Andrew Tao

Amala Sanjay Deshmukh, Kateryna Chumachenko, Tuomas Rintamaki et al. · nvidia

We introduce Nemotron Nano V2 VL, the latest model of the Nemotron vision-language series designed for strong real-world document understanding, long video comprehension, and reasoning tasks. Nemotron Nano V2 VL delivers significant improvements over our previous model, Llama-3.1-Nemotron-Nano-VL-8B, across all vision and text domains through major enhancements in model architecture, datasets, and training recipes. Nemotron Nano V2 VL builds on Nemotron Nano V2, a hybrid Mamba-Transformer LLM, and innovative token reduction techniques to achieve higher inference throughput in long document and video scenarios. We are releasing model checkpoints in BF16, FP8, and FP4 formats and sharing large parts of our datasets, recipes and training code.

Aaron Blakeman, Aarti Basant, Abhinav Khattar et al. · nvidia

As inference-time scaling becomes critical for enhanced reasoning capabilities, it is increasingly becoming important to build models that are efficient to infer. We introduce Nemotron-H, a family of 8B and 56B/47B hybrid Mamba-Transformer models designed to reduce inference cost for a given accuracy level. To achieve this goal, we replace the majority of self-attention layers in the common Transformer model architecture with Mamba layers that perform constant computation and require constant memory per generated token. We show that Nemotron-H models offer either better or on-par accuracy compared to other similarly-sized state-of-the-art open-sourced Transformer models (e.g., Qwen-2.5-7B/72B and Llama-3.1-8B/70B), while being up to 3$\times$ faster at inference. To further increase inference speed and reduce the memory required at inference time, we created Nemotron-H-47B-Base from the 56B model using a new compression via pruning and distillation technique called MiniPuzzle. Nemotron-H-47B-Base achieves similar accuracy to the 56B model, but is 20% faster to infer. In addition, we introduce an FP8-based training recipe and show that it can achieve on par results with BF16-based training. This recipe is used to train the 56B model. We are releasing Nemotron-H base model checkpoints with support in Hugging Face and NeMo.

Aakshita Chandiramani, Aaron Blakeman, Abdullahi Olaoye et al. · amazon-science, cmu

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Min Shi, Fuxiao Liu, Shihao Wang et al.

The ability to accurately interpret complex visual information is a crucial topic of multimodal large language models (MLLMs). Recent work indicates that enhanced visual perception significantly reduces hallucinations and improves performance on resolution-sensitive tasks, such as optical character recognition and document analysis. A number of recent MLLMs achieve this goal using a mixture of vision encoders. Despite their success, there is a lack of systematic comparisons and detailed ablation studies addressing critical aspects, such as expert selection and the integration of multiple vision experts. This study provides an extensive exploration of the design space for MLLMs using a mixture of vision encoders and resolutions. Our findings reveal several underlying principles common to various existing strategies, leading to a streamlined yet effective design approach. We discover that simply concatenating visual tokens from a set of complementary vision encoders is as effective as more complex mixing architectures or strategies. We additionally introduce Pre-Alignment to bridge the gap between vision-focused encoders and language tokens, enhancing model coherence. The resulting family of MLLMs, Eagle, surpasses other leading open-source models on major MLLM benchmarks.

Ali Hatamizadeh, Greg Heinrich, Hongxu Yin et al.

We design a new family of hybrid CNN-ViT neural networks, named FasterViT, with a focus on high image throughput for computer vision (CV) applications. FasterViT combines the benefits of fast local representation learning in CNNs and global modeling properties in ViT. Our newly introduced Hierarchical Attention (HAT) approach decomposes global self-attention with quadratic complexity into a multi-level attention with reduced computational costs. We benefit from efficient window-based self-attention. Each window has access to dedicated carrier tokens that participate in local and global representation learning. At a high level, global self-attentions enable the efficient cross-window communication at lower costs. FasterViT achieves a SOTA Pareto-front in terms of accuracy and image throughput. We have extensively validated its effectiveness on various CV tasks including classification, object detection and segmentation. We also show that HAT can be used as a plug-and-play module for existing networks and enhance them. We further demonstrate significantly faster and more accurate performance than competitive counterparts for images with high resolution. Code is available at https://github.com/NVlabs/FasterViT.

Aarti Basant, Abhijit Khairnar, Abhijit Paithankar et al. · nvidia

We introduce Nemotron-Nano-9B-v2, a hybrid Mamba-Transformer language model designed to increase throughput for reasoning workloads while achieving state-of-the-art accuracy compared to similarly-sized models. Nemotron-Nano-9B-v2 builds on the Nemotron-H architecture, in which the majority of the self-attention layers in the common Transformer architecture are replaced with Mamba-2 layers, to achieve improved inference speed when generating the long thinking traces needed for reasoning. We create Nemotron-Nano-9B-v2 by first pre-training a 12-billion-parameter model (Nemotron-Nano-12B-v2-Base) on 20 trillion tokens using an FP8 training recipe. After aligning Nemotron-Nano-12B-v2-Base, we employ the Minitron strategy to compress and distill the model with the goal of enabling inference on up to 128k tokens on a single NVIDIA A10G GPU (22GiB of memory, bfloat16 precision). Compared to existing similarly-sized models (e.g., Qwen3-8B), we show that Nemotron-Nano-9B-v2 achieves on-par or better accuracy on reasoning benchmarks while achieving up to 6x higher inference throughput in reasoning settings like 8k input and 16k output tokens. We are releasing Nemotron-Nano-9B-v2, Nemotron-Nano12B-v2-Base, and Nemotron-Nano-9B-v2-Base checkpoints along with the majority of our pre- and post-training datasets on Hugging Face.

Arushi Goel, Sreyan Ghosh, Vatsal Agarwal et al.

