Howard Zhou

Thomas Mensink, Jasper Uijlings, Lluis Castrejon et al. · deepmind

We propose Encyclopedic-VQA, a large scale visual question answering (VQA) dataset featuring visual questions about detailed properties of fine-grained categories and instances. It contains 221k unique question+answer pairs each matched with (up to) 5 images, resulting in a total of 1M VQA samples. Moreover, our dataset comes with a controlled knowledge base derived from Wikipedia, marking the evidence to support each answer. Empirically, we show that our dataset poses a hard challenge for large vision+language models as they perform poorly on our dataset: PaLI [14] is state-of-the-art on OK-VQA [37], yet it only achieves 13.0% accuracy on our dataset. Moreover, we experimentally show that progress on answering our encyclopedic questions can be achieved by augmenting large models with a mechanism that retrieves relevant information from the knowledge base. An oracle experiment with perfect retrieval achieves 87.0% accuracy on the single-hop portion of our dataset, and an automatic retrieval-augmented prototype yields 48.8%. We believe that our dataset enables future research on retrieval-augmented vision+language models. It is available at https://github.com/google-research/google-research/tree/master/encyclopedic_vqa .

Varun Jampani, Kevis-Kokitsi Maninis, Andreas Engelhardt et al. · deepmind

Recent advances in neural reconstruction enable high-quality 3D object reconstruction from casually captured image collections. Current techniques mostly analyze their progress on relatively simple image collections where Structure-from-Motion (SfM) techniques can provide ground-truth (GT) camera poses. We note that SfM techniques tend to fail on in-the-wild image collections such as image search results with varying backgrounds and illuminations. To enable systematic research progress on 3D reconstruction from casual image captures, we propose NAVI: a new dataset of category-agnostic image collections of objects with high-quality 3D scans along with per-image 2D-3D alignments providing near-perfect GT camera parameters. These 2D-3D alignments allow us to extract accurate derivative annotations such as dense pixel correspondences, depth and segmentation maps. We demonstrate the use of NAVI image collections on different problem settings and show that NAVI enables more thorough evaluations that were not possible with existing datasets. We believe NAVI is beneficial for systematic research progress on 3D reconstruction and correspondence estimation. Project page: https://navidataset.github.io

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

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

Arjun Karpur, Guilherme Perrotta, Ricardo Martin-Brualla et al. · deepmind

Finding localized correspondences across different images of the same object is crucial to understand its geometry. In recent years, this problem has seen remarkable progress with the advent of deep learning-based local image features and learnable matchers. Still, learnable matchers often underperform when there exists only small regions of co-visibility between image pairs (i.e. wide camera baselines). To address this problem, we leverage recent progress in coarse single-view geometry estimation methods. We propose LFM-3D, a Learnable Feature Matching framework that uses models based on graph neural networks and enhances their capabilities by integrating noisy, estimated 3D signals to boost correspondence estimation. When integrating 3D signals into the matcher model, we show that a suitable positional encoding is critical to effectively make use of the low-dimensional 3D information. We experiment with two different 3D signals - normalized object coordinates and monocular depth estimates - and evaluate our method on large-scale (synthetic and real) datasets containing object-centric image pairs across wide baselines. We observe strong feature matching improvements compared to 2D-only methods, with up to +6% total recall and +28% precision at fixed recall. Additionally, we demonstrate that the resulting improved correspondences lead to much higher relative posing accuracy for in-the-wild image pairs - up to 8.6% compared to the 2D-only approach.

Valentin Gabeur, Shangbang Long, Songyou Peng et al.

