually probes target domains without labels. \ours introduces a Self-supervised Prompt Enhancement Module (SPEM), which derives defect-aware prompts from unlabeled target data to guide a frozen ViT backbone, and a Domain-Aware Prompt Alignment (DAPA) objective, which aligns prompt-conditioned source and target representations. Experiments on four challenging benchmarks show that \ours consistently outperforms strong supervised, self-supervised, and adaptation baselines, achieving robust zero-shot transfer, improved resilience to domain variations, and high data efficiency in few-shot adaptation. These results highlight self-supervised prompting as a practical direction for building scalable and adaptive visual inspection systems. Source code is publicly available: https://github.com/xixiaouab/PROBE/tree/main

One of the principal challenges in building VLM-powered GUI agents is visual grounding, i.e., localizing the appropriate screen region for action execution based on both the visual content and the textual plans. Most existing work formulates this as a text-based coordinate generation task. However, these approaches suffer from several limitations: weak spatial-semantic alignment, inability to handle ambiguous supervision targets, and a mismatch between the dense nature of screen coordinates and the coarse, patch-level granularity of visual features extracted by models like Vision Transformers. In this paper, we propose GUI-Actor, a VLM-based method for coordinate-free GUI grounding. At its core, GUI-Actor introduces an attention-based action head that learns to align a dedicated <ACTOR> token with all relevant visual patch tokens, enabling the model to propose one or more action regions in a single forward pass. In line with this, we further design a grounding verifier to evaluate and select the most plausible action region from the candidates proposed for action execution. Extensive experiments show that GUI-Actor outperforms prior state-of-the-art methods on multiple GUI action grounding benchmarks, with improved generalization to unseen screen resolutions and layouts. Notably, GUI-Actor-7B even surpasses UI-TARS-72B (38.1) on ScreenSpot-Pro, achieving scores of 40.7 with Qwen2-VL and 44.6 with Qwen2.5-VL as backbones. Furthermore, by incorporating the verifier, we find that fine-tuning only the newly introduced action head (~100M parameters for 7B model) while keeping the VLM backbone frozen is sufficient to achieve performance comparable to previous state-of-the-art models, highlighting that GUI-Actor can endow the underlying VLM with effective grounding capabilities without compromising its general-purpose strengths.

This paper explores feature prediction as a stand-alone objective for unsupervised learning from video and introduces V-JEPA, a collection of vision models trained solely using a feature prediction objective, without the use of pretrained image encoders, text, negative examples, reconstruction, or other sources of supervision. The models are trained on 2 million videos collected from public datasets and are evaluated on downstream image and video tasks. Our results show that learning by predicting video features leads to versatile visual representations that perform well on both motion and appearance-based tasks, without adaption of the model's parameters; e.g., using a frozen backbone. Our largest model, a ViT-H/16 trained only on videos, obtains 81.9% on Kinetics-400, 72.2% on Something-Something-v2, and 77.9% on ImageNet1K.

We present Diversity-Aware Meta Visual Prompting~(DAM-VP), an efficient and effective prompting method for transferring pre-trained models to downstream tasks with frozen backbone. A challenging issue in visual prompting is that image datasets sometimes have a large data diversity whereas a per-dataset generic prompt can hardly handle the complex distribution shift toward the original pretraining data distribution properly. To address this issue, we propose a dataset Diversity-Aware prompting strategy whose initialization is realized by a Meta-prompt. Specifically, we cluster the downstream dataset into small homogeneity subsets in a diversity-adaptive way, with each subset has its own prompt optimized separately. Such a divide-and-conquer design reduces the optimization difficulty greatly and significantly boosts the prompting performance. Furthermore, all the prompts are initialized with a meta-prompt, which is learned across several datasets. It is a bootstrapped paradigm, with the key observation that the prompting knowledge learned from previous datasets could help the prompt to converge faster and perform better on a new dataset. During inference, we dynamically select a proper prompt for each input, based on the feature distance between the input and each subset. Through extensive experiments, our DAM-VP demonstrates superior efficiency and effectiveness, clearly surpassing previous prompting methods in a series of downstream datasets for different pretraining models. Our code is available at: https://github.com/shikiw/DAM-VP.

Hand motion plays a central role in human interaction, yet modeling realistic 4D hand motion (i.e., 3D hand pose sequences over time) remains challenging. Research in this area is typically divided into two tasks: (1) Estimation approaches reconstruct precise motion from visual observations, but often fail under hand occlusion or absence; (2) Generation approaches focus on synthesizing hand poses by exploiting generative priors under multi-modal structured inputs and infilling motion from incomplete sequences. However, this separation not only limits the effective use of heterogeneous condition signals that frequently arise in practice, but also prevents knowledge transfer between the two tasks. We present UniHand, a unified diffusion-based framework that formulates both estimation and generation as conditional motion synthesis. UniHand integrates heterogeneous inputs by embedding structured signals into a shared latent space through a joint variational autoencoder, which aligns conditions such as MANO parameters and 2D skeletons. Visual observations are encoded with a frozen vision backbone, while a dedicated hand perceptron extracts hand-specific cues directly from image features, removing the need for complex detection and cropping pipelines. A latent diffusion model then synthesizes consistent motion sequences from these diverse conditions. Extensive experiments across multiple benchmarks demonstrate that UniHand delivers robust and accurate hand motion modeling, maintaining performance under severe occlusions and temporally incomplete inputs.

The clinical translation of dermatological AI is hindered by opaque reasoning and systematic performance disparities across skin tones. Here we present SkinGPT-R1, a multimodal large language model that integrates chain-of-thought diagnostic reasoning with a fairness-aware mixture-of-experts architecture for interpretable and equitable skin disease diagnosis. Through parameter-efficient adaptation of a frozen reasoning backbone, SkinGPT-R1 generates structured diagnostic reports comprising visual findings, differential reasoning, and final diagnosis. Across seven external datasets spanning diverse pathologies and imaging conditions, SkinGPT-R1 achieves state-of-the-art accuracy on six benchmarks, including 82.50\% on a challenging 40-class long-tail classification task (+19.30\% over leading baselines). Blinded evaluation by five board-certified dermatologists on 1,000 phenotypically balanced cases yields a mean score of 3.6 out of 5, with the highest ratings in safety (3.8) and reasoning coherence (3.6), indicating that the generated rationales are clinically safe, logically grounded, and suitable for supporting diagnostic decision-making. Critically, SkinGPT-R1 mitigates algorithmic bias across the full Fitzpatrick spectrum, achieving a robust worst-group performance of 41.40\% on the Fitz17k benchmark and a five-fold relative improvement in lower-bound accuracy on the DDI dataset compared to standard multimodal baselines. These results establish a framework for trustworthy, fair, and explainable AI-assisted dermatological diagnosis.

Instruction-based image editing has achieved remarkable success in semantic alignment, yet state-of-the-art models frequently fail to render physically plausible results when editing involves complex causal dynamics, such as refraction or material deformation. We attribute this limitation to the dominant paradigm that treats editing as a discrete mapping between image pairs, which provides only boundary conditions and leaves transition dynamics underspecified. To address this, we reformulate physics-aware editing as predictive physical state transitions and introduce PhysicTran38K, a large-scale video-based dataset comprising 38K transition trajectories across five physical domains, constructed via a two-stage filtering and constraint-aware annotation pipeline. Building on this supervision, we propose PhysicEdit, an end-to-end framework equipped with a textual-visual dual-thinking mechanism. It combines a frozen Qwen2.5-VL for physically grounded reasoning with learnable transition queries that provide timestep-adaptive visual guidance to a diffusion backbone. Experiments show that PhysicEdit improves over Qwen-Image-Edit by 5.9% in physical realism and 10.1% in knowledge-grounded editing, setting a new state-of-the-art for open-source methods, while remaining competitive with leading proprietary models.

With the bloom of Large Language Models (LLMs), Multimodal Large Language Models (MLLMs) that incorporate LLMs with pre-trained vision models have recently demonstrated impressive performance across diverse vision-language tasks. However, they fall short to comprehend context involving multiple images. A primary reason for this shortcoming is that the visual features for each images are encoded individually by frozen encoders before feeding into the LLM backbone, lacking awareness of other images and the multimodal instructions. We term this issue as prior-LLM modality isolation and propose a two phase paradigm, browse-and-concentrate, to enable in-depth multimodal context fusion prior to feeding the features into LLMs. This paradigm initially "browses" through the inputs for essential insights, and then revisits the inputs to "concentrate" on crucial details, guided by these insights, to achieve a more comprehensive understanding of the multimodal inputs. Additionally, we develop training strategies specifically to enhance the understanding of multi-image inputs. Our method markedly boosts the performance on 7 multi-image scenarios, contributing to increments on average accuracy by 2.13% and 7.60% against strong MLLMs baselines with 3B and 11B LLMs, respectively.

Recent dense audio-visual (AV) models achieve impressive retrieval and emergent localization, but almost all evidence comes from English-centric, caption-rich web video. It is unclear whether these objectives survive in low-resource, code-switched, and noisy multilingual settings that typify developing regions. We show they do**-**and that the choice of aggregation function becomes even more critical. Using a multilingual subset of Project Vaani spanning dozens of Indian languages and dialectal variants, we compare three contrastive objectives: (i) a global mean-pooled loss (CLIP-style), (ii) a dense max-mean token matcher (DenseAV-style), and (iii) a simple hybrid (motivated by frozen-vision alignment strategies). The dense objective delivers a +59% relative R@1 (Audio Visual) improvement over global pooling and substantially lower mean/median ranks, while consistently producing sharp zero-shot localization heatmaps of spoken objects-despite keeping the vision backbone entirely frozen (no LoRA / partial fine-tuning). Our results demonstrate that dense token routing is not a luxury of high-resource English corpora; it is more decisive when annotations and acoustic cleanliness are scarce. We release the codebase and trained models.

We present EgoNav, a system that enables a humanoid robot to traverse diverse, unseen environments by learning entirely from 5 hours of human walking data, with no robot data or finetuning. A diffusion model predicts distributions of plausible future trajectories conditioned on past trajectory, a 360 deg visual memory fusing color, depth, and semantics, and video features from a frozen DINOv3 backbone that capture appearance cues invisible to depth sensors. A hybrid sampling scheme achieves real-time inference in 10 denoising steps, and a receding-horizon controller selects paths from the predicted distribution. We validate EgoNav through offline evaluations, where it outperforms baselines in collision avoidance and multi-modal coverage, and through zero-shot deployment on a Unitree G1 humanoid across unseen indoor and outdoor environments. Behaviors such as waiting for doors to open, navigating around crowds, and avoiding glass walls emerge naturally from the learned prior. We will release the dataset and trained models. Our website: https://egonav.weizhuowang.com

Adapting large-scale Vision-Language Models (VLMs) like CLIP to downstream tasks often suffers from a "one-size-fits-all" architectural approach, where visual and textual tokens are processed uniformly by wide, generic adapters. We argue that this homogeneity ignores the distinct structural nature of the modalities -- spatial locality in images versus semantic density in text. To address this, we propose HeBA (Heterogeneous Bottleneck Adapter), a unified architectural framework that introduces modality-specific structural inductive biases. HeBA departs from conventional designs through three key architectural innovations: (1) Heterogeneity: It processes visual tokens via 2D depthwise-separable convolutions to preserve spatial correlations, while distinctively processing text tokens via dense linear projections to capture semantic relationships; (2) Bottleneck Regularization: Unlike standard expanding adapters, HeBA employs a compression bottleneck (D -> D/4) that explicitly forces the model to learn compact, robust features and acts as a structural regularizer; and (3) Active Gradient Initialization: We challenge the restrictive zero-initialization paradigm, utilizing a Kaiming initialization strategy that ensures sufficient initial gradient flow to accelerate convergence without compromising the frozen backbone's pre-trained knowledge. Extensive experiments demonstrate that HeBA's architecturally specialized design achieves superior stability and accuracy, establishing a new state-of-the-art on 11 few-shot benchmarks. Code is available at https://github.com/Jahid12012021/VLM-HeBA.

Current popular backbones in computer vision, such as Vision Transformers (ViT) and ResNets are trained to perceive the world from 2D images. However, to more effectively understand 3D structural priors in 2D backbones, we propose Mask3D to leverage existing large-scale RGB-D data in a self-supervised pre-training to embed these 3D priors into 2D learned feature representations. In contrast to traditional 3D contrastive learning paradigms requiring 3D reconstructions or multi-view correspondences, our approach is simple: we formulate a pre-text reconstruction task by masking RGB and depth patches in individual RGB-D frames. We demonstrate the Mask3D is particularly effective in embedding 3D priors into the powerful 2D ViT backbone, enabling improved representation learning for various scene understanding tasks, such as semantic segmentation, instance segmentation and object detection. Experiments show that Mask3D notably outperforms existing self-supervised 3D pre-training approaches on ScanNet, NYUv2, and Cityscapes image understanding tasks, with an improvement of +6.5% mIoU against the state-of-the-art Pri3D on ScanNet image semantic segmentation.

Foundation models have emerged as effective backbones for many vision tasks. However, current self-supervised features entangle high-level semantics with low-level physical factors, such as geometry and illumination, hindering their use in tasks requiring explicit physical reasoning. In this paper, we introduce Φeat, a novel physically-grounded visual backbone that encourages a representation sensitive to material identity, including reflectance cues and geometric mesostructure. Our key idea is to employ a pretraining strategy that contrasts spatial crops and physical augmentations of the same material under varying shapes and lighting conditions. While similar data have been used in high-end supervised tasks such as intrinsic decomposition or material estimation, we demonstrate that a pure self-supervised training strategy, without explicit labels, already provides a strong prior for tasks requiring robust features invariant to external physical factors. We evaluate the learned representations through feature similarity analysis and material selection, showing that Φeat captures physically-grounded structure beyond semantic grouping. These findings highlight the promise of unsupervised physical feature learning as a foundation for physics-aware perception in vision and graphics. These findings highlight the promise of unsupervised physical feature learning as a foundation for physics-aware perception in vision and graphics.

Click-based interactive image segmentation aims at extracting objects with a limited user clicking. A hierarchical backbone is the de-facto architecture for current methods. Recently, the plain, non-hierarchical Vision Transformer (ViT) has emerged as a competitive backbone for dense prediction tasks. This design allows the original ViT to be a foundation model that can be finetuned for downstream tasks without redesigning a hierarchical backbone for pretraining. Although this design is simple and has been proven effective, it has not yet been explored for interactive image segmentation. To fill this gap, we propose SimpleClick, the first interactive segmentation method that leverages a plain backbone. Based on the plain backbone, we introduce a symmetric patch embedding layer that encodes clicks into the backbone with minor modifications to the backbone itself. With the plain backbone pretrained as a masked autoencoder (MAE), SimpleClick achieves state-of-the-art performance. Remarkably, our method achieves 4.15 NoC@90 on SBD, improving 21.8% over the previous best result. Extensive evaluation on medical images demonstrates the generalizability of our method. We further develop an extremely tiny ViT backbone for SimpleClick and provide a detailed computational analysis, highlighting its suitability as a practical annotation tool.