Multimodal Large Language Models (MLLMs) have shown strong performance in visual and audio understanding when evaluated in isolation. However, their ability to jointly reason over omni-modal (visual, audio, and textual) signals in long and complex videos remains largely unexplored. We introduce MMOU, a new benchmark designed to systematically evaluate multimodal understanding and reasoning under these challenging, real-world conditions. MMOU consists of 15,000 carefully curated questions paired with 9038 web-collected videos of varying length, spanning diverse domains and exhibiting rich, tightly coupled audio-visual content. The benchmark covers 13 fundamental skill categories, all of which require integrating evidence across modalities and time. All questions are manually annotated across multiple turns by professional annotators, ensuring high quality and reasoning fidelity. We evaluate 20+ state-of-the-art open-source and proprietary multimodal models on MMOU. The results expose substantial performance gaps: the best closed-source model achieves only 64.2% accuracy, while the strongest open-source model reaches just 46.8%. Our results highlight the challenges of long-form omni-modal understanding, revealing that current models frequently fail to apply even fundamental skills in long videos. Through detailed analysis, we further identify systematic failure modes and provide insights into where and why current models break.

Amala Sanjay Deshmukh, Kateryna Chumachenko, Tuomas Rintamaki et al. · amazon-science, nvidia

We introduce Nemotron 3 Nano Omni, the latest model in the Nemotron multimodal series and the first to natively support audio inputs alongside text, images, and video. Nemotron 3 Nano Omni delivers consistent accuracy improvements over its predecessor, Nemotron Nano V2 VL, across all modalities, enabled by advances in architecture, training data and recipes. In particular, Nemotron 3 delivers leading results in real-world document understanding, long audio-video comprehension, and agentic computer use. Built on the highly efficient Nemotron 3 Nano 30B-A3B backbone, Nemotron 3 Nano Omni further incorporates innovative multimodal token-reduction techniques to deliver substantially lower inference latency and higher throughput than other models of similar size. We are releasing model checkpoints in BF16, FP8, and FP4 formats, along with portions of the training data and codebase to facilitate further research and development.

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We present Nemotron 3 Nano 30B-A3B, a Mixture-of-Experts hybrid Mamba-Transformer language model. Nemotron 3 Nano was pretrained on 25 trillion text tokens, including more than 3 trillion new unique tokens over Nemotron 2, followed by supervised fine tuning and large-scale RL on diverse environments. Nemotron 3 Nano achieves better accuracy than our previous generation Nemotron 2 Nano while activating less than half of the parameters per forward pass. It achieves up to 3.3x higher inference throughput than similarly-sized open models like GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507, while also being more accurate on popular benchmarks. Nemotron 3 Nano demonstrates enhanced agentic, reasoning, and chat abilities and supports context lengths up to 1M tokens. We release both our pretrained Nemotron 3 Nano 30B-A3B Base and post-trained Nemotron 3 Nano 30B-A3B checkpoints on Hugging Face.

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.

Aysegul Dundar, Jun Gao, Andrew Tao et al.

This paper presents a method to reconstruct high-quality textured 3D models from both multi-view and single-view images. The reconstruction is posed as an adaptation problem and is done progressively where in the first stage, we focus on learning accurate geometry, whereas in the second stage, we focus on learning the texture with a generative adversarial network. In the generative learning pipeline, we propose two improvements. First, since the learned textures should be spatially aligned, we propose an attention mechanism that relies on the learnable positions of pixels. Secondly, since discriminator receives aligned texture maps, we augment its input with a learnable embedding which improves the feedback to the generator. We achieve significant improvements on multi-view sequences from Tripod dataset as well as on single-view image datasets, Pascal 3D+ and CUB. We demonstrate that our method achieves superior 3D textured models compared to the previous works. Please visit our web-page for 3D visuals.

Shihao Wang, Shilong Liu, Yuanguo Kuang et al.

Vision-language models (VLMs) commonly formulate visual grounding and detection as a coordinate-token generation problem, serializing each 2D box into multiple 1D tokens that are learned and decoded largely independently. This token-by-token decoding mismatches the coupled structure of box geometry and creates a practical inference bottleneck due to strictly sequential generation. We introduce LocateAnything, a unified generative grounding and detection framework based on Parallel Box Decoding (PBD). By decoding geometric elements such as bounding boxes and points as atomic units in a single step, LocateAnything preserves intra-box geometric coherence and unlocks substantial parallelism. We show that PBD improves both decoding throughput and localization accuracy. We further develop a scalable data engine and curate LocateAnything-Data, a large-scale dataset with more than 138 million training samples, substantially increasing data diversity for high-precision localization. Extensive evaluations show that LocateAnything advances the speed-accuracy frontier, achieving significantly higher decoding throughput while improving high-IoU localization quality across diverse benchmarks. The results highlight the complementary benefits of Parallel Box Decoding and large-scale training data in enabling efficient and precise unified visual grounding and detection.

Boyi Li, Ligeng Zhu, Ran Tian et al.

We propose Wolf, a WOrLd summarization Framework for accurate video captioning. Wolf is an automated captioning framework that adopts a mixture-of-experts approach, leveraging complementary strengths of Vision Language Models (VLMs). By utilizing both image and video models, our framework captures different levels of information and summarizes them efficiently. Our approach can be applied to enhance video understanding, auto-labeling, and captioning. To evaluate caption quality, we introduce CapScore, an LLM-based metric to assess the similarity and quality of generated captions compared to the ground truth captions. We further build four human-annotated datasets in three domains: autonomous driving, general scenes, and robotics, to facilitate comprehensive comparisons. We show that Wolf achieves superior captioning performance compared to state-of-the-art approaches from the research community (VILA1.5, CogAgent) and commercial solutions (Gemini-Pro-1.5, GPT-4V). For instance, in comparison with GPT-4V, Wolf improves CapScore both quality-wise by 55.6% and similarity-wise by 77.4% on challenging driving videos. Finally, we establish a benchmark for video captioning and introduce a leaderboard, aiming to accelerate advancements in video understanding, captioning, and data alignment. Webpage: https://wolfv0.github.io/.