Recent works show that image and video generators exhibit zero-shot visual understanding behaviors, in a way reminiscent of how LLMs develop emergent capabilities of language understanding and reasoning from generative pretraining. While it has long been conjectured that the ability to create visual content implies an ability to understand it, there has been limited evidence that generative vision models have developed strong understanding capabilities. In this work, we demonstrate that image generation training serves a role similar to LLM pretraining, and lets models learn powerful and general visual representations that enable SOTA performance on various vision tasks. We introduce Vision Banana, a generalist model built by instruction-tuning Nano Banana Pro (NBP) on a mixture of its original training data alongside a small amount of vision task data. By parameterizing the output space of vision tasks as RGB images, we seamlessly reframe perception as image generation. Our generalist model, Vision Banana, achieves SOTA results on a variety of vision tasks involving both 2D and 3D understanding, beating or rivaling zero-shot domain-specialists, including Segment Anything Model 3 on segmentation tasks, and the Depth Anything series on metric depth estimation. We show that these results can be achieved with lightweight instruction-tuning without sacrificing the base model's image generation capabilities. The superior results suggest that image generation pretraining is a generalist vision learner. It also shows that image generation serves as a unified and universal interface for vision tasks, similar to text generation's role in language understanding and reasoning. We could be witnessing a major paradigm shift for computer vision, where generative vision pretraining takes a central role in building Foundational Vision Models for both generation and understanding.

Bingyi Cao, Koert Chen, Kevis-Kokitsi Maninis et al.

Recent progress in vision-language pretraining has enabled significant improvements to many downstream computer vision applications, such as classification, retrieval, segmentation and depth prediction. However, a fundamental capability that these models still struggle with is aligning dense patch representations with text embeddings of corresponding concepts. In this work, we investigate this critical issue and propose novel techniques to enhance this capability in foundational vision-language models. First, we reveal that a patch-level distillation procedure significantly boosts dense patch-text alignment -- surprisingly, the patch-text alignment of the distilled student model strongly surpasses that of the teacher model. This observation inspires us to consider modifications to pretraining recipes, leading us to propose iBOT++, an upgrade to the commonly-used iBOT masked image objective, where unmasked tokens also contribute directly to the loss. This dramatically enhances patch-text alignment of pretrained models. Additionally, to improve vision-language pretraining efficiency and effectiveness, we modify the exponential moving average setup in the learning recipe, and introduce a caption sampling strategy to benefit from synthetic captions at different granularities. Combining these components, we develop TIPSv2, a new family of image-text encoder models suitable for a wide range of downstream applications. Through comprehensive experiments on 9 tasks and 20 datasets, we demonstrate strong performance, generally on par with or better than recent vision encoder models. Code and models are released via our project page at https://gdm-tipsv2.github.io/ .

Xia Hu, Honglei Zhuang, Brian Potetz et al. · deepmind

The powerful reasoning of modern Vision Language Models open a new frontier for advanced personalization study. However, progress in this area is critically hampered by the lack of suitable benchmarks. To address this gap, we introduce Life-Bench, a comprehensive, synthetically generated multimodal benchmark built on simulated user digital footprints. Life-Bench features over questions evaluating a wide spectrum of capabilities, from persona understanding to complex reasoning over historical data. These capabilities expand far beyond prior benchmarks, reflecting the critical demands essential for real-world applications. Furthermore, we propose LifeGraph, an end-to-end framework that organizes personal context into a knowledge graph to facilitate structured retrieval and reasoning. Our experiments on Life-Bench reveal that existing methods falter significantly on complex personalized tasks, exposing a large performance headroom, especially in relational, temporal and aggregative reasoning. While LifeGraph closes this gap by leveraging structured knowledge and demonstrates a promising direction, these advanced personalization tasks remain a critical open challenge, motivating new research in this area.

