Fangyin Wei

Niket Agarwal, Arslan Ali, Maciej Bala et al. · nvidia

Physical AI needs to be trained digitally first. It needs a digital twin of itself, the policy model, and a digital twin of the world, the world model. In this paper, we present the Cosmos World Foundation Model Platform to help developers build customized world models for their Physical AI setups. We position a world foundation model as a general-purpose world model that can be fine-tuned into customized world models for downstream applications. Our platform covers a video curation pipeline, pre-trained world foundation models, examples of post-training of pre-trained world foundation models, and video tokenizers. To help Physical AI builders solve the most critical problems of our society, we make Cosmos open-source and our models open-weight with permissive licenses available via https://github.com/nvidia-cosmos/cosmos-predict1.

Alisson Azzolini, Junjie Bai, Hannah Brandon et al. · nvidia

Physical AI systems need to perceive, understand, and perform complex actions in the physical world. In this paper, we present the Cosmos-Reason1 models that can understand the physical world and generate appropriate embodied decisions (e.g., next step action) in natural language through long chain-of-thought reasoning processes. We begin by defining key capabilities for Physical AI reasoning, with a focus on physical common sense and embodied reasoning. To represent physical common sense, we use a hierarchical ontology that captures fundamental knowledge about space, time, and physics. For embodied reasoning, we rely on a two-dimensional ontology that generalizes across different physical embodiments. Building on these capabilities, we develop two multimodal large language models, Cosmos-Reason1-7B and Cosmos-Reason1-56B. We curate data and train our models in two stages: Physical AI supervised fine-tuning (SFT) and Physical AI reinforcement learning (RL). To evaluate our models, we build comprehensive benchmarks for physical common sense and embodied reasoning according to our ontologies. Evaluation results show that Physical AI SFT and RL bring significant improvements. To facilitate the development of Physical AI, we make our code and pre-trained models available under the NVIDIA Open Model License at https://github.com/nvidia-cosmos/cosmos-reason1.

Arslan Ali, Junjie Bai, Maciej Bala et al. · nvidia

We introduce [Cosmos-Predict2.5], the latest generation of the Cosmos World Foundation Models for Physical AI. Built on a flow-based architecture, [Cosmos-Predict2.5] unifies Text2World, Image2World, and Video2World generation in a single model and leverages [Cosmos-Reason1], a Physical AI vision-language model, to provide richer text grounding and finer control of world simulation. Trained on 200M curated video clips and refined with reinforcement learning-based post-training, [Cosmos-Predict2.5] achieves substantial improvements over [Cosmos-Predict1] in video quality and instruction alignment, with models released at 2B and 14B scales. These capabilities enable more reliable synthetic data generation, policy evaluation, and closed-loop simulation for robotics and autonomous systems. We further extend the family with [Cosmos-Transfer2.5], a control-net style framework for Sim2Real and Real2Real world translation. Despite being 3.5$\times$ smaller than [Cosmos-Transfer1], it delivers higher fidelity and robust long-horizon video generation. Together, these advances establish [Cosmos-Predict2.5] and [Cosmos-Transfer2.5] as versatile tools for scaling embodied intelligence. To accelerate research and deployment in Physical AI, we release source code, pretrained checkpoints, and curated benchmarks under the NVIDIA Open Model License at https://github.com/nvidia-cosmos/cosmos-predict2.5 and https://github.com/nvidia-cosmos/cosmos-transfer2.5. We hope these open resources lower the barrier to adoption and foster innovation in building the next generation of embodied intelligence.

Aditi, Niket Agarwal, Arslan Ali et al.

We introduce Cosmos 3, a family of omnimodal world models designed to jointly process and generate language, image, video, audio, and action sequences within a unified mixture-of-transformers architecture. By supporting highly flexible input-output configurations, Cosmos 3 seamlessly unifies critical modalities for Physical AI -- effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework. Our evaluation demonstrates that Cosmos 3 establishes a new state-of-the-art across a diverse suite of understanding and generation tasks, demonstrating omnimodal world models as scalable, general-purpose backbones for embodied agents. Our post-trained Cosmos 3 models were ranked as the best open-source Text-to-Image and Image-to-Video models by Artificial Analysis, and the best policy model by RoboArena at the time the technical report was written. To accelerate open research and deployment in Physical AI, we make our code, model checkpoints, curated synthetic datasets, and evaluation benchmark available under the Linux Foundation's OpenMDW-1.1 https://openmdw.ai/license/1-1/ License at https://github.com/nvidia/cosmos}{github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3 . The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos3 .