Zhiqi Li, Guo Chen, Shilong Liu et al.

Recently, promising progress has been made by open-source vision-language models (VLMs) in bringing their capabilities closer to those of proprietary frontier models. However, most open-source models only publish their final model weights, leaving the critical details of data strategies and implementation largely opaque. In this work, we address VLM post-training from a data-centric perspective, showing the key role of data strategy in developing frontier VLMs. By studying and building our post-training data strategy from scratch, we share detailed insights into the development processes, aiming to benefit the development of competitive models for the open-source community. Our introduced data strategy, together with training recipes and model design, leads to a family of performant VLMs named Eagle2. Specifically, Eagle2-9B achieves state-of-the-art results across various multimodal benchmarks, matching certain competitive models with up to 70B parameters.

Guo Chen, Zhiqi Li, Shihao Wang et al.

We introduce Eagle 2.5, a family of frontier vision-language models (VLMs) for long-context multimodal learning. Our work addresses the challenges in long video comprehension and high-resolution image understanding, introducing a generalist framework for both tasks. The proposed training framework incorporates Automatic Degrade Sampling and Image Area Preservation, two techniques that preserve contextual integrity and visual details. The framework also includes numerous efficiency optimizations in the pipeline for long-context data training. Finally, we propose Eagle-Video-110K, a novel dataset that integrates both story-level and clip-level annotations, facilitating long-video understanding. Eagle 2.5 demonstrates substantial improvements on long-context multimodal benchmarks, providing a robust solution to the limitations of existing VLMs. Notably, our best model Eagle 2.5-8B achieves 72.4% on Video-MME with 512 input frames, matching the results of top-tier commercial model such as GPT-4o and large-scale open-source models like Qwen2.5-VL-72B and InternVL2.5-78B.

Ji Lin, Hongxu Yin, Wei Ping et al.

Visual language models (VLMs) rapidly progressed with the recent success of large language models. There have been growing efforts on visual instruction tuning to extend the LLM with visual inputs, but lacks an in-depth study of the visual language pre-training process, where the model learns to perform joint modeling on both modalities. In this work, we examine the design options for VLM pre-training by augmenting LLM towards VLM through step-by-step controllable comparisons. We introduce three main findings: (1) freezing LLMs during pre-training can achieve decent zero-shot performance, but lack in-context learning capability, which requires unfreezing the LLM; (2) interleaved pre-training data is beneficial whereas image-text pairs alone are not optimal; (3) re-blending text-only instruction data to image-text data during instruction fine-tuning not only remedies the degradation of text-only tasks, but also boosts VLM task accuracy. With an enhanced pre-training recipe we build VILA, a Visual Language model family that consistently outperforms the state-of-the-art models, e.g., LLaVA-1.5, across main benchmarks without bells and whistles. Multi-modal pre-training also helps unveil appealing properties of VILA, including multi-image reasoning, enhanced in-context learning, and better world knowledge.

Hanrong Ye, Chao-Han Huck Yang, Arushi Goel et al.

Advancing machine intelligence requires developing the ability to perceive across multiple modalities, much as humans sense the world. We introduce OmniVinci, an initiative to build a strong, open-source, omni-modal LLM. We carefully study the design choices across model architecture and data curation. For model architecture, we present three key innovations: (i) OmniAlignNet for strengthening alignment between vision and audio embeddings in a shared omni-modal latent space; (ii) Temporal Embedding Grouping for capturing relative temporal alignment between vision and audio signals; and (iii) Constrained Rotary Time Embedding for encoding absolute temporal information in omni-modal embeddings. We introduce a curation and synthesis pipeline that generates 24M single-modal and omni-modal conversations. We find that modalities reinforce one another in both perception and reasoning. Our model, OmniVinci, outperforms Qwen2.5-Omni with +19.05 on DailyOmni (cross-modal understanding), +1.7 on MMAR (audio), and +3.9 on Video-MME (vision), while using just 0.2T training tokens - a 6 times reduction compared to Qwen2.5-Omni's 1.2T. We finally demonstrate omni-modal advantages in downstream applications spanning robotics, medical AI, and smart factory.

Mike Ranzinger, Greg Heinrich, Pavlo Molchanov et al.

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 $224 \times 224$px, while the "high-resolution" versions are around $378-448$px, 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 .

Ning Yu, Guilin Liu, Aysegul Dundar et al.

Generative Adversarial Networks (GANs) produce impressive results on unconditional image generation when powered with large-scale image datasets. Yet generated images are still easy to spot especially on datasets with high variance (e.g. bedroom, church). In this paper, we propose various improvements to further push the boundaries in image generation. Specifically, we propose a novel dual contrastive loss and show that, with this loss, discriminator learns more generalized and distinguishable representations to incentivize generation. In addition, we revisit attention and extensively experiment with different attention blocks in the generator. We find attention to be still an important module for successful image generation even though it was not used in the recent state-of-the-art models. Lastly, we study different attention architectures in the discriminator, and propose a reference attention mechanism. By combining the strengths of these remedies, we improve the compelling state-of-the-art Fréchet Inception Distance (FID) by at least 17.5% on several benchmark datasets. We obtain even more significant improvements on compositional synthetic scenes (up to 47.5% in FID). Code and models are available at https://github.com/ningyu1991/AttentionDualContrastGAN .

Greg Heinrich, Mike Ranzinger, Hongxu et al.

