Wonmin Byeon

Yitong Jiang, Hongjun Wang, Collin McCarthy et al.

Vision foundation models are bottlenecked by the quadratic cost of self-attention, which limits usable resolution and increases the cost of large-scale pretraining. Subquadratic alternatives such as linear attention and state-space models reduce this cost, but often serialize images into 1D token streams and weaken the 2D spatial structure important for vision. Generalized Spatial Propagation Networks (GSPN) instead propagate context directly on the 2D grid through line-scan recurrences, achieving near-linear complexity without positional embeddings, but have seen little use as foundation-scale encoders. We present C-GSPN, a foundation-scale vision encoder based on 2D spatial propagation. C-GSPN makes the operator practical through three improvements: (1) a fast GSPN CUDA kernel that fuses per-step launches into a single warp-specialized implementation with shared-memory tiling, coalesced access, and a compact multi-channel propagation, reaching over 90% of peak memory bandwidth and running up to 40--52x faster than the original GSPN implementation; (2) a compressed latent-space propagation block with fused normalization, which turns kernel-level speed into block- and model-level efficiency; and (3) a two-stage cross-operator distillation recipe that trains the new architecture from an attention teacher without the cost of from-scratch foundation-scale training. Distilled with 600M image-text pairs, C-GSPN matches an isomorphic ViT baseline with 15% fewer parameters, improves ADE20K segmentation by +2.1%, transfers to high resolution with a fraction of the data needed from scratch, and delivers a 4x end-to-end block speedup at 2K with single-pass, tiling-free inference.

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.

Jiarui Xu, Sifei Liu, Arash Vahdat et al.

We present ODISE: Open-vocabulary DIffusion-based panoptic SEgmentation, which unifies pre-trained text-image diffusion and discriminative models to perform open-vocabulary panoptic segmentation. Text-to-image diffusion models have the remarkable ability to generate high-quality images with diverse open-vocabulary language descriptions. This demonstrates that their internal representation space is highly correlated with open concepts in the real world. Text-image discriminative models like CLIP, on the other hand, are good at classifying images into open-vocabulary labels. We leverage the frozen internal representations of both these models to perform panoptic segmentation of any category in the wild. Our approach outperforms the previous state of the art by significant margins on both open-vocabulary panoptic and semantic segmentation tasks. In particular, with COCO training only, our method achieves 23.4 PQ and 30.0 mIoU on the ADE20K dataset, with 8.3 PQ and 7.9 mIoU absolute improvement over the previous state of the art. We open-source our code and models at https://github.com/NVlabs/ODISE .

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.

Jimmy T. H. Smith, Shalini De Mello, Jan Kautz et al.

Effectively modeling long spatiotemporal sequences is challenging due to the need to model complex spatial correlations and long-range temporal dependencies simultaneously. ConvLSTMs attempt to address this by updating tensor-valued states with recurrent neural networks, but their sequential computation makes them slow to train. In contrast, Transformers can process an entire spatiotemporal sequence, compressed into tokens, in parallel. However, the cost of attention scales quadratically in length, limiting their scalability to longer sequences. Here, we address the challenges of prior methods and introduce convolutional state space models (ConvSSM) that combine the tensor modeling ideas of ConvLSTM with the long sequence modeling approaches of state space methods such as S4 and S5. First, we demonstrate how parallel scans can be applied to convolutional recurrences to achieve subquadratic parallelization and fast autoregressive generation. We then establish an equivalence between the dynamics of ConvSSMs and SSMs, which motivates parameterization and initialization strategies for modeling long-range dependencies. The result is ConvS5, an efficient ConvSSM variant for long-range spatiotemporal modeling. ConvS5 significantly outperforms Transformers and ConvLSTM on a long horizon Moving-MNIST experiment while training 3X faster than ConvLSTM and generating samples 400X faster than Transformers. In addition, ConvS5 matches or exceeds the performance of state-of-the-art methods on challenging DMLab, Minecraft and Habitat prediction benchmarks and enables new directions for modeling long spatiotemporal sequences.

Seung Hyun Lee, Gyeongrok Oh, Wonmin Byeon et al.