Aarti Basant, Amlan Kar, Despoina Paschalidou et al. · nvidia

As autonomous vehicle capabilities advance, the safe evaluation of driving policies in long-tail scenarios remains a critical bottleneck. In closed-loop simulation, the driving policy model actively interacts with the environment, where its actions dynamically update the simulator state and directly influence the next set of generated sensor observations. While recent reconstruction-based neural simulators offer photorealism, they are fundamentally constrained by their initial captured data and struggle to generalize to highly dynamic or novel scenes. To overcome these limitations, we introduce OmniDreams, a foundation generative world model mid- and post-trained from the Cosmos diffusion model to autoregressively generate action-conditioned videos in real time. By leveraging the rich visual priors of Cosmos and mid- and post-training on 21k hours of driving scenarios, OmniDreams synthesizes complex, unobserved phenomena that are hard for traditional simulators to capture, such as extreme weather and unpredictable dynamic agent behaviors. Crucially, it autoregressively conditions its photorealistic sensor generation on past frames, the current simulator state, and immediate driving actions. Deployed in a closed-loop system with the Alpamayo 1 policy model and AlpaSim orchestrator, OmniDreams acts as a highly responsive, reactive environment, providing a scalable and comprehensive solution for training and evaluating next-generation autonomous driving policies. We additionally show preliminary results indicating that a world-action model (WAM) post-trained from OmniDreams achieves strong performance on the Physical AI Autonomous Vehicles NuRec dataset, surpassing the VLA-based Alpamayo 1.5 research policy model while using only 1/5 the total parameters. These results highlight the potential for a real-time world model like OmniDreams to also serve as a backbone for policy architectures.

Yuval Atzmon, Maciej Bala, Yogesh Balaji et al. · nvidia

We introduce Edify Image, a family of diffusion models capable of generating photorealistic image content with pixel-perfect accuracy. Edify Image utilizes cascaded pixel-space diffusion models trained using a novel Laplacian diffusion process, in which image signals at different frequency bands are attenuated at varying rates. Edify Image supports a wide range of applications, including text-to-image synthesis, 4K upsampling, ControlNets, 360 HDR panorama generation, and finetuning for image customization.

Maciej Bala, Yin Cui, Yifan Ding et al. · nvidia

We introduce Edify 3D, an advanced solution designed for high-quality 3D asset generation. Our method first synthesizes RGB and surface normal images of the described object at multiple viewpoints using a diffusion model. The multi-view observations are then used to reconstruct the shape, texture, and PBR materials of the object. Our method can generate high-quality 3D assets with detailed geometry, clean shape topologies, high-resolution textures, and materials within 2 minutes of runtime.

Fangyin Wei, Rohan Chabra, Lingni Ma et al.

Learning geometry, motion, and appearance priors of object classes is important for the solution of a large variety of computer vision problems. While the majority of approaches has focused on static objects, dynamic objects, especially with controllable articulation, are less explored. We propose a novel approach for learning a representation of the geometry, appearance, and motion of a class of articulated objects given only a set of color images as input. In a self-supervised manner, our novel representation learns shape, appearance, and articulation codes that enable independent control of these semantic dimensions. Our model is trained end-to-end without requiring any articulation annotations. Experiments show that our approach performs well for different joint types, such as revolute and prismatic joints, as well as different combinations of these joints. Compared to state of the art that uses direct 3D supervision and does not output appearance, we recover more faithful geometry and appearance from 2D observations only. In addition, our representation enables a large variety of applications, such as few-shot reconstruction, the generation of novel articulations, and novel view-synthesis.

Fangyin Wei, Thomas Funkhouser, Szymon Rusinkiewicz

Removing clutter from scenes is essential in many applications, ranging from privacy-concerned content filtering to data augmentation. In this work, we present an automatic system that removes clutter from 3D scenes and inpaints with coherent geometry and texture. We propose techniques for its two key components: 3D segmentation from shared properties and 3D inpainting, both of which are important problems. The definition of 3D scene clutter (frequently-moving objects) is not well captured by commonly-studied object categories in computer vision. To tackle the lack of well-defined clutter annotations, we group noisy fine-grained labels, leverage virtual rendering, and impose an instance-level area-sensitive loss. Once clutter is removed, we inpaint geometry and texture in the resulting holes by merging inpainted RGB-D images. This requires novel voting and pruning strategies that guarantee multi-view consistency across individually inpainted images for mesh reconstruction. Experiments on ScanNet and Matterport dataset show that our method outperforms baselines for clutter segmentation and 3D inpainting, both visually and quantitatively.