Kevis-Kokitsi Maninis, Kaifeng Chen, Soham Ghosh et al. · cmu, deepmind

While image-text representation learning has become very popular in recent years, existing models tend to lack spatial awareness and have limited direct applicability for dense understanding tasks. For this reason, self-supervised image-only pretraining is still the go-to method for many dense vision applications (e.g. depth estimation, semantic segmentation), despite the lack of explicit supervisory signals. In this paper, we close this gap between image-text and self-supervised learning, by proposing a novel general-purpose image-text model, which can be effectively used off the shelf for dense and global vision tasks. Our method, which we refer to as Text-Image Pretraining with Spatial awareness (TIPS), leverages two simple and effective insights. First, on textual supervision: we reveal that replacing noisy web image captions by synthetically generated textual descriptions boosts dense understanding performance significantly, due to a much richer signal for learning spatially aware representations. We propose an adapted training method that combines noisy and synthetic captions, resulting in improvements across both dense and global understanding tasks. Second, on the learning technique: we propose to combine contrastive image-text learning with self-supervised masked image modeling, to encourage spatial coherence, unlocking substantial enhancements for downstream applications. Building on these two ideas, we scale our model using the transformer architecture, trained on a curated set of public images. Our experiments are conducted on 8 tasks involving 16 datasets in total, demonstrating strong off-the-shelf performance on both dense and global understanding, for several image-only and image-text tasks. Code and models are released at https://github.com/google-deepmind/tips.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Jiageng Mao, Sicheng He, Hao-Ning Wu et al.

We introduce PhysWorld, a framework that enables robot learning from video generation through physical world modeling. Recent video generation models can synthesize photorealistic visual demonstrations from language commands and images, offering a powerful yet underexplored source of training signals for robotics. However, directly retargeting pixel motions from generated videos to robots neglects physics, often resulting in inaccurate manipulations. PhysWorld addresses this limitation by coupling video generation with physical world reconstruction. Given a single image and a task command, our method generates task-conditioned videos and reconstructs the underlying physical world from the videos, and the generated video motions are grounded into physically accurate actions through object-centric residual reinforcement learning with the physical world model. This synergy transforms implicit visual guidance into physically executable robotic trajectories, eliminating the need for real robot data collection and enabling zero-shot generalizable robotic manipulation. Experiments on diverse real-world tasks demonstrate that PhysWorld substantially improves manipulation accuracy compared to previous approaches. Visit \href{https://pointscoder.github.io/PhysWorld_Web/}{the project webpage} for details.

Xia Hu, Zhenrui Yue, Brian Potetz et al.

As current Multimodal Large Language Models rapidly saturate canonical visual reasoning benchmarks, a key question emerges: do these strong scores genuinely reflect robust visual understanding? We identify a pervasive vulnerability, the \textbf{Cartesian Shortcut}: visual reasoning benchmarks prevalently build on orthogonal grid-based layouts that can be readily discretized into explicit textual coordinates. Models systematically exploit this property, heavily leveraging text-based deductive reasoning to assist visual problem-solving. To systematically dismantle this shortcut, we introduce \textbf{Polaris-Bench}, which re-formulates 53 visual reasoning tasks in Polar coordinate space with paired Cartesian counterparts as reference, while preserving consistent logical constraints and task semantics -- thus fundamentally breaking the orthogonal prior that models exploit. Comprehensive evaluation across $14$ state-of-the-art MLLMs reveals that frontier models achieving $70$--$83\%$ on Cartesian layouts collapse to $31$--$39\%$ on Polar equivalents, with degradation persisting even under complete logical equivalence. Moreover, reasoning gains observed on Cartesian layouts are severely diminished on Polar equivalents. These findings expose a critical deficiency in current MLLMs: the lack of topology-invariant visual reasoning.

Imad Eddine Toubal, Aditya Avinash, Neil Gordon Alldrin et al.

From content moderation to wildlife conservation, the number of applications that require models to recognize nuanced or subjective visual concepts is growing. Traditionally, developing classifiers for such concepts requires substantial manual effort measured in hours, days, or even months to identify and annotate data needed for training. Even with recently proposed Agile Modeling techniques, which enable rapid bootstrapping of image classifiers, users are still required to spend 30 minutes or more of monotonous, repetitive data labeling just to train a single classifier. Drawing on Fiske's Cognitive Miser theory, we propose a new framework that alleviates manual effort by replacing human labeling with natural language interactions, reducing the total effort required to define a concept by an order of magnitude: from labeling 2,000 images to only 100 plus some natural language interactions. Our framework leverages recent advances in foundation models, both large language models and vision-language models, to carve out the concept space through conversation and by automatically labeling training data points. Most importantly, our framework eliminates the need for crowd-sourced annotations. Moreover, our framework ultimately produces lightweight classification models that are deployable in cost-sensitive scenarios. Across 15 subjective concepts and across 2 public image classification datasets, our trained models outperform traditional Agile Modeling as well as state-of-the-art zero-shot classification models like ALIGN, CLIP, CuPL, and large visual question-answering models like PaLI-X.