The recent success in StyleGAN demonstrates that pre-trained StyleGAN latent space is useful for realistic video generation. However, the generated motion in the video is usually not semantically meaningful due to the difficulty of determining the direction and magnitude in the StyleGAN latent space. In this paper, we propose a framework to generate realistic videos by leveraging multimodal (sound-image-text) embedding space. As sound provides the temporal contexts of the scene, our framework learns to generate a video that is semantically consistent with sound. First, our sound inversion module maps the audio directly into the StyleGAN latent space. We then incorporate the CLIP-based multimodal embedding space to further provide the audio-visual relationships. Finally, the proposed frame generator learns to find the trajectory in the latent space which is coherent with the corresponding sound and generates a video in a hierarchical manner. We provide the new high-resolution landscape video dataset (audio-visual pair) for the sound-guided video generation task. The experiments show that our model outperforms the state-of-the-art methods in terms of video quality. We further show several applications including image and video editing to verify the effectiveness of our method.

Yujin Jeong, Wonjeong Ryoo, Seunghyun Lee et al.

In recent years, video generation has become a prominent generative tool and has drawn significant attention. However, there is little consideration in audio-to-video generation, though audio contains unique qualities like temporal semantics and magnitude. Hence, we propose The Power of Sound (TPoS) model to incorporate audio input that includes both changeable temporal semantics and magnitude. To generate video frames, TPoS utilizes a latent stable diffusion model with textual semantic information, which is then guided by the sequential audio embedding from our pretrained Audio Encoder. As a result, this method produces audio reactive video contents. We demonstrate the effectiveness of TPoS across various tasks and compare its results with current state-of-the-art techniques in the field of audio-to-video generation. More examples are available at https://ku-vai.github.io/TPoS/

Divyam Madaan, Hongxu Yin, Wonmin Byeon et al.

We propose a novel framework and a solution to tackle the continual learning (CL) problem with changing network architectures. Most CL methods focus on adapting a single architecture to a new task/class by modifying its weights. However, with rapid progress in architecture design, the problem of adapting existing solutions to novel architectures becomes relevant. To address this limitation, we propose Heterogeneous Continual Learning (HCL), where a wide range of evolving network architectures emerge continually together with novel data/tasks. As a solution, we build on top of the distillation family of techniques and modify it to a new setting where a weaker model takes the role of a teacher; meanwhile, a new stronger architecture acts as a student. Furthermore, we consider a setup of limited access to previous data and propose Quick Deep Inversion (QDI) to recover prior task visual features to support knowledge transfer. QDI significantly reduces computational costs compared to previous solutions and improves overall performance. In summary, we propose a new setup for CL with a modified knowledge distillation paradigm and design a quick data inversion method to enhance distillation. Our evaluation of various benchmarks shows a significant improvement on accuracy in comparison to state-of-the-art methods over various networks architectures.

Seung Hyun Lee, Gyeongrok Oh, Wonmin Byeon et al.

Recent successes suggest that an image can be manipulated by a text prompt, e.g., a landscape scene on a sunny day is manipulated into the same scene on a rainy day driven by a text input "raining". These approaches often utilize a StyleCLIP-based image generator, which leverages multi-modal (text and image) embedding space. However, we observe that such text inputs are often bottlenecked in providing and synthesizing rich semantic cues, e.g., differentiating heavy rain from rain with thunderstorms. To address this issue, we advocate leveraging an additional modality, sound, which has notable advantages in image manipulation as it can convey more diverse semantic cues (vivid emotions or dynamic expressions of the natural world) than texts. In this paper, we propose a novel approach that first extends the image-text joint embedding space with sound and applies a direct latent optimization method to manipulate a given image based on audio input, e.g., the sound of rain. Our extensive experiments show that our sound-guided image manipulation approach produces semantically and visually more plausible manipulation results than the state-of-the-art text and sound-guided image manipulation methods, which are further confirmed by our human evaluations. Our downstream task evaluations also show that our learned image-text-sound joint embedding space effectively encodes sound inputs.

Seung Hyun Lee, Chanyoung Kim, Wonmin Byeon et al.

We present a novel framework, Localized Image Stylization with Audio (LISA) which performs audio-driven localized image stylization. Sound often provides information about the specific context of the scene and is closely related to a certain part of the scene or object. However, existing image stylization works have focused on stylizing the entire image using an image or text input. Stylizing a particular part of the image based on audio input is natural but challenging. In this work, we propose a framework that a user provides an audio input to localize the sound source in the input image and another for locally stylizing the target object or scene. LISA first produces a delicate localization map with an audio-visual localization network by leveraging CLIP embedding space. We then utilize implicit neural representation (INR) along with the predicted localization map to stylize the target object or scene based on sound information. The proposed INR can manipulate the localized pixel values to be semantically consistent with the provided audio input. Through a series of experiments, we show that the proposed framework outperforms the other audio-guided stylization methods. Moreover, LISA constructs concise localization maps and naturally manipulates the target object or scene in accordance with the given audio input.