Hongchi Xia, Xuan Li, Zhaoshuo Li et al.

Real-world data collection for embodied agents remains costly and unsafe, calling for scalable, realistic, and simulator-ready 3D environments. However, existing scene-generation systems often rely on rule-based or task-specific pipelines, yielding artifacts and physically invalid scenes. We present SAGE, an agentic framework that, given a user-specified embodied task (e.g., "pick up a bowl and place it on the table"), understands the intent and automatically generates simulation-ready environments at scale. The agent couples multiple generators for layout and object composition with critics that evaluate semantic plausibility, visual realism, and physical stability. Through iterative reasoning and adaptive tool selection, it self-refines the scenes until meeting user intent and physical validity. The resulting environments are realistic, diverse, and directly deployable in modern simulators for policy training. Policies trained purely on this data exhibit clear scaling trends and generalize to unseen objects and layouts, demonstrating the promise of simulation-driven scaling for embodied AI. Code, demos, and the SAGE-10k dataset can be found on the project page here: https://nvlabs.github.io/sage.

Hongru Yan, Xiang Zhang, Zeyuan Chen et al.

Branches swaying in the breeze, flags rippling in the wind, and boats rocking on the water all show how aerodynamics shape natural motion -- an effect crucial for realism in vision and graphics. In this paper, we present Gaussian Swaying, a surface-based framework for aerodynamic simulation using 3D Gaussians. Unlike mesh-based methods that require costly meshing, or particle-based approaches that rely on discrete positional data, Gaussian Swaying models surfaces continuously with 3D Gaussians, enabling efficient and fine-grained aerodynamic interaction. Our framework unifies simulation and rendering on the same representation: Gaussian patches, which support force computation for dynamics while simultaneously providing normals for lightweight shading. Comprehensive experiments on both synthetic and real-world datasets across multiple metrics demonstrate that Gaussian Swaying achieves state-of-the-art performance and efficiency, offering a scalable approach for realistic aerodynamic scene simulation.

Yuanxing Duan, Fangyin Wei, Qiyu Dai et al.

We consider the problem of novel-view synthesis (NVS) for dynamic scenes. Recent neural approaches have accomplished exceptional NVS results for static 3D scenes, but extensions to 4D time-varying scenes remain non-trivial. Prior efforts often encode dynamics by learning a canonical space plus implicit or explicit deformation fields, which struggle in challenging scenarios like sudden movements or generating high-fidelity renderings. In this paper, we introduce 4D Gaussian Splatting (4DRotorGS), a novel method that represents dynamic scenes with anisotropic 4D XYZT Gaussians, inspired by the success of 3D Gaussian Splatting in static scenes. We model dynamics at each timestamp by temporally slicing the 4D Gaussians, which naturally compose dynamic 3D Gaussians and can be seamlessly projected into images. As an explicit spatial-temporal representation, 4DRotorGS demonstrates powerful capabilities for modeling complicated dynamics and fine details--especially for scenes with abrupt motions. We further implement our temporal slicing and splatting techniques in a highly optimized CUDA acceleration framework, achieving real-time inference rendering speeds of up to 277 FPS on an RTX 3090 GPU and 583 FPS on an RTX 4090 GPU. Rigorous evaluations on scenes with diverse motions showcase the superior efficiency and effectiveness of 4DRotorGS, which consistently outperforms existing methods both quantitatively and qualitatively.

Yu Liu, Hongyang Li, Junjie Yan et al.

Since convolutional neural network (CNN) lacks an inherent mechanism to handle large scale variations, we always need to compute feature maps multiple times for multi-scale object detection, which has the bottleneck of computational cost in practice. To address this, we devise a recurrent scale approximation (RSA) to compute feature map once only, and only through this map can we approximate the rest maps on other levels. At the core of RSA is the recursive rolling out mechanism: given an initial map at a particular scale, it generates the prediction at a smaller scale that is half the size of input. To further increase efficiency and accuracy, we (a): design a scale-forecast network to globally predict potential scales in the image since there is no need to compute maps on all levels of the pyramid. (b): propose a landmark retracing network (LRN) to trace back locations of the regressed landmarks and generate a confidence score for each landmark; LRN can effectively alleviate false positives caused by the accumulated error in RSA. The whole system can be trained end-to-end in a unified CNN framework. Experiments demonstrate that our proposed algorithm is superior against state-of-the-art methods on face detection benchmarks and achieves comparable results for generic proposal generation. The source code of RSA is available at github.com/sciencefans/RSA-for-object-detection.