Agglomerative models have recently emerged as a powerful approach to training vision foundation models, leveraging multi-teacher distillation from existing models such as CLIP, DINO, and SAM. This strategy enables the efficient creation of robust models, combining the strengths of individual teachers while significantly reducing computational and resource demands. In this paper, we thoroughly analyze state-of-the-art agglomerative models, identifying critical challenges including resolution mode shifts, teacher imbalance, idiosyncratic teacher artifacts, and an excessive number of output tokens. To address these issues, we propose several novel solutions: multi-resolution training, mosaic augmentation, and improved balancing of teacher loss functions. Specifically, in the context of Vision Language Models, we introduce a token compression technique to maintain high-resolution information within a fixed token count. We release our top-performing variants at multiple scales (-B, -L, -H, and -g), along with inference code and pretrained weights

Shaokun Zhang, Yi Dong, Jieyu Zhang et al.

Enabling large language models with external tools has become a pivotal strategy for extending their functionality beyond text space. To enhance LLMs' tool-calling abilities, previous approaches primarily rely on supervised fine-tuning (SFT) with trajectories distilled from stronger models, often resulting in imitative reasoning that limits generalization. In this work, we explore rule-based reinforcement learning to enhance tool-calling in LLMs, resulting in Nemotron-Research-Tool-N1, a series of tool-calling reasoning models. Rather than enforcing supervision over intermediate distilled reasoning traces, Tool-N1 is trained with a binary RL reward that assesses only the format validity and functional correctness of tool invocations. This lightweight supervision allows the model to develop reasoning strategies independently, without relying on annotated trajectories. Experiments on several major benchmarks show that Tool-N1-7B/14B clearly outperform GPT-4o. We conduct a systematic study on the design of rule-based reinforcement learning strategies for training tool-calling models. Using 5,518 distilled reasoning trajectories, we compare SFT, RL, and the SFT-then-RL pipeline, finding that the widely adopted SFT-then-RL paradigm does not necessarily outperform pure RL.

Guo Chen, Lidong Lu, Yicheng Liu et al.

While datasets for video understanding have scaled to hour-long durations, they typically consist of densely concatenated clips that differ from natural, unscripted daily life. To bridge this gap, we introduce MM-Lifelong, a dataset designed for Multimodal Lifelong Understanding. Comprising 181.1 hours of footage, it is structured across Day, Week, and Month scales to capture varying temporal densities. Extensive evaluations reveal two critical failure modes in current paradigms: end-to-end MLLMs suffer from a Working Memory Bottleneck due to context saturation, while representative agentic baselines experience Global Localization Collapse when navigating sparse, month-long timelines. To address this, we propose the Recursive Multimodal Agent (ReMA), which employs dynamic memory management to iteratively update a recursive belief state, significantly outperforming existing methods. Finally, we establish dataset splits designed to isolate temporal and domain biases, providing a rigorous foundation for future research in supervised learning and out-of-distribution generalization.

Ilia Karmanov, Amala Sanjay Deshmukh, Lukas Voegtle et al.

Optical Character Recognition (OCR) technology is widely used to extract text from images of documents, facilitating efficient digitization and data retrieval. However, merely extracting text is insufficient when dealing with complex documents. Fully comprehending such documents requires an understanding of their structure -- including formatting, formulas, tables, and the reading order of multiple blocks and columns across multiple pages -- as well as semantic information for detecting elements like footnotes and image captions. This comprehensive understanding is crucial for downstream tasks such as retrieval, document question answering, and data curation for training Large Language Models (LLMs) and Vision Language Models (VLMs). To address this, we introduce Éclair, a general-purpose text-extraction tool specifically designed to process a wide range of document types. Given an image, Éclair is able to extract formatted text in reading order, along with bounding boxes and their corresponding semantic classes. To thoroughly evaluate these novel capabilities, we introduce our diverse human-annotated benchmark for document-level OCR and semantic classification. Éclair achieves state-of-the-art accuracy on this benchmark, outperforming other methods across key metrics. Additionally, we evaluate Éclair on established benchmarks, demonstrating its versatility and strength across several evaluation standards.

Arushi Goel, Karan Sapra, Matthieu Le et al.

Large Language Models (LLMs) have made significant strides in text generation and comprehension, with recent advancements extending into multimodal LLMs that integrate visual and audio inputs. However, these models continue to struggle with fine-grained, cross-modal temporal understanding, particularly when correlating events across audio and video streams. We address these challenges with two key contributions: a new dataset and model, called OCTAV and OMCAT respectively. OCTAV (Omni Context and Temporal Audio Video) is a novel dataset designed to capture event transitions across audio and video. Second, OMCAT (Omni Context Aware Transformer) is a powerful model that leverages RoTE (Rotary Time Embeddings), an innovative extension of RoPE, to enhance temporal grounding and computational efficiency in time-anchored tasks. Through a robust three-stage training pipeline-feature alignment, instruction tuning, and OCTAV-specific training-OMCAT excels in cross-modal temporal understanding. Our model demonstrates state-of-the-art performance on Audio-Visual Question Answering (AVQA) tasks and the OCTAV benchmark, showcasing significant gains in temporal reasoning and cross-modal alignment, as validated through comprehensive experiments and ablation studies. Our dataset and code will be made publicly available. The link to our demo page is https://om-cat.github.io.

Natan Bagrov, Eugene Khvedchenia, Borys Tymchenko et al.

Vision-language models (VLMs) have recently expanded from static image understanding to video reasoning, but their scalability is fundamentally limited by the quadratic cost of processing dense frame sequences. Long videos often exceed the token budget of modern language models, leading to severe context limitations and latency issues. We introduce Efficient Video Sampling (EVS), a simple, plug-and-play method for reducing token redundancy in videos by identifying and pruning temporally static patches -- spatial regions that remain unchanged across consecutive frames. EVS preserves positional identity, requires no architectural changes or retraining. We show that EVS substantially reduces token count while maintaining semantic fidelity, enabling faster inference and longer input sequences. Applied at inference time, EVS reduces large language model (LLM) time-to-first-token (TTFT) by up to 4x with minimal accuracy loss. When combined with an uptraining phase using stochastic pruning rates, EVS yields models that are robust to varying compression levels and retain full performance under aggressive pruning. Extensive experiments demonstrate that EVS consistently improves efficiency-accuracy trade-offs, unlocking scalable video-language understanding without sacrificing quality.