Jean Kossaifi, Nikola Kovachki, Morteza Mardani et al.

The recent revolution in data-driven methods for weather forecasting has lead to a fragmented landscape of complex, bespoke architectures and training strategies, obscuring the fundamental drivers of forecast accuracy. Here, we demonstrate that state-of-the-art probabilistic skill requires neither intricate architectural constraints nor specialized training heuristics. We introduce a scalable framework for learning multi-scale atmospheric dynamics by combining a directly downsampled latent space with a history-conditioned local projector that resolves high-resolution physics. We find that our framework design is robust to the choice of probabilistic estimator, seamlessly supporting stochastic interpolants, diffusion models, and CRPS-based ensemble training. Validated against the Integrated Forecasting System and the deep learning probabilistic model GenCast, our framework achieves statistically significant improvements on most of the variables. These results suggest scaling a general-purpose model is sufficient for state-of-the-art medium-range prediction, eliminating the need for tailored training recipes and proving effective across the full spectrum of probabilistic frameworks.

Hyeongcheol Park, Jiyoung Seo, Jaewon Mun et al.

Retrieval-Augmented Generation (RAG) has recently been extended to multimodal settings, connecting multimodal large language models (MLLMs) with vast corpora of external knowledge such as multimodal knowledge graphs (MMKGs). Despite their recent success, multimodal RAG in the audio-visual domain remains challenging due to 1) limited modality coverage and multi-hop connectivity of existing MMKGs, and 2) retrieval based solely on similarity in a shared multimodal embedding space, which fails to filter out off-topic or redundant knowledge. To address these limitations, we propose M$^3$KG-RAG, a Multi-hop Multimodal Knowledge Graph-enhanced RAG that retrieves query-aligned audio-visual knowledge from MMKGs, improving reasoning depth and answer faithfulness in MLLMs. Specifically, we devise a lightweight multi-agent pipeline to construct multi-hop MMKG (M$^3$KG), which contains context-enriched triplets of multimodal entities, enabling modality-wise retrieval based on input queries. Furthermore, we introduce GRASP (Grounded Retrieval And Selective Pruning), which ensures precise entity grounding to the query, evaluates answer-supporting relevance, and prunes redundant context to retain only knowledge essential for response generation. Extensive experiments across diverse multimodal benchmarks demonstrate that M$^3$KG-RAG significantly enhances MLLMs' multimodal reasoning and grounding over existing approaches.

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.

Jiarui Xu, Shalini De Mello, Sifei Liu et al.

Grouping and recognition are important components of visual scene understanding, e.g., for object detection and semantic segmentation. With end-to-end deep learning systems, grouping of image regions usually happens implicitly via top-down supervision from pixel-level recognition labels. Instead, in this paper, we propose to bring back the grouping mechanism into deep networks, which allows semantic segments to emerge automatically with only text supervision. We propose a hierarchical Grouping Vision Transformer (GroupViT), which goes beyond the regular grid structure representation and learns to group image regions into progressively larger arbitrary-shaped segments. We train GroupViT jointly with a text encoder on a large-scale image-text dataset via contrastive losses. With only text supervision and without any pixel-level annotations, GroupViT learns to group together semantic regions and successfully transfers to the task of semantic segmentation in a zero-shot manner, i.e., without any further fine-tuning. It achieves a zero-shot accuracy of 52.3% mIoU on the PASCAL VOC 2012 and 22.4% mIoU on PASCAL Context datasets, and performs competitively to state-of-the-art transfer-learning methods requiring greater levels of supervision. We open-source our code at https://github.com/NVlabs/GroupViT .

Rakshit Kothari, Shalini De Mello, Umar Iqbal et al.

A major challenge for physically unconstrained gaze estimation is acquiring training data with 3D gaze annotations for in-the-wild and outdoor scenarios. In contrast, videos of human interactions in unconstrained environments are abundantly available and can be much more easily annotated with frame-level activity labels. In this work, we tackle the previously unexplored problem of weakly-supervised gaze estimation from videos of human interactions. We leverage the insight that strong gaze-related geometric constraints exist when people perform the activity of "looking at each other" (LAEO). To acquire viable 3D gaze supervision from LAEO labels, we propose a training algorithm along with several novel loss functions especially designed for the task. With weak supervision from two large scale CMU-Panoptic and AVA-LAEO activity datasets, we show significant improvements in (a) the accuracy of semi-supervised gaze estimation and (b) cross-domain generalization on the state-of-the-art physically unconstrained in-the-wild Gaze360 gaze estimation benchmark. We open source our code at https://github.com/NVlabs/weakly-supervised-gaze.