Jiaxiang Tang, Ruijie Lu, Zhaoshuo Li et al.

Recent progress in 3D object generation has greatly improved both the quality and efficiency. However, most existing methods generate a single mesh with all parts fused together, which limits the ability to edit or manipulate individual parts. A key challenge is that different objects may have a varying number of parts. To address this, we propose a new end-to-end framework for part-level 3D object generation. Given a single input image, our method generates high-quality 3D objects with an arbitrary number of complete and semantically meaningful parts. We introduce a dual volume packing strategy that organizes all parts into two complementary volumes, allowing for the creation of complete and interleaved parts that assemble into the final object. Experiments show that our model achieves better quality, diversity, and generalization than previous image-based part-level generation methods.

Zeqi Gu, Yin Cui, Zhaoshuo Li et al.

Designing 3D scenes is traditionally a challenging task that demands both artistic expertise and proficiency with complex software. Recent advances in text-to-3D generation have greatly simplified this process by letting users create scenes based on simple text descriptions. However, as these methods generally require extra training or in-context learning, their performance is often hindered by the limited availability of high-quality 3D data. In contrast, modern text-to-image models learned from web-scale images can generate scenes with diverse, reliable spatial layouts and consistent, visually appealing styles. Our key insight is that instead of learning directly from 3D scenes, we can leverage generated 2D images as an intermediary to guide 3D synthesis. In light of this, we introduce ArtiScene, a training-free automated pipeline for scene design that integrates the flexibility of free-form text-to-image generation with the diversity and reliability of 2D intermediary layouts. First, we generate 2D images from a scene description, then extract the shape and appearance of objects to create 3D models. These models are assembled into the final scene using geometry, position, and pose information derived from the same intermediary image. Being generalizable to a wide range of scenes and styles, ArtiScene outperforms state-of-the-art benchmarks by a large margin in layout and aesthetic quality by quantitative metrics. It also averages a 74.89% winning rate in extensive user studies and 95.07% in GPT-4o evaluation. Project page: https://artiscene-cvpr.github.io/

Hanlin Chen, Fangyin Wei, Gim Hee Lee

Humans naturally interact with their 3D surroundings using language, and modeling 3D language fields for scene understanding and interaction has gained growing interest. This paper introduces ChatSplat, a system that constructs a 3D language field, enabling rich chat-based interaction within 3D space. Unlike existing methods that primarily use CLIP-derived language features focused solely on segmentation, ChatSplat facilitates interaction on three levels: objects, views, and the entire 3D scene. For view-level interaction, we designed an encoder that encodes the rendered feature map of each view into tokens, which are then processed by a large language model (LLM) for conversation. At the scene level, ChatSplat combines multi-view tokens, enabling interactions that consider the entire scene. For object-level interaction, ChatSplat uses a patch-wise language embedding, unlike LangSplat's pixel-wise language embedding that implicitly includes mask and embedding. Here, we explicitly decouple the language embedding into separate mask and feature map representations, allowing more flexible object-level interaction. To address the challenge of learning 3D Gaussians posed by the complex and diverse distribution of language embeddings used in the LLM, we introduce a learnable normalization technique to standardize these embeddings, facilitating effective learning. Extensive experimental results demonstrate that ChatSplat supports multi-level interactions -- object, view, and scene -- within 3D space, enhancing both understanding and engagement.

Hanlin Chen, Fangyin Wei, Chen Li et al.

Although 3D Gaussian Splatting has been widely studied because of its realistic and efficient novel-view synthesis, it is still challenging to extract a high-quality surface from the point-based representation. Previous works improve the surface by incorporating geometric priors from the off-the-shelf normal estimator. However, there are two main limitations: 1) Supervising normals rendered from 3D Gaussians effectively updates the rotation parameter but is less effective for other geometric parameters; 2) The inconsistency of predicted normal maps across multiple views may lead to severe reconstruction artifacts. In this paper, we propose a Depth-Normal regularizer that directly couples normal with other geometric parameters, leading to full updates of the geometric parameters from normal regularization. We further propose a confidence term to mitigate inconsistencies of normal predictions across multiple views. Moreover, we also introduce a densification and splitting strategy to regularize the size and distribution of 3D Gaussians for more accurate surface modeling. Compared with Gaussian-based baselines, experiments show that our approach obtains better reconstruction quality and maintains competitive appearance quality at faster training speed and 100+ FPS rendering.