Ming-Chang Chiu, Fuxiao Liu, Karan Sapra et al.

The enhancement of Visual Language Models (VLMs) has traditionally relied on knowledge distillation from larger, more capable models. This dependence creates a fundamental bottleneck for improving state-of-the-art systems, particularly when no superior models exist. We introduce AIDE (Agentic Improvement through Domain Experts), a novel framework that enables VLMs to autonomously enhance their capabilities by leveraging specialized domain expert models. AIDE operates through a four-stage process: (1) identifying instances for refinement, (2) engaging domain experts for targeted analysis, (3) synthesizing expert outputs with existing data, and (4) integrating enhanced instances into the training pipeline. Experiments on multiple benchmarks, including MMMU, MME, MMBench, etc., demonstrate AIDE's ability to achieve notable performance gains without relying on larger VLMs nor human supervision. Our framework provides a scalable, resource-efficient approach to continuous VLM improvement, addressing critical limitations in current methodologies, particularly valuable when larger models are unavailable to access.

Kateryna Chumachenko, Amala Sanjay Deshmukh, Jarno Seppanen et al.

We introduce Nemotron-Parse-1.1, a lightweight document parsing and OCR model that advances the capabilities of its predecessor, Nemoretriever-Parse-1.0. Nemotron-Parse-1.1 delivers improved capabilities across general OCR, markdown formatting, structured table parsing, and text extraction from pictures, charts, and diagrams. It also supports a longer output sequence length for visually dense documents. As with its predecessor, it extracts bounding boxes of text segments, as well as corresponding semantic classes. Nemotron-Parse-1.1 follows an encoder-decoder architecture with 885M parameters, including a compact 256M-parameter language decoder. It achieves competitive accuracy on public benchmarks making it a strong lightweight OCR solution. We release the model weights publicly on Huggingface, as well as an optimized NIM container, along with a subset of the training data as part of the broader Nemotron-VLM-v2 dataset. Additionally, we release Nemotron-Parse-1.1-TC which operates on a reduced vision token length, offering a 20% speed improvement with minimal quality degradation.

Aysegul Dundar, Jun Gao, Andrew Tao et al.

This paper presents a method to reconstruct high-quality textured 3D models from single images. Current methods rely on datasets with expensive annotations; multi-view images and their camera parameters. Our method relies on GAN generated multi-view image datasets which have a negligible annotation cost. However, they are not strictly multi-view consistent and sometimes GANs output distorted images. This results in degraded reconstruction qualities. In this work, to overcome these limitations of generated datasets, we have two main contributions which lead us to achieve state-of-the-art results on challenging objects: 1) A robust multi-stage learning scheme that gradually relies more on the models own predictions when calculating losses, 2) A novel adversarial learning pipeline with online pseudo-ground truth generations to achieve fine details. Our work provides a bridge from 2D supervisions of GAN models to 3D reconstruction models and removes the expensive annotation efforts. We show significant improvements over previous methods whether they were trained on GAN generated multi-view images or on real images with expensive annotations. Please visit our web-page for 3D visuals: https://research.nvidia.com/labs/adlr/progressive-3d-learning

Songwei Ge, Seungjun Nah, Guilin Liu et al.

Despite tremendous progress in generating high-quality images using diffusion models, synthesizing a sequence of animated frames that are both photorealistic and temporally coherent is still in its infancy. While off-the-shelf billion-scale datasets for image generation are available, collecting similar video data of the same scale is still challenging. Also, training a video diffusion model is computationally much more expensive than its image counterpart. In this work, we explore finetuning a pretrained image diffusion model with video data as a practical solution for the video synthesis task. We find that naively extending the image noise prior to video noise prior in video diffusion leads to sub-optimal performance. Our carefully designed video noise prior leads to substantially better performance. Extensive experimental validation shows that our model, Preserve Your Own Correlation (PYoCo), attains SOTA zero-shot text-to-video results on the UCF-101 and MSR-VTT benchmarks. It also achieves SOTA video generation quality on the small-scale UCF-101 benchmark with a $10\times$ smaller model using significantly less computation than the prior art.

Max Ehrlich, Jon Barker, Namitha Padmanabhan et al.

Video compression is a central feature of the modern internet powering technologies from social media to video conferencing. While video compression continues to mature, for many compression settings, quality loss is still noticeable. These settings nevertheless have important applications to the efficient transmission of videos over bandwidth constrained or otherwise unstable connections. In this work, we develop a deep learning architecture capable of restoring detail to compressed videos which leverages the underlying structure and motion information embedded in the video bitstream. We show that this improves restoration accuracy compared to prior compression correction methods and is competitive when compared with recent deep-learning-based video compression methods on rate-distortion while achieving higher throughput. Furthermore, we condition our model on quantization data which is readily available in the bitstream. This allows our single model to handle a variety of different compression quality settings which required an ensemble of models in prior work.

John Guibas, Morteza Mardani, Zongyi Li et al.