We present Muskie, a native multi-view vision backbone designed for 3D vision tasks. Unlike existing models, which are frame-wise and exhibit limited multi-view consistency, Muskie is designed to process multiple views simultaneously and introduce multi-view consistency in pre-training stage. Muskie is trained to reconstruct heavily masked content in one view by finding and utilizing geometric correspondences from other views. Through this pretext task and our proposed aggressive masking strategy, the model implicitly to learn view-invariant features and develop strong geometric understanding without any 3D supervision. Compared with state-of-the-art frame-wise backbones such as DINO, Muskie achieves higher multi-view correspondence accuracy. Furthermore, we demonstrate that using Muskie as a backbone consistently enhances performance on downstream 3D tasks, including camera pose estimation and pointmap reconstruction. Codes are publicly available at https://leo-frank.github.io/Muskie/

Estimating the 6D pose of objects unseen during training is highly desirable yet challenging. Zero-shot object 6D pose estimation methods address this challenge by leveraging additional task-specific supervision provided by large-scale, photo-realistic synthetic datasets. However, their performance heavily depends on the quality and diversity of rendered data and they require extensive training. In this work, we show how to tackle the same task but without training on specific data. We propose FreeZe, a novel solution that harnesses the capabilities of pre-trained geometric and vision foundation models. FreeZe leverages 3D geometric descriptors learned from unrelated 3D point clouds and 2D visual features learned from web-scale 2D images to generate discriminative 3D point-level descriptors. We then estimate the 6D pose of unseen objects by 3D registration based on RANSAC. We also introduce a novel algorithm to solve ambiguous cases due to geometrically symmetric objects that is based on visual features. We comprehensively evaluate FreeZe across the seven core datasets of the BOP Benchmark, which include over a hundred 3D objects and 20,000 images captured in various scenarios. FreeZe consistently outperforms all state-of-the-art approaches, including competitors extensively trained on synthetic 6D pose estimation data. Code will be publicly available at https://andreacaraffa.github.io/freeze.

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.

Generating multiview images from a single view facilitates the rapid generation of a 3D mesh conditioned on a single image. Recent methods that introduce 3D global representation into diffusion models have shown the potential to generate consistent multiviews, but they have reduced generation speed and face challenges in maintaining generalizability and quality. To address this issue, we propose EpiDiff, a localized interactive multiview diffusion model. At the core of the proposed approach is to insert a lightweight epipolar attention block into the frozen diffusion model, leveraging epipolar constraints to enable cross-view interaction among feature maps of neighboring views. The newly initialized 3D modeling module preserves the original feature distribution of the diffusion model, exhibiting compatibility with a variety of base diffusion models. Experiments show that EpiDiff generates 16 multiview images in just 12 seconds, and it surpasses previous methods in quality evaluation metrics, including PSNR, SSIM and LPIPS. Additionally, EpiDiff can generate a more diverse distribution of views, improving the reconstruction quality from generated multiviews. Please see our project page at https://huanngzh.github.io/EpiDiff/.

Vision language models (VLMs) have seen growing adoption in recent years, but many still struggle with basic spatial reasoning errors. We hypothesize that this is due to VLMs adopting pre-trained vision backbones, specifically vision transformers (ViTs) trained with image-level supervision and minimal inductive biases. Such models may fail to encode the class contents at each position in the image, and our goal is to resolve this by ensuring that the vision backbone effectively captures both local and global image semantics. Our main insight is that we do not require new supervision to learn this capability -- pre-trained models contain significant knowledge of local semantics that we can extract and use for scalable self-supervision. We propose a new efficient post-training stage for ViTs called locality alignment and a novel fine-tuning procedure called MaskEmbed that uses a masked reconstruction loss to learn semantic contributions for each image patch. We first evaluate locality alignment with a vision-only benchmark, finding that it improves a model's performance at a patch-level semantic segmentation task, especially for strong backbones trained with image-caption pairs (e.g., CLIP and SigLIP). We then train a series of VLMs with and without locality alignment, and show that locality-aligned backbones improve performance across a range of benchmarks, particularly ones that involve spatial understanding (e.g., RefCOCO, OCID-Ref, TallyQA, VSR, AI2D). Overall, we demonstrate that we can efficiently learn local semantic extraction via a locality alignment stage, and that this procedure complements existing VLM training recipes that use off-the-shelf vision backbones.

Neural network based computer vision systems are typically built on a backbone, a pretrained or randomly initialized feature extractor. Several years ago, the default option was an ImageNet-trained convolutional neural network. However, the recent past has seen the emergence of countless backbones pretrained using various algorithms and datasets. While this abundance of choice has led to performance increases for a range of systems, it is difficult for practitioners to make informed decisions about which backbone to choose. Battle of the Backbones (BoB) makes this choice easier by benchmarking a diverse suite of pretrained models, including vision-language models, those trained via self-supervised learning, and the Stable Diffusion backbone, across a diverse set of computer vision tasks ranging from classification to object detection to OOD generalization and more. Furthermore, BoB sheds light on promising directions for the research community to advance computer vision by illuminating strengths and weakness of existing approaches through a comprehensive analysis conducted on more than 1500 training runs. While vision transformers (ViTs) and self-supervised learning (SSL) are increasingly popular, we find that convolutional neural networks pretrained in a supervised fashion on large training sets still perform best on most tasks among the models we consider. Moreover, in apples-to-apples comparisons on the same architectures and similarly sized pretraining datasets, we find that SSL backbones are highly competitive, indicating that future works should perform SSL pretraining with advanced architectures and larger pretraining datasets. We release the raw results of our experiments along with code that allows researchers to put their own backbones through the gauntlet here: https://github.com/hsouri/Battle-of-the-Backbones

Synthesizing high-fidelity frozen 3D scenes from monocular Mannequin-Challenge (MC) videos is a unique problem distinct from standard dynamic scene reconstruction. Instead of focusing on modeling motion, our goal is to create a frozen scene while strategically preserving subtle dynamics to enable user-controlled instant selection. To achieve this, we introduce a novel application of dynamic Gaussian splatting: the scene is modeled dynamically, which retains nearby temporal variation, and a static scene is rendered by fixing the model's time parameter. However, under this usage, monocular capture with sparse temporal supervision introduces artifacts like ghosting and blur for Gaussians that become unobserved or occluded at weakly supervised timestamps. We propose Splannequin, an architecture-agnostic regularization that detects two states of Gaussian primitives, hidden and defective, and applies temporal anchoring. Under predominantly forward camera motion, hidden states are anchored to their recent well-observed past states, while defective states are anchored to future states with stronger supervision. Our method integrates into existing dynamic Gaussian pipelines via simple loss terms, requires no architectural changes, and adds zero inference overhead. This results in markedly improved visual quality, enabling high-fidelity, user-selectable frozen-time renderings, validated by a 96% user preference. Project page: https://chien90190.github.io/splannequin/

Forecasting what will happen next is a critical skill for general-purpose systems that plan or act in the world at different levels of abstraction. In this paper, we identify a strong correlation between a vision model's perceptual ability and its generalist forecasting performance over short time horizons. This trend holds across a diverse set of pretrained models-including those trained generatively-and across multiple levels of abstraction, from raw pixels to depth, point tracks, and object motion. The result is made possible by a novel generalist forecasting framework that operates on any frozen vision backbone: we train latent diffusion models to forecast future features in the frozen representation space, which are then decoded via lightweight, task-specific readouts. To enable consistent evaluation across tasks, we introduce distributional metrics that compare distributional properties directly in the space of downstream tasks and apply this framework to nine models and four tasks. Our results highlight the value of bridging representation learning and generative modeling for temporally grounded video understanding.

Modern visual agents require representations that are general, causal, and physically structured to operate in real-time streaming environments. However, current vision foundation models remain fragmented, specializing narrowly in image semantic perception, offline temporal modeling, or spatial geometry. This paper introduces OmniStream, a unified streaming visual backbone that effectively perceives, reconstructs, and acts from diverse visual inputs. By incorporating causal spatiotemporal attention and 3D rotary positional embeddings (3D-RoPE), our model supports efficient, frame-by-frame online processing of video streams via a persistent KV-cache. We pre-train OmniStream using a synergistic multi-task framework coupling static and temporal representation learning, streaming geometric reconstruction, and vision-language alignment on 29 datasets. Extensive evaluations show that, even with a strictly frozen backbone, OmniStream achieves consistently competitive performance with specialized experts across image and video probing, streaming geometric reconstruction, complex video and spatial reasoning, as well as robotic manipulation (unseen at training). Rather than pursuing benchmark-specific dominance, our work demonstrates the viability of training a single, versatile vision backbone that generalizes across semantic, spatial, and temporal reasoning, i.e., a more meaningful step toward general-purpose visual understanding for interactive and embodied agents.

A fundamental problem in learning robust and expressive visual representations lies in efficiently estimating the spatial relationships of visual semantics throughout the entire image. In this study, we propose vHeat, a novel vision backbone model that simultaneously achieves both high computational efficiency and global receptive field. The essential idea, inspired by the physical principle of heat conduction, is to conceptualize image patches as heat sources and model the calculation of their correlations as the diffusion of thermal energy. This mechanism is incorporated into deep models through the newly proposed module, the Heat Conduction Operator (HCO), which is physically plausible and can be efficiently implemented using DCT and IDCT operations with a complexity of O(N^{1.5}). Extensive experiments demonstrate that vHeat surpasses Vision Transformers (ViTs) across various vision tasks, while also providing higher inference speeds, reduced FLOPs, and lower GPU memory usage for high-resolution images. The code will be released at https://github.com/MzeroMiko/vHeat.

Recent vision foundation models can extract universal representations and show impressive abilities in various tasks. However, their application on object detection is largely overlooked, especially without fine-tuning them. In this work, we show that frozen foundation models can be a versatile feature enhancer, even though they are not pre-trained for object detection. Specifically, we explore directly transferring the high-level image understanding of foundation models to detectors in the following two ways. First, the class token in foundation models provides an in-depth understanding of the complex scene, which facilitates decoding object queries in the detector's decoder by providing a compact context. Additionally, the patch tokens in foundation models can enrich the features in the detector's encoder by providing semantic details. Utilizing frozen foundation models as plug-and-play modules rather than the commonly used backbone can significantly enhance the detector's performance while preventing the problems caused by the architecture discrepancy between the detector's backbone and the foundation model. With such a novel paradigm, we boost the SOTA query-based detector DINO from 49.0% AP to 51.9% AP (+2.9% AP) and further to 53.8% AP (+4.8% AP) by integrating one or two foundation models respectively, on the COCO validation set after training for 12 epochs with R50 as the detector's backbone.

Automated radiology report generation from 3D computed tomography (CT) volumes is challenging due to extreme sequence lengths, severe class imbalance, and the tendency of large language models (LLMs) to ignore visual tokens in favor of linguistic priors. We present Ker-VLJEPA-3B, a four-phase curriculum learning framework for free-text report generation from thoracic CT volumes. A phased training curriculum progressively adapts a Llama 3.2 3B decoder to ground its output in visual features from a frozen, self-supervised encoder. Our visual backbone (LeJEPA ViT-Large) is trained via self-supervised joint-embedding prediction on unlabeled CTs, without text supervision. Unlike contrastive models (CLIP, BiomedCLIP), this language-free backbone yields modality-pure representations. Vision-language alignment is deferred to the curriculum's bridge and generation phases. This modality-agnostic design can integrate any self-supervised encoder into an LLM without paired text during foundation training. Methodological innovations include: (1) zone-constrained cross-attention compressing slice embeddings into 32 spatially-grounded visual tokens; (2) PCA whitening of anisotropic LLM embeddings; (3) a positive-findings-only strategy eliminating posterior collapse; (4) warm bridge initialization transferring projection weights; and (5) selective cross-attention freezing with elastic weight consolidation to prevent catastrophic forgetting. Evaluated on the CT-RATE benchmark (2,984 validation volumes, 18 classes), Ker-VLJEPA-3B achieves a macro F1 of 0.429, surpassing the state-of-the-art (U-VLM, macro F1 = 0.414) by 3.6%, and reaching 0.448 (+8.2%) with threshold optimization. Ablation studies confirm 56.6% of generation quality derives from patient-specific visual content. Code and weights are available.

Text-to-image diffusion models understand spatial relationship between objects, but do they represent the true 3D structure of the world from only 2D supervision? We demonstrate that yes, 3D knowledge is encoded in 2D image diffusion models like Stable Diffusion, and we show that this structure can be exploited for 3D vision tasks. Our method, Viewpoint Neural Textual Inversion (ViewNeTI), controls the 3D viewpoint of objects in generated images from frozen diffusion models. We train a small neural mapper to take camera viewpoint parameters and predict text encoder latents; the latents then condition the diffusion generation process to produce images with the desired camera viewpoint. ViewNeTI naturally addresses Novel View Synthesis (NVS). By leveraging the frozen diffusion model as a prior, we can solve NVS with very few input views; we can even do single-view novel view synthesis. Our single-view NVS predictions have good semantic details and photorealism compared to prior methods. Our approach is well suited for modeling the uncertainty inherent in sparse 3D vision problems because it can efficiently generate diverse samples. Our view-control mechanism is general, and can even change the camera view in images generated by user-defined prompts.

Self-supervised image backbones can be used to address complex 2D tasks (e.g., semantic segmentation, object discovery) very efficiently and with little or no downstream supervision. Ideally, 3D backbones for lidar should be able to inherit these properties after distillation of these powerful 2D features. The most recent methods for image-to-lidar distillation on autonomous driving data show promising results, obtained thanks to distillation methods that keep improving. Yet, we still notice a large performance gap when measuring the quality of distilled and fully supervised features by linear probing. In this work, instead of focusing only on the distillation method, we study the effect of three pillars for distillation: the 3D backbone, the pretrained 2D backbones, and the pretraining dataset. In particular, thanks to our scalable distillation method named ScaLR, we show that scaling the 2D and 3D backbones and pretraining on diverse datasets leads to a substantial improvement of the feature quality. This allows us to significantly reduce the gap between the quality of distilled and fully-supervised 3D features, and to improve the robustness of the pretrained backbones to domain gaps and perturbations.