Fangyin Wei, Elena Sizikova, Avneesh Sud et al.

Many applications in 3D shape design and augmentation require the ability to make specific edits to an object's semantic parameters (e.g., the pose of a person's arm or the length of an airplane's wing) while preserving as much existing details as possible. We propose to learn a deep network that infers the semantic parameters of an input shape and then allows the user to manipulate those parameters. The network is trained jointly on shapes from an auxiliary synthetic template and unlabeled realistic models, ensuring robustness to shape variability while relieving the need to label realistic exemplars. At testing time, edits within the parameter space drive deformations to be applied to the original shape, which provides semantically-meaningful manipulation while preserving the details. This is in contrast to prior methods that either use autoencoders with a limited latent-space dimensionality, failing to preserve arbitrary detail, or drive deformations with purely-geometric controls, such as cages, losing the ability to update local part regions. Experiments with datasets of chairs, airplanes, and human bodies demonstrate that our method produces more natural edits than prior work.

Nicolas Scheiner, Florian Kraus, Fangyin Wei et al.

Conventional sensor systems record information about directly visible objects, whereas occluded scene components are considered lost in the measurement process. Non-line-of-sight (NLOS) methods try to recover such hidden objects from their indirect reflections - faint signal components, traditionally treated as measurement noise. Existing NLOS approaches struggle to record these low-signal components outside the lab, and do not scale to large-scale outdoor scenes and high-speed motion, typical in automotive scenarios. In particular, optical NLOS capture is fundamentally limited by the quartic intensity falloff of diffuse indirect reflections. In this work, we depart from visible-wavelength approaches and demonstrate detection, classification, and tracking of hidden objects in large-scale dynamic environments using Doppler radars that can be manufactured at low-cost in series production. To untangle noisy indirect and direct reflections, we learn from temporal sequences of Doppler velocity and position measurements, which we fuse in a joint NLOS detection and tracking network over time. We validate the approach on in-the-wild automotive scenes, including sequences of parked cars or house facades as relay surfaces, and demonstrate low-cost, real-time NLOS in dynamic automotive environments.

Zhisheng Zhong, Fangyin Wei, Zhouchen Lin et al.

Despite the recent success of deep learning models in numerous applications, their widespread use on mobile devices is seriously impeded by storage and computational requirements. In this paper, we propose a novel network compression method called Adaptive Dimension Adjustment Tucker decomposition (ADA-Tucker). With learnable core tensors and transformation matrices, ADA-Tucker performs Tucker decomposition of arbitrary-order tensors. Furthermore, we propose that weight tensors in networks with proper order and balanced dimension are easier to be compressed. Therefore, the high flexibility in decomposition choice distinguishes ADA-Tucker from all previous low-rank models. To compress more, we further extend the model to Shared Core ADA-Tucker (SCADA-Tucker) by defining a shared core tensor for all layers. Our methods require no overhead of recording indices of non-zero elements. Without loss of accuracy, our methods reduce the storage of LeNet-5 and LeNet-300 by ratios of 691 times and 233 times, respectively, significantly outperforming state of the art. The effectiveness of our methods is also evaluated on other three benchmarks (CIFAR-10, SVHN, ILSVRC12) and modern newly deep networks (ResNet, Wide-ResNet).

Yu Liu, Fangyin Wei, Jing Shao et al.

This paper proposes learning disentangled but complementary face features with minimal supervision by face identification. Specifically, we construct an identity Distilling and Dispelling Autoencoder (D2AE) framework that adversarially learns the identity-distilled features for identity verification and the identity-dispelled features to fool the verification system. Thanks to the design of two-stream cues, the learned disentangled features represent not only the identity or attribute but the complete input image. Comprehensive evaluations further demonstrate that the proposed features not only maintain state-of-the-art identity verification performance on LFW, but also acquire competitive discriminative power for face attribute recognition on CelebA and LFWA. Moreover, the proposed system is ready to semantically control the face generation/editing based on various identities and attributes in an unsupervised manner.

Xiao Sun, Bin Xiao, Fangyin Wei et al.

State-of-the-art human pose estimation methods are based on heat map representation. In spite of the good performance, the representation has a few issues in nature, such as not differentiable and quantization error. This work shows that a simple integral operation relates and unifies the heat map representation and joint regression, thus avoiding the above issues. It is differentiable, efficient, and compatible with any heat map based methods. Its effectiveness is convincingly validated via comprehensive ablation experiments under various settings, specifically on 3D pose estimation, for the first time.