Vision transformers have delivered tremendous success in representation learning. This is primarily due to effective token mixing through self attention. However, this scales quadratically with the number of pixels, which becomes infeasible for high-resolution inputs. To cope with this challenge, we propose Adaptive Fourier Neural Operator (AFNO) as an efficient token mixer that learns to mix in the Fourier domain. AFNO is based on a principled foundation of operator learning which allows us to frame token mixing as a continuous global convolution without any dependence on the input resolution. This principle was previously used to design FNO, which solves global convolution efficiently in the Fourier domain and has shown promise in learning challenging PDEs. To handle challenges in visual representation learning such as discontinuities in images and high resolution inputs, we propose principled architectural modifications to FNO which results in memory and computational efficiency. This includes imposing a block-diagonal structure on the channel mixing weights, adaptively sharing weights across tokens, and sparsifying the frequency modes via soft-thresholding and shrinkage. The resulting model is highly parallel with a quasi-linear complexity and has linear memory in the sequence size. AFNO outperforms self-attention mechanisms for few-shot segmentation in terms of both efficiency and accuracy. For Cityscapes segmentation with the Segformer-B3 backbone, AFNO can handle a sequence size of 65k and outperforms other efficient self-attention mechanisms.

Anand Bhattad, Aysegul Dundar, Guilin Liu et al.

Humans can easily infer the underlying 3D geometry and texture of an object only from a single 2D image. Current computer vision methods can do this, too, but suffer from view generalization problems - the models inferred tend to make poor predictions of appearance in novel views. As for generalization problems in machine learning, the difficulty is balancing single-view accuracy (cf. training error; bias) with novel view accuracy (cf. test error; variance). We describe a class of models whose geometric rigidity is easily controlled to manage this tradeoff. We describe a cycle consistency loss that improves view generalization (roughly, a model from a generated view should predict the original view well). View generalization of textures requires that models share texture information, so a car seen from the back still has headlights because other cars have headlights. We describe a cycle consistency loss that encourages model textures to be aligned, so as to encourage sharing. We compare our method against the state-of-the-art method and show both qualitative and quantitative improvements.

Guilin Liu, Rohan Taori, Ting-Chun Wang et al.

Conventional CNNs for texture synthesis consist of a sequence of (de)-convolution and up/down-sampling layers, where each layer operates locally and lacks the ability to capture the long-term structural dependency required by texture synthesis. Thus, they often simply enlarge the input texture, rather than perform reasonable synthesis. As a compromise, many recent methods sacrifice generalizability by training and testing on the same single (or fixed set of) texture image(s), resulting in huge re-training time costs for unseen images. In this work, based on the discovery that the assembling/stitching operation in traditional texture synthesis is analogous to a transposed convolution operation, we propose a novel way of using transposed convolution operation. Specifically, we directly treat the whole encoded feature map of the input texture as transposed convolution filters and the features' self-similarity map, which captures the auto-correlation information, as input to the transposed convolution. Such a design allows our framework, once trained, to be generalizable to perform synthesis of unseen textures with a single forward pass in nearly real-time. Our method achieves state-of-the-art texture synthesis quality based on various metrics. While self-similarity helps preserve the input textures' regular structural patterns, our framework can also take random noise maps for irregular input textures instead of self-similarity maps as transposed convolution inputs. It allows to get more diverse results as well as generate arbitrarily large texture outputs by directly sampling large noise maps in a single pass as well.

Andrew Tao, Karan Sapra, Bryan Catanzaro

Multi-scale inference is commonly used to improve the results of semantic segmentation. Multiple images scales are passed through a network and then the results are combined with averaging or max pooling. In this work, we present an attention-based approach to combining multi-scale predictions. We show that predictions at certain scales are better at resolving particular failures modes, and that the network learns to favor those scales for such cases in order to generate better predictions. Our attention mechanism is hierarchical, which enables it to be roughly 4x more memory efficient to train than other recent approaches. In addition to enabling faster training, this allows us to train with larger crop sizes which leads to greater model accuracy. We demonstrate the result of our method on two datasets: Cityscapes and Mapillary Vistas. For Cityscapes, which has a large number of weakly labelled images, we also leverage auto-labelling to improve generalization. Using our approach we achieve a new state-of-the-art results in both Mapillary (61.1 IOU val) and Cityscapes (85.1 IOU test).

Aysegul Dundar, Karan Sapra, Guilin Liu et al.

Conditional image synthesis for generating photorealistic images serves various applications for content editing to content generation. Previous conditional image synthesis algorithms mostly rely on semantic maps, and often fail in complex environments where multiple instances occlude each other. We propose a panoptic aware image synthesis network to generate high fidelity and photorealistic images conditioned on panoptic maps which unify semantic and instance information. To achieve this, we efficiently use panoptic maps in convolution and upsampling layers. We show that with the proposed changes to the generator, we can improve on the previous state-of-the-art methods by generating images in complex instance interaction environments in higher fidelity and tiny objects in more details. Furthermore, our proposed method also outperforms the previous state-of-the-art methods in metrics of mean IoU (Intersection over Union), and detAP (Detection Average Precision).

Aysegul Dundar, Kevin J. Shih, Animesh Garg et al.

Unsupervised landmark learning is the task of learning semantic keypoint-like representations without the use of expensive input keypoint-level annotations. A popular approach is to factorize an image into a pose and appearance data stream, then to reconstruct the image from the factorized components. The pose representation should capture a set of consistent and tightly localized landmarks in order to facilitate reconstruction of the input image. Ultimately, we wish for our learned landmarks to focus on the foreground object of interest. However, the reconstruction task of the entire image forces the model to allocate landmarks to model the background. This work explores the effects of factorizing the reconstruction task into separate foreground and background reconstructions, conditioning only the foreground reconstruction on the unsupervised landmarks. Our experiments demonstrate that the proposed factorization results in landmarks that are focused on the foreground object of interest. Furthermore, the rendered background quality is also improved, as the background rendering pipeline no longer requires the ill-suited landmarks to model its pose and appearance. We demonstrate this improvement in the context of the video-prediction task.

Rafael Valle, Fitsum Reda, Mohammad Shoeybi et al.