Transformers have recently gained significant attention in the computer vision community. However, the lack of scalability of self-attention mechanisms with respect to image size has limited their wide adoption in state-of-the-art vision backbones. In this paper we introduce an efficient and scalable attention model we call multi-axis attention, which consists of two aspects: blocked local and dilated global attention. These design choices allow global-local spatial interactions on arbitrary input resolutions with only linear complexity. We also present a new architectural element by effectively blending our proposed attention model with convolutions, and accordingly propose a simple hierarchical vision backbone, dubbed MaxViT, by simply repeating the basic building block over multiple stages. Notably, MaxViT is able to ''see'' globally throughout the entire network, even in earlier, high-resolution stages. We demonstrate the effectiveness of our model on a broad spectrum of vision tasks. On image classification, MaxViT achieves state-of-the-art performance under various settings: without extra data, MaxViT attains 86.5% ImageNet-1K top-1 accuracy; with ImageNet-21K pre-training, our model achieves 88.7% top-1 accuracy. For downstream tasks, MaxViT as a backbone delivers favorable performance on object detection as well as visual aesthetic assessment. We also show that our proposed model expresses strong generative modeling capability on ImageNet, demonstrating the superior potential of MaxViT blocks as a universal vision module. The source code and trained models will be available at https://github.com/google-research/maxvit.

Frozen-backbone transfer with Vision Transformers faces two under-addressed issues: optimization instability when adapters are naively inserted into a fixed feature extractor, and the absence of principled guidance for setting adapter capacity. We introduce AdapterTune, which augments each transformer block with a residual low-rank bottleneck whose up-projection is zero-initialized, guaranteeing that the adapted network starts exactly at the pretrained function and eliminates early-epoch representation drift. On the analytical side, we formalize adapter rank as a capacity budget for approximating downstream task shifts in feature space. The resulting excess-risk decomposition predicts monotonic but diminishing accuracy gains with increasing rank, an ``elbow'' behavior we confirm through controlled sweeps. We evaluate on 9 datasets and 3 backbone scales with multi-seed reporting throughout. On a core 5 dataset transfer suite, AdapterTune improves top-1 accuracy over head-only transfer by +14.9 points on average while training only 0.92 of the parameters required by full fine-tuning, and outperforms full fine-tuning on 10 of 15 dataset-backbone pairs. Across the full benchmark, AdapterTune improves over head-only transfer on every dataset-backbone pair tested. Ablations on rank, placement, and initialization isolate each design choice. The code is available at: https://github.com/salimkhazem/adaptertune

In this paper, we first assess and harness various Vision Foundation Models (VFMs) in the context of Domain Generalized Semantic Segmentation (DGSS). Driven by the motivation that Leveraging Stronger pre-trained models and Fewer trainable parameters for Superior generalizability, we introduce a robust fine-tuning approach, namely Rein, to parameter-efficiently harness VFMs for DGSS. Built upon a set of trainable tokens, each linked to distinct instances, Rein precisely refines and forwards the feature maps from each layer to the next layer within the backbone. This process produces diverse refinements for different categories within a single image. With fewer trainable parameters, Rein efficiently fine-tunes VFMs for DGSS tasks, surprisingly surpassing full parameter fine-tuning. Extensive experiments across various settings demonstrate that Rein significantly outperforms state-of-the-art methods. Remarkably, with just an extra 1% of trainable parameters within the frozen backbone, Rein achieves a mIoU of 78.4% on the Cityscapes, without accessing any real urban-scene datasets.Code is available at https://github.com/w1oves/Rein.git.

In this work, we introduce frozen-encoder model composition, a novel approach to multimodal embedding models. We build on the VLM-style architecture, in which non-text encoders are adapted to produce input for a language model, which in turn generates embeddings for all varieties of input. We present the result: the jina-embeddings-v5-omni suite, a pair of models that encode text, image, audio, and video input into a single semantic embedding space. Our method is to extend the two Jina Embeddings v5 Text models to support additional media by adding encoders for images and audio. The backbone text embedding models and the added non-text media encoders remain frozen. We only trained the connecting components, representing 0.35% of the total weights of the joint model. Training is therefore much more efficient than full-parameter retraining. Additionally, the language model remains effectively unaltered, producing exactly the same embeddings for text inputs as the Jina Embeddings v5 Text models. Our evaluations show that this approach produces results that are competitive with the state-of-the-art, yielding nearly equal performance to larger multimodal embedding models.

Multimodal Large Language Models (MLLMs) have recently achieved remarkable success in visual-language understanding, demonstrating superior high-level semantic alignment within their vision encoders. An important question thus arises: Can these encoders serve as versatile vision backbones, capable of reliably performing classic vision-centric tasks as well? To address the question, we make the following contributions: (i) we identify that the vision encoders within MLLMs exhibit deficiencies in their dense feature representations, as evidenced by their suboptimal performance on dense prediction tasks (e.g., semantic segmentation, depth estimation); (ii) we propose VersaViT, a well-rounded vision transformer that instantiates a novel multi-task framework for collaborative post-training. This framework facilitates the optimization of the vision backbone via lightweight task heads with multi-granularity supervision; (iii) extensive experiments across various downstream tasks demonstrate the effectiveness of our method, yielding a versatile vision backbone suited for both language-mediated reasoning and pixel-level understanding.

This paper delves into the interplay between vision backbones and optimizers, unvealing an inter-dependent phenomenon termed \textbf{backbone-optimizer coupling bias} (BOCB). We observe that canonical CNNs, such as VGG and ResNet, exhibit a marked co-dependency with SGD families, while recent architectures like ViTs and ConvNeXt share a tight coupling with the adaptive learning rate ones. We further show that BOCB can be introduced by both optimizers and certain backbone designs and may significantly impact the pre-training and downstream fine-tuning of vision models. Through in-depth empirical analysis, we summarize takeaways on recommended optimizers and insights into robust vision backbone architectures. We hope this work can inspire the community to question long-held assumptions on backbones and optimizers, stimulate further explorations, and thereby contribute to more robust vision systems. The source code and models are publicly available at https://bocb-ai.github.io/.

3D content creation has achieved significant progress in terms of both quality and speed. Although current feed-forward models can produce 3D objects in seconds, their resolution is constrained by the intensive computation required during training. In this paper, we introduce Large Multi-View Gaussian Model (LGM), a novel framework designed to generate high-resolution 3D models from text prompts or single-view images. Our key insights are two-fold: 1) 3D Representation: We propose multi-view Gaussian features as an efficient yet powerful representation, which can then be fused together for differentiable rendering. 2) 3D Backbone: We present an asymmetric U-Net as a high-throughput backbone operating on multi-view images, which can be produced from text or single-view image input by leveraging multi-view diffusion models. Extensive experiments demonstrate the high fidelity and efficiency of our approach. Notably, we maintain the fast speed to generate 3D objects within 5 seconds while boosting the training resolution to 512, thereby achieving high-resolution 3D content generation.

In the field of multimodal large language models (MLLMs), common methods typically involve unfreezing the language model during training to foster profound visual understanding. However, the fine-tuning of such models with vision-language data often leads to a diminution of their natural language processing (NLP) capabilities. To avoid this performance degradation, a straightforward solution is to freeze the language model while developing multimodal competencies. Unfortunately, previous works have not attained satisfactory outcomes. Building on the strategy of freezing the language model, we conduct thorough structural exploration and introduce the Inner-Adaptor Architecture (IAA). Specifically, the architecture incorporates multiple multimodal adaptors at varying depths within the large language model to facilitate direct interaction with the inherently text-oriented transformer layers, thereby enabling the frozen language model to acquire multimodal capabilities. Unlike previous approaches of freezing language models that require large-scale aligned data, our proposed architecture is able to achieve superior performance on small-scale datasets. We conduct extensive experiments to improve the general multimodal capabilities and visual grounding abilities of the MLLM. Our approach remarkably outperforms previous state-of-the-art methods across various vision-language benchmarks without sacrificing performance on NLP tasks. Code and models are available at https://github.com/360CVGroup/Inner-Adaptor-Architecture.

Diffusion-based text-to-image models have demonstrated impressive achievements in diversity and aesthetics but struggle to generate images with legible visual texts. Existing backbone models have limitations such as misspelling, failing to generate texts, and lack of support for Chinese text, but their development shows promising potential. In this paper, we propose a series of methods, aiming to empower backbone models to generate visual texts in English and Chinese. We first conduct a preliminary study revealing that Byte Pair Encoding (BPE) tokenization and the insufficient learning of cross-attention modules restrict the performance of the backbone models. Based on these observations, we make the following improvements: (1) We design a mixed granularity input strategy to provide more suitable text representations; (2) We propose to augment the conventional training objective with three glyph-aware training losses, which enhance the learning of cross-attention modules and encourage the model to focus on visual texts. Through experiments, we demonstrate that our methods can effectively empower backbone models to generate semantic relevant, aesthetically appealing, and accurate visual text images, while maintaining their fundamental image generation quality.

Large foundation models have emerged in the last years and are pushing performance boundaries for a variety of tasks. Training or even finetuning such models demands vast datasets and computational resources, which are often scarce and costly. Adaptation methods provide a computationally efficient solution to address these limitations by allowing such models to be finetuned on small amounts of data and computing power. This is achieved by appending new trainable modules to frozen backbones with only a fraction of the trainable parameters and fitting only these modules on novel tasks. Recently, the VeRA adapter was shown to excel in parameter-efficient adaptations by utilizing a pair of frozen random low-rank matrices shared across all layers. In this paper, we propose PVeRA, a probabilistic version of the VeRA adapter, which modifies the low-rank matrices of VeRA in a probabilistic manner. This modification naturally allows handling inherent ambiguities in the input and allows for different sampling configurations during training and testing. A comprehensive evaluation was performed on the VTAB-1k benchmark and seven adapters, with PVeRA outperforming VeRA and other adapters. Our code for training models with PVeRA and benchmarking all adapters is available https://github.com/leofillioux/pvera.

Recently the state space models (SSMs) with efficient hardware-aware designs, i.e., Mamba, have shown great potential for long sequence modeling. Building efficient and generic vision backbones purely upon SSMs is an appealing direction. However, representing visual data is challenging for SSMs due to the position-sensitivity of visual data and the requirement of global context for visual understanding. In this paper, we show that the reliance of visual representation learning on self-attention is not necessary and propose a new generic vision backbone with bidirectional Mamba blocks (Vim), which marks the image sequences with position embeddings and compresses the visual representation with bidirectional state space models. On ImageNet classification, COCO object detection, and ADE20k semantic segmentation tasks, Vim achieves higher performance compared to well-established vision transformers like DeiT, while also demonstrating significantly improved computation & memory efficiency. For example, Vim is 2.8times faster than DeiT and saves 86.8% GPU memory when performing batch inference to extract features on images with a resolution of 1248times1248. The results demonstrate that Vim is capable of overcoming the computation & memory constraints on performing Transformer-style understanding for high-resolution images and it has great potential to become the next-generation backbone for vision foundation models. Code is available at https://github.com/hustvl/Vim.

Multi-camera 3D object detection for autonomous driving is a challenging problem that has garnered notable attention from both academia and industry. An obstacle encountered in vision-based techniques involves the precise extraction of geometry-conscious features from RGB images. Recent approaches have utilized geometric-aware image backbones pretrained on depth-relevant tasks to acquire spatial information. However, these approaches overlook the critical aspect of view transformation, resulting in inadequate performance due to the misalignment of spatial knowledge between the image backbone and view transformation. To address this issue, we propose a novel geometric-aware pretraining framework called GAPretrain. Our approach incorporates spatial and structural cues to camera networks by employing the geometric-rich modality as guidance during the pretraining phase. The transference of modal-specific attributes across different modalities is non-trivial, but we bridge this gap by using a unified bird's-eye-view (BEV) representation and structural hints derived from LiDAR point clouds to facilitate the pretraining process. GAPretrain serves as a plug-and-play solution that can be flexibly applied to multiple state-of-the-art detectors. Our experiments demonstrate the effectiveness and generalization ability of the proposed method. We achieve 46.2 mAP and 55.5 NDS on the nuScenes val set using the BEVFormer method, with a gain of 2.7 and 2.1 points, respectively. We also conduct experiments on various image backbones and view transformations to validate the efficacy of our approach. Code will be released at https://github.com/OpenDriveLab/BEVPerception-Survey-Recipe.

Foley Control is a lightweight approach to video-guided Foley that keeps pretrained single-modality models frozen and learns only a small cross-attention bridge between them. We connect V-JEPA2 video embeddings to a frozen Stable Audio Open DiT text-to-audio (T2A) model by inserting compact video cross-attention after the model's existing text cross-attention, so prompts set global semantics while video refines timing and local dynamics. The frozen backbones retain strong marginals (video; audio given text) and the bridge learns the audio-video dependency needed for synchronization -- without retraining the audio prior. To cut memory and stabilize training, we pool video tokens before conditioning. On curated video-audio benchmarks, Foley Control delivers competitive temporal and semantic alignment with far fewer trainable parameters than recent multi-modal systems, while preserving prompt-driven controllability and production-friendly modularity (swap/upgrade encoders or the T2A backbone without end-to-end retraining). Although we focus on Video-to-Foley, the same bridge design can potentially extend to other audio modalities (e.g., speech).

Existing dynamic scene generation methods mostly rely on distilling knowledge from pre-trained 3D generative models, which are typically fine-tuned on synthetic object datasets. As a result, the generated scenes are often object-centric and lack photorealism. To address these limitations, we introduce a novel pipeline designed for photorealistic text-to-4D scene generation, discarding the dependency on multi-view generative models and instead fully utilizing video generative models trained on diverse real-world datasets. Our method begins by generating a reference video using the video generation model. We then learn the canonical 3D representation of the video using a freeze-time video, delicately generated from the reference video. To handle inconsistencies in the freeze-time video, we jointly learn a per-frame deformation to model these imperfections. We then learn the temporal deformation based on the canonical representation to capture dynamic interactions in the reference video. The pipeline facilitates the generation of dynamic scenes with enhanced photorealism and structural integrity, viewable from multiple perspectives, thereby setting a new standard in 4D scene generation.

This paper reveals that large language models (LLMs), despite being trained solely on textual data, are surprisingly strong encoders for purely visual tasks in the absence of language. Even more intriguingly, this can be achieved by a simple yet previously overlooked strategy -- employing a frozen transformer block from pre-trained LLMs as a constituent encoder layer to directly process visual tokens. Our work pushes the boundaries of leveraging LLMs for computer vision tasks, significantly departing from conventional practices that typically necessitate a multi-modal vision-language setup with associated language prompts, inputs, or outputs. We demonstrate that our approach consistently enhances performance across a diverse range of tasks, encompassing pure 2D and 3D visual recognition tasks (e.g., image and point cloud classification), temporal modeling tasks (e.g., action recognition), non-semantic tasks (e.g., motion forecasting), and multi-modal tasks (e.g., 2D/3D visual question answering and image-text retrieval). Such improvements are a general phenomenon, applicable to various types of LLMs (e.g., LLaMA and OPT) and different LLM transformer blocks. We additionally propose the information filtering hypothesis to explain the effectiveness of pre-trained LLMs in visual encoding -- the pre-trained LLM transformer blocks discern informative visual tokens and further amplify their effect. This hypothesis is empirically supported by the observation that the feature activation, after training with LLM transformer blocks, exhibits a stronger focus on relevant regions. We hope that our work inspires new perspectives on utilizing LLMs and deepening our understanding of their underlying mechanisms. Code is available at https://github.com/ziqipang/LM4VisualEncoding.