Ian Huang, Yanan Bao, Karen Truong et al.

Scene generation with 3D assets presents a complex challenge, requiring both high-level semantic understanding and low-level geometric reasoning. While Multimodal Large Language Models (MLLMs) excel at semantic tasks, their application to 3D scene generation is hindered by their limited grounding on 3D geometry. In this paper, we investigate how to best work with MLLMs in an object placement task. Towards this goal, we introduce a novel framework, FirePlace, that applies existing MLLMs in (1) 3D geometric reasoning and the extraction of relevant geometric details from the 3D scene, (2) constructing and solving geometric constraints on the extracted low-level geometry, and (3) pruning for final placements that conform to common sense. By combining geometric reasoning with real-world understanding of MLLMs, our method can propose object placements that satisfy both geometric constraints as well as high-level semantic common-sense considerations. Our experiments show that these capabilities allow our method to place objects more effectively in complex scenes with intricate geometry, surpassing the quality of prior work.

Lluis Castrejon, Thomas Mensink, Howard Zhou et al.

Combining Large Language Models (LLMs) with external specialized tools (LLMs+tools) is a recent paradigm to solve multimodal tasks such as Visual Question Answering (VQA). While this approach was demonstrated to work well when optimized and evaluated for each individual benchmark, in practice it is crucial for the next generation of real-world AI systems to handle a broad range of multimodal problems. Therefore we pose the VQA problem from a unified perspective and evaluate a single system on a varied suite of VQA tasks including counting, spatial reasoning, OCR-based reasoning, visual pointing, external knowledge, and more. In this setting, we demonstrate that naively applying the LLM+tools approach using the combined set of all tools leads to poor results. This motivates us to introduce HAMMR: HierArchical MultiModal React. We start from a multimodal ReAct-based system and make it hierarchical by enabling our HAMMR agents to call upon other specialized agents. This enhances the compositionality of the LLM+tools approach, which we show to be critical for obtaining high accuracy on generic VQA. Concretely, on our generic VQA suite, HAMMR outperforms the naive LLM+tools approach by 19.5%. Additionally, HAMMR achieves state-of-the-art results on this task, outperforming the generic standalone PaLI-X VQA model by 5.0%.

Yijia Weng, Zhicheng Wang, Songyou Peng et al.

We perceive our surroundings with an active focus, paying more attention to regions of interest, such as the shelf labels in a grocery store. When it comes to scene reconstruction, this human perception trait calls for spatially varying degrees of detail ready for closer inspection in critical regions, preferably reconstructed on demand. While recent works in 3D Gaussian Splatting (3DGS) achieve fast, generalizable reconstruction from sparse views, their uniform resolution output leads to high computational costs unscalable to high-resolution training. As a result, they cannot leverage available images at their original high resolution to reconstruct details. Per-scene optimization methods reconstruct finer details with adaptive density control, yet require dense observations and lengthy offline optimization. To bridge the gap between the prohibitive cost of high-resolution holistic reconstructions and the user needs for localized fine details, we propose the problem of localized high-resolution reconstruction via on-demand Gaussian densification. Given a low-resolution 3DGS reconstruction, the goal is to learn a generalizable network that densifies the initial 3DGS to capture fine details in a user-specified local region of interest (RoI), based on sparse high-resolution observations of the RoI. This formulation avoids the high cost and redundancy of uniformly high-resolution reconstructions and fully leverages high-resolution captures in critical regions. We propose GaussianLens, a feed-forward densification framework that fuses multi-modal information from the initial 3DGS and multi-view images. We further design a pixel-guided densification mechanism that effectively captures details under large resolution increases. Experiments demonstrate our method's superior performance in local fine detail reconstruction and strong scalability to images of up to $1024\times1024$ resolution.