We propose a novel approach for image segmentation that combines Neural Ordinary Differential Equations (NODEs) and the Level Set method. Our approach parametrizes the evolution of an initial contour with a NODE that implicitly learns from data a speed function describing the evolution. In addition, for cases where an initial contour is not available and to alleviate the need for careful choice or design of contour embedding functions, we propose a NODE-based method that evolves an image embedding into a dense per-pixel semantic label space. We evaluate our methods on kidney segmentation (KiTS19) and on salient object detection (PASCAL-S, ECSSD and HKU-IS). In addition to improving initial contours provided by deep learning models while using a fraction of their number of parameters, our approach achieves F scores that are higher than several state-of-the-art deep learning algorithms.

Ting-Chun Wang, Ming-Yu Liu, Andrew Tao et al.

Video-to-video synthesis (vid2vid) aims at converting an input semantic video, such as videos of human poses or segmentation masks, to an output photorealistic video. While the state-of-the-art of vid2vid has advanced significantly, existing approaches share two major limitations. First, they are data-hungry. Numerous images of a target human subject or a scene are required for training. Second, a learned model has limited generalization capability. A pose-to-human vid2vid model can only synthesize poses of the single person in the training set. It does not generalize to other humans that are not in the training set. To address the limitations, we propose a few-shot vid2vid framework, which learns to synthesize videos of previously unseen subjects or scenes by leveraging few example images of the target at test time. Our model achieves this few-shot generalization capability via a novel network weight generation module utilizing an attention mechanism. We conduct extensive experimental validations with comparisons to strong baselines using several large-scale video datasets including human-dancing videos, talking-head videos, and street-scene videos. The experimental results verify the effectiveness of the proposed framework in addressing the two limitations of existing vid2vid approaches.

Kevin J. Shih, Aysegul Dundar, Animesh Garg et al.

Prediction and interpolation for long-range video data involves the complex task of modeling motion trajectories for each visible object, occlusions and dis-occlusions, as well as appearance changes due to viewpoint and lighting. Optical flow based techniques generalize but are suitable only for short temporal ranges. Many methods opt to project the video frames to a low dimensional latent space, achieving long-range predictions. However, these latent representations are often non-interpretable, and therefore difficult to manipulate. This work poses video prediction and interpolation as unsupervised latent structure inference followed by a temporal prediction in this latent space. The latent representations capture foreground semantics without explicit supervision such as keypoints or poses. Further, as each landmark can be mapped to a coordinate indicating where a semantic part is positioned, we can reliably interpolate within the coordinate domain to achieve predictable motion interpolation. Given an image decoder capable of mapping these landmarks back to the image domain, we are able to achieve high-quality long-range video interpolation and extrapolation by operating on the landmark representation space.

Fitsum A. Reda, Deqing Sun, Aysegul Dundar et al.

Learning to synthesize high frame rate videos via interpolation requires large quantities of high frame rate training videos, which, however, are scarce, especially at high resolutions. Here, we propose unsupervised techniques to synthesize high frame rate videos directly from low frame rate videos using cycle consistency. For a triplet of consecutive frames, we optimize models to minimize the discrepancy between the center frame and its cycle reconstruction, obtained by interpolating back from interpolated intermediate frames. This simple unsupervised constraint alone achieves results comparable with supervision using the ground truth intermediate frames. We further introduce a pseudo supervised loss term that enforces the interpolated frames to be consistent with predictions of a pre-trained interpolation model. The pseudo supervised loss term, used together with cycle consistency, can effectively adapt a pre-trained model to a new target domain. With no additional data and in a completely unsupervised fashion, our techniques significantly improve pre-trained models on new target domains, increasing PSNR values from 32.84dB to 33.05dB on the Slowflow and from 31.82dB to 32.53dB on the Sintel evaluation datasets.

Ji Zhang, Kevin J. Shih, Ahmed Elgammal et al.

Most scene graph parsers use a two-stage pipeline to detect visual relationships: the first stage detects entities, and the second predicts the predicate for each entity pair using a softmax distribution. We find that such pipelines, trained with only a cross entropy loss over predicate classes, suffer from two common errors. The first, Entity Instance Confusion, occurs when the model confuses multiple instances of the same type of entity (e.g. multiple cups). The second, Proximal Relationship Ambiguity, arises when multiple subject-predicate-object triplets appear in close proximity with the same predicate, and the model struggles to infer the correct subject-object pairings (e.g. mis-pairing musicians and their instruments). We propose a set of contrastive loss formulations that specifically target these types of errors within the scene graph parsing problem, collectively termed the Graphical Contrastive Losses. These losses explicitly force the model to disambiguate related and unrelated instances through margin constraints specific to each type of confusion. We further construct a relationship detector, called RelDN, using the aforementioned pipeline to demonstrate the efficacy of our proposed losses. Our model outperforms the winning method of the OpenImages Relationship Detection Challenge by 4.7\% (16.5\% relative) on the test set. We also show improved results over the best previous methods on the Visual Genome and Visual Relationship Detection datasets.

Yi Zhu, Karan Sapra, Fitsum A. Reda et al.

Semantic segmentation requires large amounts of pixel-wise annotations to learn accurate models. In this paper, we present a video prediction-based methodology to scale up training sets by synthesizing new training samples in order to improve the accuracy of semantic segmentation networks. We exploit video prediction models' ability to predict future frames in order to also predict future labels. A joint propagation strategy is also proposed to alleviate mis-alignments in synthesized samples. We demonstrate that training segmentation models on datasets augmented by the synthesized samples leads to significant improvements in accuracy. Furthermore, we introduce a novel boundary label relaxation technique that makes training robust to annotation noise and propagation artifacts along object boundaries. Our proposed methods achieve state-of-the-art mIoUs of 83.5% on Cityscapes and 82.9% on CamVid. Our single model, without model ensembles, achieves 72.8% mIoU on the KITTI semantic segmentation test set, which surpasses the winning entry of the ROB challenge 2018. Our code and videos can be found at https://nv-adlr.github.io/publication/2018-Segmentation.