The success of large language models (LLMs) has fostered a new research trend of multi-modality large language models (MLLMs), which changes the paradigm of various fields in computer vision. Though MLLMs have shown promising results in numerous high-level vision and vision-language tasks such as VQA and text-to-image, no works have demonstrated how low-level vision tasks can benefit from MLLMs. We find that most current MLLMs are blind to low-level features due to their design of vision modules, thus are inherently incapable for solving low-level vision tasks. In this work, we purpose LM4LV, a framework that enables a FROZEN LLM to solve a range of low-level vision tasks without any multi-modal data or prior. This showcases the LLM's strong potential in low-level vision and bridges the gap between MLLMs and low-level vision tasks. We hope this work can inspire new perspectives on LLMs and deeper understanding of their mechanisms.

This paper introduces VLAP, a novel approach that bridges pretrained vision models and large language models (LLMs) to make frozen LLMs understand the visual world. VLAP transforms the embedding space of pretrained vision models into the LLMs' word embedding space using a single linear layer for efficient and general-purpose visual and language understanding. Specifically, we harness well-established word embeddings to bridge two modality embedding spaces. The visual and text representations are simultaneously assigned to a set of word embeddings within pretrained LLMs by formulating the assigning procedure as an optimal transport problem. We predict the assignment of one modality from the representation of another modality data, enforcing consistent assignments for paired multimodal data. This allows vision and language representations to contain the same information, grounding the frozen LLMs' word embedding space in visual data. Moreover, a robust semantic taxonomy of LLMs can be preserved with visual data since the LLMs interpret and reason linguistic information from correlations between word embeddings. Experimental results show that VLAP achieves substantial improvements over the previous linear transformation-based approaches across a range of vision-language tasks, including image captioning, visual question answering, and cross-modal retrieval. We also demonstrate the learned visual representations hold a semantic taxonomy of LLMs, making visual semantic arithmetic possible.

The use of pretrained backbones with fine-tuning has been successful for 2D vision and natural language processing tasks, showing advantages over task-specific networks. In this work, we introduce a pretrained 3D backbone, called {\SST}, for 3D indoor scene understanding. We design a 3D Swin transformer as our backbone network, which enables efficient self-attention on sparse voxels with linear memory complexity, making the backbone scalable to large models and datasets. We also introduce a generalized contextual relative positional embedding scheme to capture various irregularities of point signals for improved network performance. We pretrained a large {\SST} model on a synthetic Structured3D dataset, which is an order of magnitude larger than the ScanNet dataset. Our model pretrained on the synthetic dataset not only generalizes well to downstream segmentation and detection on real 3D point datasets, but also outperforms state-of-the-art methods on downstream tasks with +2.3 mIoU and +2.2 mIoU on S3DIS Area5 and 6-fold semantic segmentation, +1.8 mIoU on ScanNet segmentation (val), +1.9 mAP@0.5 on ScanNet detection, and +8.1 mAP@0.5 on S3DIS detection. A series of extensive ablation studies further validate the scalability, generality, and superior performance enabled by our approach. The code and models are available at https://github.com/microsoft/Swin3D .

The feature maps of vision encoders are fundamental to myriad modern AI tasks, ranging from core perception algorithms (e.g. semantic segmentation, object detection, depth perception, etc.) to modern multimodal understanding in vision-language models (VLMs). Currently, in computer vision, the frontier of general purpose vision backbones is Vision Transformers (ViT), typically trained using contrastive loss (e.g. CLIP). A key problem with most off-the-shelf ViTs, particularly CLIP, is that these models are inflexibly low resolution. Most run at 224 times 224px, while the "high-resolution" versions are around 378-448px, but still inflexible. We introduce a novel method to coherently and cheaply upsample the feature maps of low-resolution vision encoders while picking up on fine-grained details that would otherwise be lost due to resolution. We demonstrate the effectiveness of this approach on core perception tasks as well as within agglomerative model training using RADIO as a way of providing richer targets for distillation. Code available at https://github.com/NVlabs/FeatSharp .

Being data-driven is one of the most iconic properties of deep learning algorithms. The birth of ImageNet drives a remarkable trend of "learning from large-scale data" in computer vision. Pretraining on ImageNet to obtain rich universal representations has been manifested to benefit various 2D visual tasks, and becomes a standard in 2D vision. However, due to the laborious collection of real-world 3D data, there is yet no generic dataset serving as a counterpart of ImageNet in 3D vision, thus how such a dataset can impact the 3D community is unraveled. To remedy this defect, we introduce MVImgNet, a large-scale dataset of multi-view images, which is highly convenient to gain by shooting videos of real-world objects in human daily life. It contains 6.5 million frames from 219,188 videos crossing objects from 238 classes, with rich annotations of object masks, camera parameters, and point clouds. The multi-view attribute endows our dataset with 3D-aware signals, making it a soft bridge between 2D and 3D vision. We conduct pilot studies for probing the potential of MVImgNet on a variety of 3D and 2D visual tasks, including radiance field reconstruction, multi-view stereo, and view-consistent image understanding, where MVImgNet demonstrates promising performance, remaining lots of possibilities for future explorations. Besides, via dense reconstruction on MVImgNet, a 3D object point cloud dataset is derived, called MVPNet, covering 87,200 samples from 150 categories, with the class label on each point cloud. Experiments show that MVPNet can benefit the real-world 3D object classification while posing new challenges to point cloud understanding. MVImgNet and MVPNet will be publicly available, hoping to inspire the broader vision community.

In contrast to numerous NLP and 2D vision foundational models, learning a 3D foundational model poses considerably greater challenges. This is primarily due to the inherent data variability and diversity of downstream tasks. In this paper, we introduce a novel universal 3D pre-training framework designed to facilitate the acquisition of efficient 3D representation, thereby establishing a pathway to 3D foundational models. Considering that informative 3D features should encode rich geometry and appearance cues that can be utilized to render realistic images, we propose to learn 3D representations by differentiable neural rendering. We train a 3D backbone with a devised volumetric neural renderer by comparing the rendered with the real images. Notably, our approach seamlessly integrates the learned 3D encoder into various downstream tasks. These tasks encompass not only high-level challenges such as 3D detection and segmentation but also low-level objectives like 3D reconstruction and image synthesis, spanning both indoor and outdoor scenarios. Besides, we also illustrate the capability of pre-training a 2D backbone using the proposed methodology, surpassing conventional pre-training methods by a large margin. For the first time, PonderV2 achieves state-of-the-art performance on 11 indoor and outdoor benchmarks, implying its effectiveness. Code and models are available at https://github.com/OpenGVLab/PonderV2.

Current 3D content generation builds on generative models that output RGB images. Modern graphics pipelines, however, require physically-based rendering (PBR) material properties. We propose to model the PBR image distribution directly to avoid photometric inaccuracies in RGB generation and the inherent ambiguity in extracting PBR from RGB. Existing paradigms for cross-modal finetuning are not suited for PBR generation due to a lack of data and the high dimensionality of the output modalities: we overcome both challenges by retaining a frozen RGB model and tightly linking a newly trained PBR model using a novel cross-network communication paradigm. As the base RGB model is fully frozen, the proposed method does not risk catastrophic forgetting during finetuning and remains compatible with techniques such as IPAdapter pretrained for the base RGB model. We validate our design choices, robustness to data sparsity, and compare against existing paradigms with an extensive experimental section.

We present a novel bird's-eye-view (BEV) detector with perspective supervision, which converges faster and better suits modern image backbones. Existing state-of-the-art BEV detectors are often tied to certain depth pre-trained backbones like VoVNet, hindering the synergy between booming image backbones and BEV detectors. To address this limitation, we prioritize easing the optimization of BEV detectors by introducing perspective space supervision. To this end, we propose a two-stage BEV detector, where proposals from the perspective head are fed into the bird's-eye-view head for final predictions. To evaluate the effectiveness of our model, we conduct extensive ablation studies focusing on the form of supervision and the generality of the proposed detector. The proposed method is verified with a wide spectrum of traditional and modern image backbones and achieves new SoTA results on the large-scale nuScenes dataset. The code shall be released soon.

We address the problem of few-shot pattern detection, which aims to detect all instances of a given pattern, typically represented by a few exemplars, from an input image. Although similar problems have been studied in few-shot object counting and detection (FSCD), previous methods and their benchmarks have narrowed patterns of interest to object categories and often fail to localize non-object patterns. In this work, we propose a simple yet effective detector based on template matching and regression, dubbed TMR. While previous FSCD methods typically represent target exemplars as spatially collapsed prototypes and lose structural information, we revisit classic template matching and regression. It effectively preserves and leverages the spatial layout of exemplars through a minimalistic structure with a small number of learnable convolutional or projection layers on top of a frozen backbone We also introduce a new dataset, dubbed RPINE, which covers a wider range of patterns than existing object-centric datasets. Our method outperforms the state-of-the-art methods on the three benchmarks, RPINE, FSCD-147, and FSCD-LVIS, and demonstrates strong generalization in cross-dataset evaluation.

Contrastive language image pretraining (CLIP) encoders have been shown to be beneficial for a range of visual tasks from classification and detection to captioning and image manipulation. We investigate the effectiveness of CLIP visual backbones for Embodied AI tasks. We build incredibly simple baselines, named EmbCLIP, with no task specific architectures, inductive biases (such as the use of semantic maps), auxiliary tasks during training, or depth maps -- yet we find that our improved baselines perform very well across a range of tasks and simulators. EmbCLIP tops the RoboTHOR ObjectNav leaderboard by a huge margin of 20 pts (Success Rate). It tops the iTHOR 1-Phase Rearrangement leaderboard, beating the next best submission, which employs Active Neural Mapping, and more than doubling the % Fixed Strict metric (0.08 to 0.17). It also beats the winners of the 2021 Habitat ObjectNav Challenge, which employ auxiliary tasks, depth maps, and human demonstrations, and those of the 2019 Habitat PointNav Challenge. We evaluate the ability of CLIP's visual representations at capturing semantic information about input observations -- primitives that are useful for navigation-heavy embodied tasks -- and find that CLIP's representations encode these primitives more effectively than ImageNet-pretrained backbones. Finally, we extend one of our baselines, producing an agent capable of zero-shot object navigation that can navigate to objects that were not used as targets during training. Our code and models are available at https://github.com/allenai/embodied-clip

Leveraging the powerful representations of pre-trained vision foundation models -- traditionally used for visual comprehension -- we explore a novel direction: building an image tokenizer directly atop such models, a largely underexplored area. Specifically, we employ a frozen vision foundation model as the encoder of our tokenizer. To enhance its effectiveness, we introduce two key components: (1) a region-adaptive quantization framework that reduces redundancy in the pre-trained features on regular 2D grids, and (2) a semantic reconstruction objective that aligns the tokenizer's outputs with the foundation model's representations to preserve semantic fidelity. Based on these designs, our proposed image tokenizer, VFMTok, achieves substantial improvements in image reconstruction and generation quality, while also enhancing token efficiency. It further boosts autoregressive (AR) generation -- achieving a gFID of 2.07 on ImageNet benchmarks, while accelerating model convergence by three times, and enabling high-fidelity class-conditional synthesis without the need for classifier-free guidance (CFG). The code will be released publicly to benefit the community.

We present a novel methodology aimed at optimizing the application of frozen large language models (LLMs) for resource-intensive vision-language (VL) pre-training. The current paradigm uses visual features as prompts to guide language models, with a focus on determining the most relevant visual features for corresponding text. Our approach diverges by concentrating on the language component, specifically identifying the optimal prompts to align with visual features. We introduce the Prompt-Transformer (P-Former), a model that predicts these ideal prompts, which is trained exclusively on linguistic data, bypassing the need for image-text pairings. This strategy subtly bifurcates the end-to-end VL training process into an additional, separate stage. Our experiments reveal that our framework significantly enhances the performance of a robust image-to-text baseline (BLIP-2), and effectively narrows the performance gap between models trained with either 4M or 129M image-text pairs. Importantly, our framework is modality-agnostic and flexible in terms of architectural design, as validated by its successful application in a video learning task using varied base modules. The code is available at https://github.com/yiren-jian/BLIText

The combination of strong visual backbones and Large Language Model (LLM) reasoning has led to Large Multimodal Models (LMMs) becoming the current standard for a wide range of vision and language (VL) tasks. However, recent research has shown that even the most advanced LMMs still struggle to capture aspects of compositional visual reasoning, such as attributes and relationships between objects. One solution is to utilize scene graphs (SGs)--a formalization of objects and their relations and attributes that has been extensively used as a bridge between the visual and textual domains. Yet, scene graph data requires scene graph annotations, which are expensive to collect and thus not easily scalable. Moreover, finetuning an LMM based on SG data can lead to catastrophic forgetting of the pretraining objective. To overcome this, inspired by chain-of-thought methods, we propose Compositional Chain-of-Thought (CCoT), a novel zero-shot Chain-of-Thought prompting method that utilizes SG representations in order to extract compositional knowledge from an LMM. Specifically, we first generate an SG using the LMM, and then use that SG in the prompt to produce a response. Through extensive experiments, we find that the proposed CCoT approach not only improves LMM performance on several vision and language VL compositional benchmarks but also improves the performance of several popular LMMs on general multimodal benchmarks, without the need for fine-tuning or annotated ground-truth SGs. Code: https://github.com/chancharikmitra/CCoT

Vision foundation models (VFMs) have emerged as powerful tools for surgical scene understanding. However, current approaches predominantly rely on unimodal RGB pre-training, overlooking the complex 3D geometry inherent to surgical environments. Although several architectures support multimodal or geometry-aware inputs in general computer vision, the benefits of incorporating depth information in surgical settings remain underexplored. We conduct a large-scale empirical study comparing eight ViT-based VFMs that differ in pre-training domain, learning objective, and input modality (RGB vs. RGB-D). For pre-training, we use a curated dataset of 1.4 million robotic surgical images paired with depth maps generated from an off-the-shelf network. We evaluate these models under both frozen-backbone and end-to-end fine-tuning protocols across eight surgical datasets spanning object detection, segmentation, depth estimation, and pose estimation. Our experiments yield several consistent findings. Models incorporating explicit geometric tokenization, such as MultiMAE, substantially outperform unimodal baselines across all tasks. Notably, geometric-aware pre-training enables remarkable data efficiency: models fine-tuned on just 25% of labeled data consistently surpass RGB-only models trained on the full dataset. Importantly, these gains require no architectural or runtime changes at inference; depth is used only during pre-training, making adoption straightforward. These findings suggest that multimodal pre-training offers a viable path towards building more capable surgical vision systems.