Qianqian Wang, Zhicheng Wang, Kyle Genova et al.

We present a method that synthesizes novel views of complex scenes by interpolating a sparse set of nearby views. The core of our method is a network architecture that includes a multilayer perceptron and a ray transformer that estimates radiance and volume density at continuous 5D locations (3D spatial locations and 2D viewing directions), drawing appearance information on the fly from multiple source views. By drawing on source views at render time, our method hearkens back to classic work on image-based rendering (IBR), and allows us to render high-resolution imagery. Unlike neural scene representation work that optimizes per-scene functions for rendering, we learn a generic view interpolation function that generalizes to novel scenes. We render images using classic volume rendering, which is fully differentiable and allows us to train using only multi-view posed images as supervision. Experiments show that our method outperforms recent novel view synthesis methods that also seek to generalize to novel scenes. Further, if fine-tuned on each scene, our method is competitive with state-of-the-art single-scene neural rendering methods. Project page: https://ibrnet.github.io/

Yang Feng, Futang Peng, Xu Zhang et al.

Deep image embedding provides a way to measure the semantic similarity of two images. It plays a central role in many applications such as image search, face verification, and zero-shot learning. It is desirable to have a universal deep embedding model applicable to various domains of images. However, existing methods mainly rely on training specialist embedding models each of which is applicable to images from a single domain. In this paper, we study an important but unexplored task: how to train a single universal image embedding model to match the performance of several specialists on each specialist's domain. Simply fusing the training data from multiple domains cannot solve this problem because some domains become overfitted sooner when trained together using existing methods. Therefore, we propose to distill the knowledge in multiple specialists into a universal embedding to solve this problem. In contrast to existing embedding distillation methods that distill the absolute distances between images, we transform the absolute distances between images into a probabilistic distribution and minimize the KL-divergence between the distributions of the specialists and the universal embedding. Using several public datasets, we validate that our proposed method accomplishes the goal of universal image embedding.

Calvin Murdock, Zhen Li, Howard Zhou et al.

Most deep architectures for image classification--even those that are trained to classify a large number of diverse categories--learn shared image representations with a single model. Intuitively, however, categories that are more similar should share more information than those that are very different. While hierarchical deep networks address this problem by learning separate features for subsets of related categories, current implementations require simplified models using fixed architectures specified via heuristic clustering methods. Instead, we propose Blockout, a method for regularization and model selection that simultaneously learns both the model architecture and parameters. A generalization of Dropout, our approach gives a novel parametrization of hierarchical architectures that allows for structure learning via back-propagation. To demonstrate its utility, we evaluate Blockout on the CIFAR and ImageNet datasets, demonstrating improved classification accuracy, better regularization performance, faster training, and the clear emergence of hierarchical network structures.

Jonathan Krause, Benjamin Sapp, Andrew Howard et al.

Current approaches for fine-grained recognition do the following: First, recruit experts to annotate a dataset of images, optionally also collecting more structured data in the form of part annotations and bounding boxes. Second, train a model utilizing this data. Toward the goal of solving fine-grained recognition, we introduce an alternative approach, leveraging free, noisy data from the web and simple, generic methods of recognition. This approach has benefits in both performance and scalability. We demonstrate its efficacy on four fine-grained datasets, greatly exceeding existing state of the art without the manual collection of even a single label, and furthermore show first results at scaling to more than 10,000 fine-grained categories. Quantitatively, we achieve top-1 accuracies of 92.3% on CUB-200-2011, 85.4% on Birdsnap, 93.4% on FGVC-Aircraft, and 80.8% on Stanford Dogs without using their annotated training sets. We compare our approach to an active learning approach for expanding fine-grained datasets.