Guilin Liu, Kevin J. Shih, Ting-Chun Wang et al.

In this paper, we present a simple yet effective padding scheme that can be used as a drop-in module for existing convolutional neural networks. We call it partial convolution based padding, with the intuition that the padded region can be treated as holes and the original input as non-holes. Specifically, during the convolution operation, the convolution results are re-weighted near image borders based on the ratios between the padded area and the convolution sliding window area. Extensive experiments with various deep network models on ImageNet classification and semantic segmentation demonstrate that the proposed padding scheme consistently outperforms standard zero padding with better accuracy.

Ji Zhang, Kevin Shih, Andrew Tao et al.

We propose an efficient and interpretable scene graph generator. We consider three types of features: visual, spatial and semantic, and we use a late fusion strategy such that each feature's contribution can be explicitly investigated. We study the key factors about these features that have the most impact on the performance, and also visualize the learned visual features for relationships and investigate the efficacy of our model. We won the champion of the OpenImages Visual Relationship Detection Challenge on Kaggle, where we outperform the 2nd place by 5\% (20\% relatively). We believe an accurate scene graph generator is a fundamental stepping stone for higher-level vision-language tasks such as image captioning and visual QA, since it provides a semantic, structured comprehension of an image that is beyond pixels and objects.

Fitsum A. Reda, Guilin Liu, Kevin J. Shih et al.

We present an approach for high-resolution video frame prediction by conditioning on both past frames and past optical flows. Previous approaches rely on resampling past frames, guided by a learned future optical flow, or on direct generation of pixels. Resampling based on flow is insufficient because it cannot deal with disocclusions. Generative models currently lead to blurry results. Recent approaches synthesis a pixel by convolving input patches with a predicted kernel. However, their memory requirement increases with kernel size. Here, we spatially-displaced convolution (SDC) module for video frame prediction. We learn a motion vector and a kernel for each pixel and synthesize a pixel by applying the kernel at a displaced location in the source image, defined by the predicted motion vector. Our approach inherits the merits of both vector-based and kernel-based approaches, while ameliorating their respective disadvantages. We train our model on 428K unlabelled 1080p video game frames. Our approach produces state-of-the-art results, achieving an SSIM score of 0.904 on high-definition YouTube-8M videos, 0.918 on Caltech Pedestrian videos. Our model handles large motion effectively and synthesizes crisp frames with consistent motion.

Ji Zhang, Kevin Shih, Andrew Tao et al.

This article describes the model we built that achieved 1st place in the OpenImage Visual Relationship Detection Challenge on Kaggle. Three key factors contribute the most to our success: 1) language bias is a powerful baseline for this task. We build the empirical distribution $P(predicate|subject,object)$ in the training set and directly use that in testing. This baseline achieved the 2nd place when submitted; 2) spatial features are as important as visual features, especially for spatial relationships such as "under" and "inside of"; 3) It is a very effective way to fuse different features by first building separate modules for each of them, then adding their output logits before the final softmax layer. We show in ablation study that each factor can improve the performance to a non-trivial extent, and the model reaches optimal when all of them are combined.

Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu et al.

We study the problem of video-to-video synthesis, whose goal is to learn a mapping function from an input source video (e.g., a sequence of semantic segmentation masks) to an output photorealistic video that precisely depicts the content of the source video. While its image counterpart, the image-to-image synthesis problem, is a popular topic, the video-to-video synthesis problem is less explored in the literature. Without understanding temporal dynamics, directly applying existing image synthesis approaches to an input video often results in temporally incoherent videos of low visual quality. In this paper, we propose a novel video-to-video synthesis approach under the generative adversarial learning framework. Through carefully-designed generator and discriminator architectures, coupled with a spatio-temporal adversarial objective, we achieve high-resolution, photorealistic, temporally coherent video results on a diverse set of input formats including segmentation masks, sketches, and poses. Experiments on multiple benchmarks show the advantage of our method compared to strong baselines. In particular, our model is capable of synthesizing 2K resolution videos of street scenes up to 30 seconds long, which significantly advances the state-of-the-art of video synthesis. Finally, we apply our approach to future video prediction, outperforming several state-of-the-art competing systems.

Guilin Liu, Fitsum A. Reda, Kevin J. Shih et al.

Existing deep learning based image inpainting methods use a standard convolutional network over the corrupted image, using convolutional filter responses conditioned on both valid pixels as well as the substitute values in the masked holes (typically the mean value). This often leads to artifacts such as color discrepancy and blurriness. Post-processing is usually used to reduce such artifacts, but are expensive and may fail. We propose the use of partial convolutions, where the convolution is masked and renormalized to be conditioned on only valid pixels. We further include a mechanism to automatically generate an updated mask for the next layer as part of the forward pass. Our model outperforms other methods for irregular masks. We show qualitative and quantitative comparisons with other methods to validate our approach.

Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu et al.

We present a new method for synthesizing high-resolution photo-realistic images from semantic label maps using conditional generative adversarial networks (conditional GANs). Conditional GANs have enabled a variety of applications, but the results are often limited to low-resolution and still far from realistic. In this work, we generate 2048x1024 visually appealing results with a novel adversarial loss, as well as new multi-scale generator and discriminator architectures. Furthermore, we extend our framework to interactive visual manipulation with two additional features. First, we incorporate object instance segmentation information, which enables object manipulations such as removing/adding objects and changing the object category. Second, we propose a method to generate diverse results given the same input, allowing users to edit the object appearance interactively. Human opinion studies demonstrate that our method significantly outperforms existing methods, advancing both the quality and the resolution of deep image synthesis and editing.