Current computer vision models, unlike the human visual system, cannot yet achieve general-purpose visual understanding. Existing efforts to create a general vision model are limited in the scope of assessed tasks and offer no overarching framework to perform them holistically. We present a new comprehensive benchmark, General-purpose Visual Understanding Evaluation (G-VUE), covering the full spectrum of visual cognitive abilities with four functional domains x2014 Perceive, Ground, Reason, and Act. The four domains are embodied in 11 carefully curated tasks, from 3D reconstruction to visual reasoning and manipulation. Along with the benchmark, we provide a general encoder-decoder framework to allow for the evaluation of arbitrary visual representation on all 11 tasks. We evaluate various pre-trained visual representations with our framework and observe that (1) Transformer-based visual backbone generally outperforms CNN-based backbone on G-VUE, (2) visual representations from vision-language pre-training are superior to those with vision-only pre-training across visual tasks. With G-VUE, we provide a holistic evaluation standard to motivate research toward building general-purpose visual systems via obtaining more general-purpose visual representations.

Multi-view visual reasoning is essential for intelligent systems that must understand complex environments from sparse and discrete viewpoints, yet existing research has largely focused on single-image or temporally dense video settings. In real-world scenarios, reasoning across views requires integrating partial observations without explicit guidance, while collecting large-scale multi-view data with accurate geometric and semantic annotations remains challenging. To address this gap, we leverage physically grounded simulation to construct diverse, high-fidelity 3D scenes with precise per-view metadata, enabling scalable data generation that remains transferable to real-world settings. Based on this engine, we introduce VIEW2SPACE, a multi-dimensional benchmark for sparse multi-view reasoning, together with a scalable, disjoint training split supporting millions of grounded question-answer pairs. Using this benchmark, a comprehensive evaluation of state-of-the-art vision-language and spatial models reveals that multi-view reasoning remains largely unsolved, with most models performing only marginally above random guessing. We further investigate whether training can bridge this gap. Our proposed Grounded Chain-of-Thought with Visual Evidence substantially improves performance under moderate difficulty, and generalizes to real-world data, outperforming existing approaches in cross-dataset evaluation. We further conduct difficulty-aware scaling analyses across model size, data scale, reasoning depth, and visibility constraints, indicating that while geometric perception can benefit from scaling under sufficient visibility, deep compositional reasoning across sparse views remains a fundamental challenge.

Training-free open-vocabulary semantic segmentation (OVSS) promises rapid adaptation to new label sets without retraining. Yet, many methods rely on heavy post-processing or handle text and vision in isolation, leaving cross-modal geometry underutilized. Others introduce auxiliary vision backbones or multi-model pipelines, which increase complexity and latency while compromising design simplicity. We present PEARL, \underline{P}rocrust\underline{e}s \underline{a}lignment with text-awa\underline{r}e \underline{L}aplacian propagation, a compact two-step inference that follows an align-then-propagate principle. The Procrustes alignment step performs an orthogonal projection inside the last self-attention block, rotating keys toward the query subspace via a stable polar iteration. The text-aware Laplacian propagation then refines per-pixel logits on a small grid through a confidence-weighted, text-guided graph solve: text provides both a data-trust signal and neighbor gating, while image gradients preserve boundaries. In this work, our method is fully training-free, plug-and-play, and uses only fixed constants, adding minimal latency with a small per-head projection and a few conjugate-gradient steps. Our approach, PEARL, sets a new state-of-the-art in training-free OVSS without extra data or auxiliary backbones across standard benchmarks, achieving superior performance under both with-background and without-background protocols.

Zero-shot captioners are recently proposed models that utilize common-space vision-language representations to caption images without relying on paired image-text data. To caption an image, they proceed by textually decoding a text-aligned image feature, but they limit their scope to global representations and whole-image captions. We present , a unified framework for zero-shot captioning that shifts from an image-centric to a patch-centric paradigm, enabling the captioning of arbitrary regions without the need of region-level supervision. Instead of relying on global image representations, we treat individual patches as atomic captioning units and aggregate them to describe arbitrary regions, from single patches to non-contiguous areas and entire images. We analyze the key ingredients that enable current latent captioners to work in our novel proposed framework. Experiments demonstrate that backbones producing meaningful, dense visual features, such as DINO, are key to achieving state-of-the-art performance in multiple region-based captioning tasks. Compared to other baselines and state-of-the-art competitors, our models achieve better performance on zero-shot dense, region-set, and a newly introduced trace captioning task, highlighting the effectiveness of patch-wise semantic representations for scalable caption generation. Project page at https://paciosoft.com/Patch-ioner/ .

In this paper, we present the Semantic Boundary Conditioned Backbone (SBCB) framework, a simple yet effective training framework that is model-agnostic and boosts segmentation performance, especially around the boundaries. Motivated by the recent development in improving semantic segmentation by incorporating boundaries as auxiliary tasks, we propose a multi-task framework that uses semantic boundary detection (SBD) as an auxiliary task. The SBCB framework utilizes the nature of the SBD task, which is complementary to semantic segmentation, to improve the backbone of the segmentation head. We apply an SBD head that exploits the multi-scale features from the backbone, where the model learns low-level features in the earlier stages, and high-level semantic understanding in the later stages. This head perfectly complements the common semantic segmentation architectures where the features from the later stages are used for classification. We can improve semantic segmentation models without additional parameters during inference by only conditioning the backbone. Through extensive evaluations, we show the effectiveness of the SBCB framework by improving various popular segmentation heads and backbones by 0.5% ~ 3.0% IoU on the Cityscapes dataset and gains 1.6% ~ 4.1% in boundary Fscores. We also apply this framework on customized backbones and the emerging vision transformer models and show the effectiveness of the SBCB framework.

This work investigates a simple yet powerful dense prediction task adapter for Vision Transformer (ViT). Unlike recently advanced variants that incorporate vision-specific inductive biases into their architectures, the plain ViT suffers inferior performance on dense predictions due to weak prior assumptions. To address this issue, we propose the ViT-Adapter, which allows plain ViT to achieve comparable performance to vision-specific transformers. Specifically, the backbone in our framework is a plain ViT that can learn powerful representations from large-scale multi-modal data. When transferring to downstream tasks, a pre-training-free adapter is used to introduce the image-related inductive biases into the model, making it suitable for these tasks. We verify ViT-Adapter on multiple dense prediction tasks, including object detection, instance segmentation, and semantic segmentation. Notably, without using extra detection data, our ViT-Adapter-L yields state-of-the-art 60.9 box AP and 53.0 mask AP on COCO test-dev. We hope that the ViT-Adapter could serve as an alternative for vision-specific transformers and facilitate future research. The code and models will be released at https://github.com/czczup/ViT-Adapter.

Multi-level feature fusion is a fundamental topic in computer vision. It has been exploited to detect, segment and classify objects at various scales. When multi-level features meet multi-modal cues, the optimal feature aggregation and multi-modal learning strategy become a hot potato. In this paper, we leverage the inherent multi-modal and multi-level nature of RGB-D salient object detection to devise a novel cascaded refinement network. In particular, first, we propose to regroup the multi-level features into teacher and student features using a bifurcated backbone strategy (BBS). Second, we introduce a depth-enhanced module (DEM) to excavate informative depth cues from the channel and spatial views. Then, RGB and depth modalities are fused in a complementary way. Our architecture, named Bifurcated Backbone Strategy Network (BBS-Net), is simple, efficient, and backbone-independent. Extensive experiments show that BBS-Net significantly outperforms eighteen SOTA models on eight challenging datasets under five evaluation measures, demonstrating the superiority of our approach (sim 4 % improvement in S-measure vs. the top-ranked model: DMRA-iccv2019). In addition, we provide a comprehensive analysis on the generalization ability of different RGB-D datasets and provide a powerful training set for future research.

Vision-Language-Action (VLA) models are driving rapid progress in robotics by enabling agents to interpret multimodal inputs and execute complex, long-horizon tasks. However, their safety and robustness against adversarial attacks remain largely underexplored. In this work, we identify and formalize a critical adversarial vulnerability in which adversarial images can "freeze" VLA models and cause them to ignore subsequent instructions. This threat effectively disconnects the robot's digital mind from its physical actions, potentially inducing inaction during critical interventions. To systematically study this vulnerability, we propose FreezeVLA, a novel attack framework that generates and evaluates action-freezing attacks via min-max bi-level optimization. Experiments on three state-of-the-art VLA models and four robotic benchmarks show that FreezeVLA attains an average attack success rate of 76.2%, significantly outperforming existing methods. Moreover, adversarial images generated by FreezeVLA exhibit strong transferability, with a single image reliably inducing paralysis across diverse language prompts. Our findings expose a critical safety risk in VLA models and highlight the urgent need for robust defense mechanisms.

Monocular depth estimation methods traditionally train deep models to infer depth directly from RGB pixels. This implicit learning often overlooks explicit monocular cues that the human visual system relies on, such as occlusion boundaries, shading, and perspective. Rather than expecting a network to discover these cues unaided, we present ThirdEye, a cue-aware pipeline that deliberately supplies each cue through specialised, pre-trained, and frozen networks. These cues are fused in a three-stage cortical hierarchy (V1->V2->V3) equipped with a key-value working-memory module that weights them by reliability. An adaptive-bins transformer head then produces a high-resolution disparity map. Because the cue experts are frozen, ThirdEye inherits large amounts of external supervision while requiring only modest fine-tuning. This extended version provides additional architectural detail, neuroscientific motivation, and an expanded experimental protocol; quantitative results will appear in a future revision.

Graph Neural Networks operate through bottom-up message-passing, fundamentally differing from human visual perception, which intuitively captures global structures first. We investigate the underappreciated potential of vision models for graph understanding, finding they achieve performance comparable to GNNs on established benchmarks while exhibiting distinctly different learning patterns. These divergent behaviors, combined with limitations of existing benchmarks that conflate domain features with topological understanding, motivate our introduction of GraphAbstract. This benchmark evaluates models' ability to perceive global graph properties as humans do: recognizing organizational archetypes, detecting symmetry, sensing connectivity strength, and identifying critical elements. Our results reveal that vision models significantly outperform GNNs on tasks requiring holistic structural understanding and maintain generalizability across varying graph scales, while GNNs struggle with global pattern abstraction and degrade with increasing graph size. This work demonstrates that vision models possess remarkable yet underutilized capabilities for graph structural understanding, particularly for problems requiring global topological awareness and scale-invariant reasoning. These findings open new avenues to leverage this underappreciated potential for developing more effective graph foundation models for tasks dominated by holistic pattern recognition.

Methods based on diffusion backbones have recently revolutionized novel view synthesis (NVS). However, those models require pretrained 2D diffusion checkpoints (e.g., Stable Diffusion) as the basis for geometrical priors. Since such checkpoints require exorbitant amounts of data and compute to train, this greatly limits the scalability of diffusion-based NVS models. We present Next-Scale Autoregression Conditioned by View (ArchonView), a method that significantly exceeds state-of-the-art methods despite being trained from scratch with 3D rendering data only and no 2D pretraining. We achieve this by incorporating both global (pose-augmented semantics) and local (multi-scale hierarchical encodings) conditioning into a backbone based on the next-scale autoregression paradigm. Our model also exhibits robust performance even for difficult camera poses where previous methods fail, and is several times faster in inference speed compared to diffusion. We experimentally verify that performance scales with model and dataset size, and conduct extensive demonstration of our method's synthesis quality across several tasks. Our code is open-sourced at https://github.com/Shiran-Yuan/ArchonView.

We present ProtoViT, a method for interpretable image classification combining deep learning and case-based reasoning. This method classifies an image by comparing it to a set of learned prototypes, providing explanations of the form ``this looks like that.'' In our model, a prototype consists of parts, which can deform over irregular geometries to create a better comparison between images. Unlike existing models that rely on Convolutional Neural Network (CNN) backbones and spatially rigid prototypes, our model integrates Vision Transformer (ViT) backbones into prototype based models, while offering spatially deformed prototypes that not only accommodate geometric variations of objects but also provide coherent and clear prototypical feature representations with an adaptive number of prototypical parts. Our experiments show that our model can generally achieve higher performance than the existing prototype based models. Our comprehensive analyses ensure that the prototypes are consistent and the interpretations are faithful.

We present a method to create diffusion-based video models from pretrained Text-to-Image (T2I) models. Recently, AnimateDiff proposed freezing the T2I model while only training temporal layers. We advance this method by proposing a unique architecture, incorporating a mapping network and frame-wise tokens, tailored for video generation while maintaining the diversity and creativity of the original T2I model. Key innovations include novel loss functions for temporal smoothness and a mitigating gradient sampling technique, ensuring realistic and temporally consistent video generation despite limited public video data. We have successfully integrated video-specific inductive biases into the architecture and loss functions. Our method, built on the frozen StableDiffusion model, simplifies training processes and allows for seamless integration with off-the-shelf models like ControlNet and DreamBooth. project page: https://kwonminki.github.io/HARIVO

3D foundation models (3DFMs) have recently transformed 3D vision, enabling joint prediction of depths, poses, and point maps directly from images. Yet their ability to reason under extreme, non-overlapping views remains largely unexplored. In this work, we study their internal representations and find that 3DFMs exhibit an emergent understanding of extreme-view geometry, despite never being trained for such conditions. To further enhance these capabilities, we introduce a lightweight alignment scheme that refines their internal 3D representation by tuning only a small subset of backbone bias terms, leaving all decoder heads frozen. This targeted adaptation substantially improves relative pose estimation under extreme viewpoints without degrading per-image depth or point quality. Additionally, we contribute MegaUnScene, a new benchmark of Internet scenes unseen by existing 3DFMs, with dedicated test splits for both relative pose estimation and dense 3D reconstruction. All code and data will be released.