Oliver Hennigh, Susheela Narasimhan, Mohammad Amin Nabian et al.

We present SimNet, an AI-driven multi-physics simulation framework, to accelerate simulations across a wide range of disciplines in science and engineering. Compared to traditional numerical solvers, SimNet addresses a wide range of use cases - coupled forward simulations without any training data, inverse and data assimilation problems. SimNet offers fast turnaround time by enabling parameterized system representation that solves for multiple configurations simultaneously, as opposed to the traditional solvers that solve for one configuration at a time. SimNet is integrated with parameterized constructive solid geometry as well as STL modules to generate point clouds. Furthermore, it is customizable with APIs that enable user extensions to geometry, physics and network architecture. It has advanced network architectures that are optimized for high-performance GPU computing, and offers scalable performance for multi-GPU and multi-Node implementation with accelerated linear algebra as well as FP32, FP64 and TF32 computations. In this paper we review the neural network solver methodology, the SimNet architecture, and the various features that are needed for effective solution of the PDEs. We present real-world use cases that range from challenging forward multi-physics simulations with turbulence and complex 3D geometries, to industrial design optimization and inverse problems that are not addressed efficiently by the traditional solvers. Extensive comparisons of SimNet results with open source and commercial solvers show good correlation.

Qiushan Guo, Shalini De Mello, Hongxu Yin et al.

Vision language models (VLMs) have experienced rapid advancements through the integration of large language models (LLMs) with image-text pairs, yet they struggle with detailed regional visual understanding due to limited spatial awareness of the vision encoder, and the use of coarse-grained training data that lacks detailed, region-specific captions. To address this, we introduce RegionGPT (short as RGPT), a novel framework designed for complex region-level captioning and understanding. RGPT enhances the spatial awareness of regional representation with simple yet effective modifications to existing visual encoders in VLMs. We further improve performance on tasks requiring a specific output scope by integrating task-guided instruction prompts during both training and inference phases, while maintaining the model's versatility for general-purpose tasks. Additionally, we develop an automated region caption data generation pipeline, enriching the training set with detailed region-level captions. We demonstrate that a universal RGPT model can be effectively applied and significantly enhancing performance across a range of region-level tasks, including but not limited to complex region descriptions, reasoning, object classification, and referring expressions comprehension.

Xin Dong, Yonggan Fu, Shizhe Diao et al.

We propose Hymba, a family of small language models featuring a hybrid-head parallel architecture that integrates transformer attention mechanisms with state space models (SSMs) for enhanced efficiency. Attention heads provide high-resolution recall, while SSM heads enable efficient context summarization. Additionally, we introduce learnable meta tokens that are prepended to prompts, storing critical information and alleviating the "forced-to-attend" burden associated with attention mechanisms. This model is further optimized by incorporating cross-layer key-value (KV) sharing and partial sliding window attention, resulting in a compact cache size. During development, we conducted a controlled study comparing various architectures under identical settings and observed significant advantages of our proposed architecture. Notably, Hymba achieves state-of-the-art results for small LMs: Our Hymba-1.5B-Base model surpasses all sub-2B public models in performance and even outperforms Llama-3.2-3B with 1.32% higher average accuracy, an 11.67x cache size reduction, and 3.49x throughput.

Gyeongrok Oh, Jaehwan Jeong, Sieun Kim et al.

We introduce a novel diffusion-based video generation method, generating a video showing multiple events given multiple individual sentences from the user. Our method does not require a large-scale video dataset since our method uses a pre-trained diffusion-based text-to-video generative model without a fine-tuning process. Specifically, we propose a last frame-aware diffusion process to preserve visual coherence between consecutive videos where each video consists of different events by initializing the latent and simultaneously adjusting noise in the latent to enhance the motion dynamic in a generated video. Furthermore, we find that the iterative update of latent vectors by referring to all the preceding frames maintains the global appearance across the frames in a video clip. To handle dynamic text input for video generation, we utilize a novel prompt generator that transfers course text messages from the user into the multiple optimal prompts for the text-to-video diffusion model. Extensive experiments and user studies show that our proposed method is superior to other video-generative models in terms of temporal coherency of content and semantics. Video examples are available on our project page: https://kuai-lab.github.io/eccv2024mevg.

Jindong Jiang, Xiuyu Li, Zhijian Liu et al.