Although using convolutional neural networks (CNNs) as backbones achieves great successes in computer vision, this work investigates a simple backbone network useful for many dense prediction tasks without convolutions. Unlike the recently-proposed Transformer model (e.g., ViT) that is specially designed for image classification, we propose Pyramid Vision Transformer~(PVT), which overcomes the difficulties of porting Transformer to various dense prediction tasks. PVT has several merits compared to prior arts. (1) Different from ViT that typically has low-resolution outputs and high computational and memory cost, PVT can be not only trained on dense partitions of the image to achieve high output resolution, which is important for dense predictions but also using a progressive shrinking pyramid to reduce computations of large feature maps. (2) PVT inherits the advantages from both CNN and Transformer, making it a unified backbone in various vision tasks without convolutions by simply replacing CNN backbones. (3) We validate PVT by conducting extensive experiments, showing that it boosts the performance of many downstream tasks, e.g., object detection, semantic, and instance segmentation. For example, with a comparable number of parameters, RetinaNet+PVT achieves 40.4 AP on the COCO dataset, surpassing RetinNet+ResNet50 (36.3 AP) by 4.1 absolute AP. We hope PVT could serve as an alternative and useful backbone for pixel-level predictions and facilitate future researches. Code is available at https://github.com/whai362/PVT.

Recent contrastive multimodal vision-language models like CLIP have demonstrated robust open-world semantic understanding, becoming the standard image backbones for vision-language applications due to their aligned latent space. However, this practice has left powerful unimodal encoders for both vision and language underutilized in multimodal applications which raises a key question: Is there a plausible way to connect unimodal backbones for zero-shot vision-language tasks? To this end, we propose a novel approach that aligns vision and language modalities using only projection layers on pretrained, frozen unimodal encoders. Our method exploits the high semantic similarity between embedding spaces of well-trained vision and language models. It involves selecting semantically similar encoders in the latent space, curating a concept-rich dataset of image-caption pairs, and training simple MLP projectors. We evaluated our approach on 12 zero-shot classification datasets and 2 image-text retrieval datasets. Our best model, utilizing DINOv2 and All-Roberta-Large text encoder, achieves 76\(\%\) accuracy on ImageNet with a 20-fold reduction in data and 65 fold reduction in compute requirements. The proposed framework enhances the accessibility of model development while enabling flexible adaptation across diverse scenarios, offering an efficient approach to building multimodal models by utilizing existing unimodal architectures. Code and datasets will be released soon.

Vision foundation models (VFMs) trained on large-scale image datasets provide high-quality features that have significantly advanced 2D visual recognition. However, their potential in 3D vision remains largely untapped, despite the common availability of 2D images alongside 3D point cloud datasets. While significant research has been dedicated to 2D-3D fusion, recent state-of-the-art 3D methods predominantly focus on 3D data, leaving the integration of VFMs into 3D models underexplored. In this work, we challenge this trend by introducing DITR, a simple yet effective approach that extracts 2D foundation model features, projects them to 3D, and finally injects them into a 3D point cloud segmentation model. DITR achieves state-of-the-art results on both indoor and outdoor 3D semantic segmentation benchmarks. To enable the use of VFMs even when images are unavailable during inference, we further propose to distill 2D foundation models into a 3D backbone as a pretraining task. By initializing the 3D backbone with knowledge distilled from 2D VFMs, we create a strong basis for downstream 3D segmentation tasks, ultimately boosting performance across various datasets.

Visual perception connects high-level semantic understanding to pixel-level perception, but most existing settings assume that the decisive evidence for identifying a target is already in the image or frozen model knowledge. We study a more practical yet harder open-world case where a visible object must first be resolved from external facts, recent events, long-tail entities, or multi-hop relations before it can be localized. We formalize this challenge as Perception Deep Research and introduce WebEye, an object-anchored benchmark with verifiable evidence, knowledge-intensive queries, precise box/mask annotations, and three task views: Search-based Grounding, Search-based Segmentation, and Search-based VQA. WebEyes contains 120 images, 473 annotated object instances, 645 unique QA pairs, and 1,927 task samples. We further propose Pixel-Searcher, an agentic search-to-pixel workflow that resolves hidden target identities and binds them to boxes, masks, or grounded answers. Experiments show that Pixel-Searcher achieves the strongest open-source performance across all three task views, while failures mainly arise from evidence acquisition, identity resolution, and visual instance binding.

Active computer vision promises efficient, biologically plausible perception through sequential, localized glimpses, but lacks scalable general-purpose architectures and pretraining pipelines. As a result, Active-Vision Foundation Models (AVFMs) have remained unexplored. We introduce CanViT, the first task- and policy-agnostic AVFM. CanViT uses scene-relative RoPE to bind a retinotopic Vision Transformer backbone and a spatiotopic scene-wide latent workspace, the canvas. Efficient interaction with this high-capacity working memory is supported by Canvas Attention, a novel asymmetric cross-attention mechanism. We decouple thinking (backbone-level) and memory (canvas-level), eliminating canvas-side self-attention and fully-connected layers to achieve low-latency sequential inference and scalability to large scenes. We propose a label-free active vision pretraining scheme, policy-agnostic passive-to-active dense latent distillation: reconstructing scene-wide DINOv3 embeddings from sequences of low-resolution glimpses with randomized locations, zoom levels, and lengths. We pretrain CanViT-B from a random initialization on 13.2 million ImageNet-21k scenes -- an order of magnitude more than previous active models -- and 1 billion random glimpses, in 166 hours on a single H100. On ADE20K segmentation, a frozen CanViT-B achieves 38.5% mIoU in a single low-resolution glimpse, outperforming the best active model's 27.6% with 19.5x fewer inference FLOPs and no fine-tuning, as well as its FLOP- or input-matched DINOv3 teacher. Given additional glimpses, CanViT-B reaches 45.9% ADE20K mIoU. On ImageNet-1k classification, CanViT-B reaches 81.2% top-1 accuracy with frozen teacher probes. CanViT generalizes to longer rollouts, larger scenes, and new policies. Our work closes the wide gap between passive and active vision on semantic segmentation and demonstrates the potential of AVFMs as a new research axis.

Large foundation models have recently emerged as a prominent focus of interest, attaining superior performance in widespread scenarios. Due to the scarcity of 3D data, many efforts have been made to adapt pre-trained transformers from vision to 3D domains. However, such 2D-to-3D approaches are still limited, due to the potential loss of spatial geometries and high computation cost. More importantly, their frameworks are mainly designed for 2D models, lacking a general any-to-3D paradigm. In this paper, we introduce Any2Point, a parameter-efficient method to empower any-modality large models (vision, language, audio) for 3D understanding. Given a frozen transformer from any source modality, we propose a 3D-to-any (1D or 2D) virtual projection strategy that correlates the input 3D points to the original 1D or 2D positions within the source modality. This mechanism enables us to assign each 3D token with a positional encoding paired with the pre-trained model, which avoids 3D geometry loss caused by the true projection and better motivates the transformer for 3D learning with 1D/2D positional priors. Then, within each transformer block, we insert an any-to-3D guided adapter module for parameter-efficient fine-tuning. The adapter incorporates prior spatial knowledge from the source modality to guide the local feature aggregation of 3D tokens, compelling the semantic adaption of any-modality transformers. We conduct extensive experiments to showcase the effectiveness and efficiency of our method. Code and models are released at https://github.com/Ivan-Tang-3D/Any2Point.

In many computer vision applications, images are acquired with arbitrary or random rotations and translations, and in such setups, it is desirable to obtain semantic representations disentangled from the image orientation. Examples of such applications include semiconductor wafer defect inspection, plankton microscope images, and inference on single-particle cryo-electron microscopy (cryo-EM) micro-graphs. In this work, we propose Invariant Representation Learning with Implicit Neural Representation (IRL-INR), which uses an implicit neural representation (INR) with a hypernetwork to obtain semantic representations disentangled from the orientation of the image. We show that IRL-INR can effectively learn disentangled semantic representations on more complex images compared to those considered in prior works and show that these semantic representations synergize well with SCAN to produce state-of-the-art unsupervised clustering results.

Large unimodal foundation models for vision and language encode rich semantic structures, yet aligning them typically requires computationally intensive multimodal fine-tuning. Such approaches depend on large-scale parameter updates, are resource intensive, and can perturb pretrained representations. Emerging evidence suggests, however, that independently trained foundation models may already exhibit latent semantic compatibility, reflecting shared structures in the data they model. This raises a fundamental question: can cross-modal alignment be achieved without modifying the models themselves? Here we introduce HDFLIM (HyperDimensional computing with Frozen Language and Image Models), a framework that establishes cross-modal mappings while keeping pretrained vision and language models fully frozen. HDFLIM projects unimodal embeddings into a shared hyperdimensional space and leverages lightweight symbolic operations -- binding, bundling, and similarity-based retrieval to construct associative cross-modal representations in a single pass over the data. Caption generation emerges from high-dimensional memory retrieval rather than iterative gradient-based optimization. We show that HDFLIM achieves performance comparable to end-to-end vision-language training methods and produces captions that are more semantically grounded than zero-shot baselines. By decoupling alignment from parameter tuning, our results suggest that semantic mapping across foundation models can be realized through symbolic operations on hyperdimensional encodings of the respective embeddings. More broadly, this work points toward an alternative paradigm for foundation model alignment in which frozen models are integrated through structured representational mappings rather than through large-scale retraining. The codebase for our implementation can be found at https://github.com/Abhishek-Dalvi410/HDFLIM.

Foundation models pre-trained on large-scale natural image datasets offer a powerful paradigm for medical image segmentation. However, effectively transferring their learned representations for precise clinical applications remains a challenge. In this work, we propose Dino U-Net, a novel encoder-decoder architecture designed to exploit the high-fidelity dense features of the DINOv3 vision foundation model. Our architecture introduces an encoder built upon a frozen DINOv3 backbone, which employs a specialized adapter to fuse the model's rich semantic features with low-level spatial details. To preserve the quality of these representations during dimensionality reduction, we design a new fidelity-aware projection module (FAPM) that effectively refines and projects the features for the decoder. We conducted extensive experiments on seven diverse public medical image segmentation datasets. Our results show that Dino U-Net achieves state-of-the-art performance, consistently outperforming previous methods across various imaging modalities. Our framework proves to be highly scalable, with segmentation accuracy consistently improving as the backbone model size increases up to the 7-billion-parameter variant. The findings demonstrate that leveraging the superior, dense-pretrained features from a general-purpose foundation model provides a highly effective and parameter-efficient approach to advance the accuracy of medical image segmentation. The code is available at https://github.com/yifangao112/DinoUNet.

We propose a novel approach for disentangling visual and semantic features from the backbones of pre-trained diffusion models, enabling visual correspondence in a manner analogous to the well-established semantic correspondence. While diffusion model backbones are known to encode semantically rich features, they must also contain visual features to support their image synthesis capabilities. However, isolating these visual features is challenging due to the absence of annotated datasets. To address this, we introduce an automated pipeline that constructs image pairs with annotated semantic and visual correspondences based on existing subject-driven image generation datasets, and design a contrastive architecture to separate the two feature types. Leveraging the disentangled representations, we propose a new metric, Visual Semantic Matching (VSM), that quantifies visual inconsistencies in subject-driven image generation. Empirical results show that our approach outperforms global feature-based metrics such as CLIP, DINO, and vision--language models in quantifying visual inconsistencies while also enabling spatial localization of inconsistent regions. To our knowledge, this is the first method that supports both quantification and localization of inconsistencies in subject-driven generation, offering a valuable tool for advancing this task. Project Page:https://abdo-eldesokey.github.io/mind-the-glitch/

Vision-language models (VLMs) are typically composed of a vision encoder, e.g. CLIP, and a language model (LM) that interprets the encoded features to solve downstream tasks. Despite remarkable progress, VLMs are subject to several shortcomings due to the limited capabilities of vision encoders, e.g. "blindness" to certain image features, visual hallucination, etc. To address these issues, we study broadening the visual encoding capabilities of VLMs. We first comprehensively benchmark several vision encoders with different inductive biases for solving VLM tasks. We observe that there is no single encoding configuration that consistently achieves top performance across different tasks, and encoders with different biases can perform surprisingly similarly. Motivated by this, we introduce a method, named BRAVE, that consolidates features from multiple frozen encoders into a more versatile representation that can be directly fed as the input to a frozen LM. BRAVE achieves state-of-the-art performance on a broad range of captioning and VQA benchmarks and significantly reduces the aforementioned issues of VLMs, while requiring a smaller number of trainable parameters than existing methods and having a more compressed representation. Our results highlight the potential of incorporating different visual biases for a more broad and contextualized visual understanding of VLMs.

We explore the plain, non-hierarchical Vision Transformer (ViT) as a backbone network for object detection. This design enables the original ViT architecture to be fine-tuned for object detection without needing to redesign a hierarchical backbone for pre-training. With minimal adaptations for fine-tuning, our plain-backbone detector can achieve competitive results. Surprisingly, we observe: (i) it is sufficient to build a simple feature pyramid from a single-scale feature map (without the common FPN design) and (ii) it is sufficient to use window attention (without shifting) aided with very few cross-window propagation blocks. With plain ViT backbones pre-trained as Masked Autoencoders (MAE), our detector, named ViTDet, can compete with the previous leading methods that were all based on hierarchical backbones, reaching up to 61.3 AP_box on the COCO dataset using only ImageNet-1K pre-training. We hope our study will draw attention to research on plain-backbone detectors. Code for ViTDet is available in Detectron2.

Existing multimodal large language models have achieved high-fidelity visual perception and exploratory visual generation. However, a precision paradox persists in complex reasoning tasks: optical perception systems transcribe symbols without capturing logical topology, while pixel-based generative models produce visual artifacts lacking mathematical exactness. To bridge this gap, we propose that reasoning over visual inputs be reconceptualized as optical decompression-the process of reconstructing latent logical structures from compressed visual tokens. Guided by the axiom that Parsing is Reasoning, we introduce Thinking with Drafting (TwD), which utilizes a minimalist Domain-Specific Language (DSL) as a grounding intermediate representation. Unlike standard approaches that hallucinate answers directly, TwD forces the model to draft its mental model into executable code, rendering deterministic visual proofs for self-verification. To validate this, we present VisAlg, a visual algebra benchmark. Experiments demonstrate that TwD serve as a superior cognitive scaffold. Our work establishes a closed-loop system where visual generation acts not as a creative output but as a logical verifier, offering a generalizable path for visual reasoning.

Visual reasoning (VR), which is crucial in many fields for enabling human-like visual understanding, remains highly challenging. Recently, compositional visual reasoning approaches, which leverage the reasoning abilities of large language models (LLMs) with integrated tools to solve problems, have shown promise as more effective strategies than end-to-end VR methods. However, these approaches face limitations, as frozen LLMs lack tool awareness in VR, leading to performance bottlenecks. While leveraging LLMs for reasoning is widely used in other domains, they are not directly applicable to VR due to limited training data, imperfect tools that introduce errors and reduce data collection efficiency in VR, and challenging in fine-tuning on noisy workflows. To address these challenges, we propose DWIM: i) Discrepancy-aware training Workflow generation, which assesses tool usage and extracts more viable workflows for training; and ii) Instruct-Masking fine-tuning, which guides the model to only clone effective actions, enabling the generation of more practical solutions. Our experiments demonstrate that DWIM achieves state-of-the-art performance across various VR tasks, exhibiting strong generalization on multiple widely-used datasets.

In-context segmentation (ICS) aims to segment arbitrary concepts, e.g., objects, parts, or personalized instances, given one annotated visual examples. Existing work relies on (i) fine-tuning vision foundation models (VFMs), which improves in-domain results but harms generalization, or (ii) combines multiple frozen VFMs, which preserves generalization but yields architectural complexity and fixed segmentation granularities. We revisit ICS from a minimalist perspective and ask: Can a single self-supervised backbone support both semantic matching and segmentation, without any supervision or auxiliary models? We show that scaled-up dense self-supervised features from DINOv3 exhibit strong spatial structure and semantic correspondence. We introduce INSID3, a training-free approach that segments concepts at varying granularities only from frozen DINOv3 features, given an in-context example. INSID3 achieves state-of-the-art results across one-shot semantic, part, and personalized segmentation, outperforming previous work by +7.5 % mIoU, while using 3x fewer parameters and without any mask or category-level supervision. Code is available at https://github.com/visinf/INSID3 .

Multimodal Large Language Models (MLLMs) have recently demonstrated impressive capabilities in connecting vision and language, yet their proficiency in fundamental visual reasoning tasks remains limited. This limitation can be attributed to the fact that MLLMs learn visual understanding primarily from textual descriptions, which constitute a subjective and inherently incomplete supervisory signal. Furthermore, the modest scale of multimodal instruction tuning compared to massive text-only pre-training leads MLLMs to overfit language priors while overlooking visual details. To address these issues, we introduce JARVIS, a JEPA-inspired framework for self-supervised visual enhancement in MLLMs. Specifically, we integrate the I-JEPA learning paradigm into the standard vision-language alignment pipeline of MLLMs training. Our approach leverages frozen vision foundation models as context and target encoders, while training the predictor, implemented as the early layers of an LLM, to learn structural and semantic regularities from images without relying exclusively on language supervision. Extensive experiments on standard MLLM benchmarks show that JARVIS consistently improves performance on vision-centric benchmarks across different LLM families, without degrading multimodal reasoning abilities. Our source code is publicly available at: https://github.com/aimagelab/JARVIS.

Contrastive Language-Image Pretraining (CLIP) stands out as a prominent method for image representation learning. Various neural architectures, spanning Transformer-based models like Vision Transformers (ViTs) to Convolutional Networks (ConvNets) like ResNets, are trained with CLIP and serve as universal backbones across diverse vision tasks. Despite utilizing the same data and training objectives, the effectiveness of representations learned by these architectures raises a critical question. Our investigation explores the differences in CLIP performance among these backbone architectures, revealing significant disparities in their classifications. Notably, normalizing these representations results in substantial performance variations. Our findings showcase a remarkable possible synergy between backbone predictions that could reach an improvement of over 20% through informed selection of the appropriate backbone. Moreover, we propose a simple, yet effective approach to combine predictions from multiple backbones, leading to a notable performance boost of up to 6.34\%. We will release the code for reproducing the results.

Video question answering (VideoQA) is a complex task that requires diverse multi-modal data for training. Manual annotation of question and answers for videos, however, is tedious and prohibits scalability. To tackle this problem, recent methods consider zero-shot settings with no manual annotation of visual question-answer. In particular, a promising approach adapts frozen autoregressive language models pretrained on Web-scale text-only data to multi-modal inputs. In contrast, we here build on frozen bidirectional language models (BiLM) and show that such an approach provides a stronger and cheaper alternative for zero-shot VideoQA. In particular, (i) we combine visual inputs with the frozen BiLM using light trainable modules, (ii) we train such modules using Web-scraped multi-modal data, and finally (iii) we perform zero-shot VideoQA inference through masked language modeling, where the masked text is the answer to a given question. Our proposed approach, FrozenBiLM, outperforms the state of the art in zero-shot VideoQA by a significant margin on a variety of datasets, including LSMDC-FiB, iVQA, MSRVTT-QA, MSVD-QA, ActivityNet-QA, TGIF-FrameQA, How2QA and TVQA. It also demonstrates competitive performance in the few-shot and fully-supervised setting. Our code and models are publicly available at https://github.com/antoyang/FrozenBiLM.

Visual reasoning is critical for a wide range of computer vision tasks that go beyond surface-level object detection and classification. Despite notable advances in relational, symbolic, temporal, causal, and commonsense reasoning, existing surveys often address these directions in isolation, lacking a unified analysis and comparison across reasoning types, methodologies, and evaluation protocols. This survey aims to address this gap by categorizing visual reasoning into five major types (relational, symbolic, temporal, causal, and commonsense) and systematically examining their implementation through architectures such as graph-based models, memory networks, attention mechanisms, and neuro-symbolic systems. We review evaluation protocols designed to assess functional correctness, structural consistency, and causal validity, and critically analyze their limitations in terms of generalizability, reproducibility, and explanatory power. Beyond evaluation, we identify key open challenges in visual reasoning, including scalability to complex scenes, deeper integration of symbolic and neural paradigms, the lack of comprehensive benchmark datasets, and reasoning under weak supervision. Finally, we outline a forward-looking research agenda for next-generation vision systems, emphasizing that bridging perception and reasoning is essential for building transparent, trustworthy, and cross-domain adaptive AI systems, particularly in critical domains such as autonomous driving and medical diagnostics.

Video Joint Embedding Predictive Architectures (V-JEPA) learn generalizable off-the-shelf video representation by predicting masked regions in latent space with an exponential moving average (EMA)-updated teacher. While EMA prevents representation collapse, it complicates scalable model selection and couples teacher and student architectures. We revisit masked-latent prediction and show that a frozen teacher suffices. Concretely, we (i) train a target encoder with a simple pixel-reconstruction objective under V-JEPA masking, then (ii) freeze it and train a student to predict the teacher's latents on masked regions. This leads to a two-stage, unregularized scheme that we refer to as SALT (Static-teacher Asymmetric Latent Training). SALT decouples optimization into pixel reconstruction (teacher) and masked latent prediction (student), increasing transparency, efficiency, and scalability while preserving the ability of representation to generalize under frozen evaluation. Empirically, our student models outperform recently proposed V-JEPA 2 encoders under frozen backbone evaluation across diverse benchmarks. They are also more compute-optimal: at matched pretraining FLOPs, our method achieves higher probing accuracy, and its scaling curves dominate V-JEPA's accuracy-FLOPs Pareto frontier. Finally, we find that student quality is remarkably robust to teacher quality: high-performing students emerge even with small, sub-optimal teachers. This points to a compute budget allocation that should overwhelmingly favor the student. These results position SALT as a simple, scalable, and compute-efficient alternative to EMA-based self-distillation for video representation learning.

Recent research on vision backbone architectures has predominantly focused on optimizing efficiency for hardware platforms with high parallel processing capabilities. This category increasingly includes embedded systems such as mobile phones and embedded AI accelerator modules. In contrast, CPUs do not have the possibility to parallelize operations in the same manner, wherefore models benefit from a specific design philosophy that balances amount of operations (MACs) and hardware-efficient execution by having high MACs per second (MACpS). In pursuit of this, we investigate two modifications to standard convolutions, aimed at reducing computational cost: grouping convolutions and reducing kernel sizes. While both adaptations substantially decrease the total number of MACs required for inference, sustaining low latency necessitates preserving hardware-efficiency. Our experiments across diverse CPU devices confirm that these adaptations successfully retain high hardware-efficiency on CPUs. Based on these insights, we introduce CPUBone, a new family of vision backbone models optimized for CPU-based inference. CPUBone achieves state-of-the-art Speed-Accuracy Trade-offs (SATs) across a wide range of CPU devices and effectively transfers its efficiency to downstream tasks such as object detection and semantic segmentation. Models and code are available at https://github.com/altair199797/CPUBone.

Vision-Language Models (VLMs) have achieved remarkable progress across tasks such as visual question answering and image captioning. Yet, the extent to which these models perform visual reasoning as opposed to relying on linguistic priors remains unclear. To address this, we introduce VisRes Bench, a benchmark designed to study visual reasoning in naturalistic settings without contextual language supervision. Analyzing model behavior across three levels of complexity, we uncover clear limitations in perceptual and relational visual reasoning capacities. VisRes isolates distinct reasoning abilities across its levels. Level 1 probes perceptual completion and global image matching under perturbations such as blur, texture changes, occlusion, and rotation; Level 2 tests rule-based inference over a single attribute (e.g., color, count, orientation); and Level 3 targets compositional reasoning that requires integrating multiple visual attributes. Across more than 19,000 controlled task images, we find that state-of-the-art VLMs perform near random under subtle perceptual perturbations, revealing limited abstraction beyond pattern recognition. We conclude by discussing how VisRes provides a unified framework for advancing abstract visual reasoning in multimodal research.

Being able to perceive the semantics and the spatial structure of the environment is essential for visual navigation of a household robot. However, most existing works only employ visual backbones pre-trained either with independent images for classification or with self-supervised learning methods to adapt to the indoor navigation domain, neglecting the spatial relationships that are essential to the learning of navigation. Inspired by the behavior that humans naturally build semantically and spatially meaningful cognitive maps in their brains during navigation, in this paper, we propose a novel navigational-specific visual representation learning method by contrasting the agent's egocentric views and semantic maps (Ego^2-Map). We apply the visual transformer as the backbone encoder and train the model with data collected from the large-scale Habitat-Matterport3D environments. Ego^2-Map learning transfers the compact and rich information from a map, such as objects, structure and transition, to the agent's egocentric representations for navigation. Experiments show that agents using our learned representations on object-goal navigation outperform recent visual pre-training methods. Moreover, our representations significantly improve vision-and-language navigation in continuous environments for both high-level and low-level action spaces, achieving new state-of-the-art results of 47% SR and 41% SPL on the test server.

Recent advances in vision-language reasoning underscore the importance of thinking with images, where models actively ground their reasoning in visual evidence. Yet, prevailing frameworks treat visual actions as optional tools, boosting metrics but leaving reasoning ungrounded and crops ineffective. This gap gives rise to the illusion of thinking with images: models seem visually grounded but rely on context-agnostic actions that neither refine perception nor guide reasoning toward correct answers. We address this problem by reframing visual actions as core reasoning primitives rather than optional tools, which we term visual rationalization, the visual analogue of textual Chain-of-Thought. Building on this insight, we propose Visual Rationale Learning (ViRL), an end-to-end paradigm that grounds training in the visual rationale itself. ViRL integrates (1) Process Supervision with ground-truth rationales, (2) Objective Alignment via step-level reward shaping, and (3) Fine-Grained Credit Assignment to distinguish correct, redundant, and erroneous actions. By ensuring each action contributes meaningfully to the reasoning chain, ViRL enables models to "get the right answer for the right visual reason". Trained purely with end-to-end RL, ViRL achieves state-of-the-art results across benchmarks spanning perception, hallucination, and reasoning. This work establishes visual rationalization as a task-agnostic, process-grounded paradigm for building transparent, verifiable, and trustworthy vision-language models.

Multimodal large language models (MLLMs) have advanced the integration of visual and linguistic modalities, establishing themselves as the dominant paradigm for visual-language tasks. Current approaches like chain of thought (CoT) reasoning have augmented the cognitive capabilities of large language models (LLMs), yet their adaptation to MLLMs is hindered by heightened risks of hallucination in cross-modality comprehension. In this paper, we find that the thinking while looking paradigm in current multimodal CoT approaches--where reasoning chains are generated alongside visual input--fails to mitigate hallucinations caused by misleading images. To address these limitations, we propose the Visual Inference Chain (VIC) framework, a novel approach that constructs reasoning chains using textual context alone before introducing visual input, effectively reducing cross-modal biases and enhancing multimodal reasoning accuracy. Comprehensive evaluations demonstrate that VIC significantly improves zero-shot performance across various vision-related tasks, mitigating hallucinations while refining the reasoning capabilities of MLLMs. Our code repository can be found at https://github.com/Terry-Xu-666/visual_inference_chain.

Text-to-image diffusion models can be fine-tuned in custom domains to adapt to specific user preferences, but such unconstrained adaptability has also been utilized for illegal purposes, such as forging public figures' portraits and duplicating copyrighted artworks. Most existing work focuses on detecting the illegally generated contents, but cannot prevent or mitigate illegal adaptations of diffusion models. Other schemes of model unlearning and reinitialization, similarly, cannot prevent users from relearning the knowledge of illegal model adaptation with custom data. In this paper, we present FreezeAsGuard, a new technique that addresses these limitations and enables irreversible mitigation of illegal adaptations of diffusion models. The basic approach is that the model publisher selectively freezes tensors in pre-trained diffusion models that are critical to illegal model adaptations, to mitigate the fine-tuned model's representation power in illegal domains but minimize the impact on legal model adaptations in other domains. Such tensor freezing can be enforced via APIs provided by the model publisher for fine-tuning, can motivate users' adoption due to its computational savings. Experiment results with datasets in multiple domains show that FreezeAsGuard provides stronger power in mitigating illegal model adaptations of generating fake public figures' portraits, while having the minimum impact on model adaptation in other legal domains. The source code is available at: https://github.com/pittisl/FreezeAsGuard/

In this work, we introduce Semantic Pyramid AutoEncoder (SPAE) for enabling frozen LLMs to perform both understanding and generation tasks involving non-linguistic modalities such as images or videos. SPAE converts between raw pixels and interpretable lexical tokens (or words) extracted from the LLM's vocabulary. The resulting tokens capture both the semantic meaning and the fine-grained details needed for visual reconstruction, effectively translating the visual content into a language comprehensible to the LLM, and empowering it to perform a wide array of multimodal tasks. Our approach is validated through in-context learning experiments with frozen PaLM 2 and GPT 3.5 on a diverse set of image understanding and generation tasks. Our method marks the first successful attempt to enable a frozen LLM to generate image content while surpassing state-of-the-art performance in image understanding tasks, under the same setting, by over 25%.

Self-supervised learning (SSL) has advanced visual representation learning, but its value in chest radiography, a high-volume imaging modality with fine-grained findings, remains unclear. Meta's DINOv3 extends earlier SSL models through Gram-anchored self-distillation. Whether these design choices improve transfer learning for chest radiography has not been systematically tested. We benchmarked DINOv3 against DINOv2 and ImageNet initialization across seven datasets (n>814,000). Two representative backbones were evaluated: ViT-B/16 and ConvNeXt-B. Images were analyzed at 224x224, 512x512, and 1024x1024 pixels. We additionally assessed frozen features from a 7B model. The primary outcome was mean AUROC across labels. At 224x224, DINOv3 and DINOv2 achieved comparable performance on adult datasets. Increasing resolution to 512x512 yielded consistent improvements for DINOv3 over both DINOv2 and ImageNet. In contrast, results in pediatric cohort showed no differences across initializations. Across all settings, ConvNeXt-B outperformed ViT-B/16. Models using frozen DINOv3-7B features underperformed relative to fully finetuned 86-89M-parameter backbones, highlighting the importance of domain adaptation. Scaling to 1024x1024 did not further improve accuracy. Resolution-related gains were most evident for boundary-dependent and small focal abnormalities. In chest radiography, higher input resolution is critical for leveraging the benefits of modern self-supervised models. 512x512 pixels represent a practical upper limit where DINOv3-initialized ConvNeXt-B networks provide the strongest performance, while larger inputs offer minimal return on cost. Clinically, these findings support use of finetuned, mid-sized backbones at 512x512 for chest radiograph interpretation, with the greatest gains expected in detecting subtle or boundary-centered lesions relevant to emergency and critical care settings.

Visual token pruning is a promising approach for reducing the computational cost of vision-language models (VLMs), and existing methods often rely on early pruning decisions to improve efficiency. While effective on coarse-grained reasoning tasks, they suffer from significant performance degradation on tasks requiring fine-grained visual details. Through layer-wise analysis, we reveal substantial discrepancies in visual token importance across layers, showing that tokens deemed unimportant at shallow layers can later become highly relevant for text-conditioned reasoning. To avoid irreversible critical information loss caused by premature pruning, we introduce a new pruning paradigm, termed bypass, which preserves unselected visual tokens and forwards them to subsequent pruning stages for re-evaluation. Building on this paradigm, we propose SwiftVLM, a simple and training-free method that performs pruning at model-specific layers with strong visual token selection capability, while enabling independent pruning decisions across layers. Experiments across multiple VLMs and benchmarks demonstrate that SwiftVLM consistently outperforms existing pruning strategies, achieving superior accuracy-efficiency trade-offs and more faithful visual token selection behavior.

Monocular 3D object detection is an essential task in computer vision, and it has several applications in robotics and virtual reality. However, 3D object detectors are typically trained in a fully supervised way, relying extensively on 3D labeled data, which is labor-intensive and costly to annotate. This work focuses on weakly-supervised 3D detection to reduce data needs using a monocular method that leverages a singlecamera system over expensive LiDAR sensors or multi-camera setups. We propose a general model Weak Cube R-CNN, which can predict objects in 3D at inference time, requiring only 2D box annotations for training by exploiting the relationship between 2D projections of 3D cubes. Our proposed method utilizes pre-trained frozen foundation 2D models to estimate depth and orientation information on a training set. We use these estimated values as pseudo-ground truths during training. We design loss functions that avoid 3D labels by incorporating information from the external models into the loss. In this way, we aim to implicitly transfer knowledge from these large foundation 2D models without having access to 3D bounding box annotations. Experimental results on the SUN RGB-D dataset show increased performance in accuracy compared to an annotation time equalized Cube R-CNN baseline. While not precise for centimetre-level measurements, this method provides a strong foundation for further research.

The cold-start initialization stage plays a pivotal role in training Multimodal Large Reasoning Models (MLRMs), yet its mechanisms remain insufficiently understood. To analyze this stage, we introduce the Visual Attention Score (VAS), an attention-based metric that quantifies how much a model attends to visual tokens. We find that reasoning performance is strongly correlated with VAS (r=0.9616): models with higher VAS achieve substantially stronger multimodal reasoning. Surprisingly, multimodal cold-start fails to elevate VAS, resulting in attention distributions close to the base model, whereas text-only cold-start leads to a clear increase. We term this counter-intuitive phenomenon Lazy Attention Localization. To validate its causal role, we design training-free interventions that directly modulate attention allocation during inference, performance gains of 1-2% without any retraining. Building on these insights, we further propose Attention-Guided Visual Anchoring and Reflection (AVAR), a comprehensive cold-start framework that integrates visual-anchored data synthesis, attention-guided objectives, and visual-anchored reward shaping. Applied to Qwen2.5-VL-7B, AVAR achieves an average gain of 7.0% across 7 multimodal reasoning benchmarks. Ablation studies further confirm that each component of AVAR contributes step-wise to the overall gains. The code, data, and models are available at https://github.com/lrlbbzl/Qwen-AVAR.

Can multi-agent communication pressure extract discrete, compositional representations of invisible physical properties from frozen video features? We show that agents communicating through a Gumbel-Softmax bottleneck with iterated learning develop positionally disentangled protocols for latent properties (elasticity, friction, mass ratio) without property labels or supervision on message structure. With 4 agents, 100% of 80 seeds converge to near-perfect compositionality (PosDis=0.999, holdout 98.3%). Controls confirm multi-agent structure -- not bandwidth or temporal coverage -- drives this effect. Causal intervention shows surgical property disruption (~15% drop on targeted property, <3% on others). A controlled backbone comparison reveals that the perceptual prior determines what is communicable: DINOv2 dominates on spatially-visible ramp physics (98.3% vs 95.1%), while V-JEPA 2 dominates on dynamics-only collision physics (87.4% vs 77.7%, d=2.74). Scale-matched (d=3.37) and frame-matched (d=6.53) controls attribute this gap entirely to video-native pretraining. The frozen protocol supports action-conditioned planning (91.5%) with counterfactual velocity reasoning (r=0.780). Validation on Physics 101 real camera footage confirms 85.6% mass-comparison accuracy on unseen objects, temporal dynamics contributing +11.2% beyond static appearance, agent-scaling compositionality replicating at 90% for 4 agents, and causal intervention extending to real video (d=1.87, p=0.022).

With the overwhelming trend of mask image modeling led by MAE, generative pre-training has shown a remarkable potential to boost the performance of fundamental models in 2D vision. However, in 3D vision, the over-reliance on Transformer-based backbones and the unordered nature of point clouds have restricted the further development of generative pre-training. In this paper, we propose a novel 3D-to-2D generative pre-training method that is adaptable to any point cloud model. We propose to generate view images from different instructed poses via the cross-attention mechanism as the pre-training scheme. Generating view images has more precise supervision than its point cloud counterpart, thus assisting 3D backbones to have a finer comprehension of the geometrical structure and stereoscopic relations of the point cloud. Experimental results have proved the superiority of our proposed 3D-to-2D generative pre-training over previous pre-training methods. Our method is also effective in boosting the performance of architecture-oriented approaches, achieving state-of-the-art performance when fine-tuning on ScanObjectNN classification and ShapeNetPart segmentation tasks. Code is available at https://github.com/wangzy22/TAP.

This paper presents a comprehensive survey of the taxonomy and evolution of multimodal foundation models that demonstrate vision and vision-language capabilities, focusing on the transition from specialist models to general-purpose assistants. The research landscape encompasses five core topics, categorized into two classes. (i) We start with a survey of well-established research areas: multimodal foundation models pre-trained for specific purposes, including two topics -- methods of learning vision backbones for visual understanding and text-to-image generation. (ii) Then, we present recent advances in exploratory, open research areas: multimodal foundation models that aim to play the role of general-purpose assistants, including three topics -- unified vision models inspired by large language models (LLMs), end-to-end training of multimodal LLMs, and chaining multimodal tools with LLMs. The target audiences of the paper are researchers, graduate students, and professionals in computer vision and vision-language multimodal communities who are eager to learn the basics and recent advances in multimodal foundation models.

The primal sketch is a fundamental representation in Marr's vision theory, which allows for parsimonious image-level processing from 2D to 2.5D perception. This paper takes a further step by computing 3D primal sketch of wireframes from a set of images with known camera poses, in which we take the 2D wireframes in multi-view images as the basis to compute 3D wireframes in a volumetric rendering formulation. In our method, we first propose a NEural Attraction (NEAT) Fields that parameterizes the 3D line segments with coordinate Multi-Layer Perceptrons (MLPs), enabling us to learn the 3D line segments from 2D observation without incurring any explicit feature correspondences across views. We then present a novel Global Junction Perceiving (GJP) module to perceive meaningful 3D junctions from the NEAT Fields of 3D line segments by optimizing a randomly initialized high-dimensional latent array and a lightweight decoding MLP. Benefitting from our explicit modeling of 3D junctions, we finally compute the primal sketch of 3D wireframes by attracting the queried 3D line segments to the 3D junctions, significantly simplifying the computation paradigm of 3D wireframe parsing. In experiments, we evaluate our approach on the DTU and BlendedMVS datasets with promising performance obtained. As far as we know, our method is the first approach to achieve high-fidelity 3D wireframe parsing without requiring explicit matching.

Recently many techniques have emerged to create high quality 3D assets and scenes. When it comes to editing of these objects, however, existing approaches are either slow, compromise on quality, or do not provide enough customization. We introduce a novel approach to quickly edit a 3D model from a single reference view. Our technique first segments the edit image, and then matches semantically corresponding regions across chosen segmented dataset views using DINO features. A color or texture change from a particular region of the edit image can then be applied to other views automatically in a semantically sensible manner. These edited views act as an updated dataset to further train and re-style the 3D scene. The end-result is therefore an edited 3D model. Our framework enables a wide variety of editing tasks such as manual local edits, correspondence based style transfer from any example image, and a combination of different styles from multiple example images. We use Gaussian Splats as our primary 3D representation due to their speed and ease of local editing, but our technique works for other methods such as NeRFs as well. We show through multiple examples that our method produces higher quality results while offering fine-grained control of editing. Project page: ice-gaussian.github.io

Category-Agnostic Pose Estimation (CAPE) localizes keypoints across diverse object categories with a single model, using one or a few annotated support images. Recent works have shown that using a pose graph (i.e., treating keypoints as nodes in a graph rather than isolated points) helps handle occlusions and break symmetry. However, these methods assume a static pose graph with equal-weight edges, leading to suboptimal results. We introduce EdgeCape, a novel framework that overcomes these limitations by predicting the graph's edge weights which optimizes localization. To further leverage structural priors, we propose integrating Markovian Structural Bias, which modulates the self-attention interaction between nodes based on the number of hops between them. We show that this improves the model's ability to capture global spatial dependencies. Evaluated on the MP-100 benchmark, which includes 100 categories and over 20K images, EdgeCape achieves state-of-the-art results in the 1-shot setting and leads among similar-sized methods in the 5-shot setting, significantly improving keypoint localization accuracy. Our code is publicly available.

We present a novel framework for high-fidelity novel view synthesis (NVS) from sparse images, addressing key limitations in recent feed-forward 3D Gaussian Splatting (3DGS) methods built on Vision Transformer (ViT) backbones. While ViT-based pipelines offer strong geometric priors, they are often constrained by low-resolution inputs due to computational costs. Moreover, existing generative enhancement methods tend to be 3D-agnostic, resulting in inconsistent structures across views, especially in unseen regions. To overcome these challenges, we design a Dual-Domain Detail Perception Module, which enables handling high-resolution images without being limited by the ViT backbone, and endows Gaussians with additional features to store high-frequency details. We develop a feature-guided diffusion network, which can preserve high-frequency details during the restoration process. We introduce a unified training strategy that enables joint optimization of the ViT-based geometric backbone and the diffusion-based refinement module. Experiments demonstrate that our method can maintain superior generation quality across multiple datasets.

We propose an efficient method to ground pretrained text-only language models to the visual domain, enabling them to process arbitrarily interleaved image-and-text data, and generate text interleaved with retrieved images. Our method leverages the abilities of language models learnt from large scale text-only pretraining, such as in-context learning and free-form text generation. We keep the language model frozen, and finetune input and output linear layers to enable cross-modality interactions. This allows our model to process arbitrarily interleaved image-and-text inputs, and generate free-form text interleaved with retrieved images. We achieve strong zero-shot performance on grounded tasks such as contextual image retrieval and multimodal dialogue, and showcase compelling interactive abilities. Our approach works with any off-the-shelf language model and paves the way towards an effective, general solution for leveraging pretrained language models in visually grounded settings.

Recent advancements in Multimodal Large Language Models (MLLMs) have revolutionized the field of vision-language understanding by integrating visual perception capabilities into Large Language Models (LLMs). The prevailing trend in this field involves the utilization of a vision encoder derived from vision-language contrastive learning (CL), showing expertise in capturing overall representations while facing difficulties in capturing detailed local patterns. In this work, we focus on enhancing the visual representations for MLLMs by combining high-frequency and detailed visual representations, obtained through masked image modeling (MIM), with semantically-enriched low-frequency representations captured by CL. To achieve this goal, we introduce X-Former which is a lightweight transformer module designed to exploit the complementary strengths of CL and MIM through an innovative interaction mechanism. Specifically, X-Former first bootstraps vision-language representation learning and multimodal-to-multimodal generative learning from two frozen vision encoders, i.e., CLIP-ViT (CL-based) and MAE-ViT (MIM-based). It further bootstraps vision-to-language generative learning from a frozen LLM to ensure visual features from X-Former can be interpreted by the LLM. To demonstrate the effectiveness of our approach, we assess its performance on tasks demanding detailed visual understanding. Extensive evaluations indicate that X-Former excels in visual reasoning tasks involving both structural and semantic categories in the GQA dataset. Assessment on fine-grained visual perception benchmark further confirms its superior capabilities in visual understanding.

Research in efficient vision backbones is evolving into models that are a mixture of convolutions and transformer blocks. A smart combination of both, architecture-wise and component-wise is mandatory to excel in the speedaccuracy trade-off. Most publications focus on maximizing accuracy and utilize MACs (multiply accumulate operations) as an efficiency metric. The latter however often do not measure accurately how fast a model actually is due to factors like memory access cost and degree of parallelism. We analyzed common modules and architectural design choices for backbones not in terms of MACs, but rather in actual throughput and latency, as the combination of the latter two is a better representation of the efficiency of models in real applications. We applied the conclusions taken from that analysis to create a recipe for increasing hardware-efficiency in macro design. Additionally we introduce a simple slimmed-down version of MultiHead Self-Attention, that aligns with our analysis. We combine both macro and micro design to create a new family of hardware-efficient backbone networks called LowFormer. LowFormer achieves a remarkable speedup in terms of throughput and latency, while achieving similar or better accuracy than current state-of-the-art efficient backbones. In order to prove the generalizability of our hardware-efficient design, we evaluate our method on GPU, mobile GPU and ARM CPU. We further show that the downstream tasks object detection and semantic segmentation profit from our hardware-efficient architecture. Code and models are available at https://github.com/ altair199797/LowFormer.