Recent advances in video-based multimodal large language models (Video-LLMs) have significantly improved video understanding by processing videos as sequences of image frames. However, many existing methods treat frames independently in the vision backbone, lacking explicit temporal modeling, which limits their ability to capture dynamic patterns and efficiently handle long videos. To address these limitations, we introduce STORM (Spatiotemporal TOken Reduction for Multimodal LLMs), a novel architecture incorporating a dedicated temporal encoder between the image encoder and the LLM. Our temporal encoder leverages the Mamba State Space Model to integrate temporal information into image tokens, generating enriched representations that preserve inter-frame dynamics across the entire video sequence. This enriched encoding not only enhances video reasoning capabilities but also enables effective token reduction strategies, including test-time sampling and training-based temporal and spatial pooling, substantially reducing computational demands on the LLM without sacrificing key temporal information. By integrating these techniques, our approach simultaneously reduces training and inference latency while improving performance, enabling efficient and robust video understanding over extended temporal contexts. Extensive evaluations show that STORM achieves state-of-the-art results across various long video understanding benchmarks (more than 5% improvement on MLVU and LongVideoBench) while reducing the computation costs by up to $8\times$ and the decoding latency by 2.4-2.9$\times$ for the fixed numbers of input frames. Project page is available at https://research.nvidia.com/labs/lpr/storm

Hongjun Wang, Wonmin Byeon, Jiarui Xu et al.

We present the Generalized Spatial Propagation Network (GSPN), a new attention mechanism optimized for vision tasks that inherently captures 2D spatial structures. Existing attention models, including transformers, linear attention, and state-space models like Mamba, process multi-dimensional data as 1D sequences, compromising spatial coherence and efficiency. GSPN overcomes these limitations by directly operating on spatially coherent image data and forming dense pairwise connections through a line-scan approach. Central to GSPN is the Stability-Context Condition, which ensures stable, context-aware propagation across 2D sequences and reduces the effective sequence length to $\sqrt{N}$ for a square map with N elements, significantly enhancing computational efficiency. With learnable, input-dependent weights and no reliance on positional embeddings, GSPN achieves superior spatial fidelity and state-of-the-art performance in vision tasks, including ImageNet classification, class-guided image generation, and text-to-image generation. Notably, GSPN accelerates SD-XL with softmax-attention by over $84\times$ when generating 16K images.

Hongjun Wang, Yitong Jiang, Collin McCarthy et al.

Efficient vision transformer remains a bottleneck for high-resolution images and long-video related real-world applications. Generalized Spatial Propagation Network (GSPN) addresses this by replacing quadratic self-attention with a line-scan propagation scheme, bringing the cost close to linear in the number of rows or columns, while retaining accuracy. Despite this advancement, the existing GSPN implementation still suffers from (i) heavy overhead due to repeatedly launching GPU kernels, (ii) excessive data transfers from global GPU memory, and (iii) redundant computations caused by maintaining separate propagation weights for each channel. We introduce GSPN-2, a joint algorithm-system redesign. In particular, we eliminate thousands of micro-launches from the previous implementation into one single 2D kernel, explicitly pin one warp to each channel slice, and stage the previous column's activations in shared memory. On the model side, we introduce a compact channel propagation strategy that replaces per-channel matrices, trimming parameters, and align naturally with the affinity map used in transformer attention. Experiments demonstrate GSPN-2's effectiveness across image classification and text-to-image synthesis tasks, matching transformer-level accuracy with significantly lower computational cost. GSPN-2 establishes a new efficiency frontier for modeling global spatial context in vision applications through its unique combination of structured matrix transformations and GPU-optimized implementation. Project page: https://whj363636.github.io/GSPN2/

Yonggan Fu, Xin Dong, Shizhe Diao et al.

Efficient deployment of small language models (SLMs) is essential for numerous real-world applications with stringent latency constraints. While previous work on SLM design has primarily focused on reducing the number of parameters to achieve parameter-optimal SLMs, parameter efficiency does not necessarily translate into proportional real-device speed-ups. This work aims to identify the key determinants of SLMs' real-device latency and offer generalizable principles and methodologies for SLM design and training when real-device latency is the primary consideration. Specifically, we identify two central architectural factors: depth-width ratios and operator choices. The former is crucial for small-batch-size latency, while the latter affects both latency and large-batch-size throughput. In light of this, we first study latency-optimal depth-width ratios, with the key finding that although deep-thin models generally achieve better accuracy under the same parameter budget, they may not lie on the accuracy-latency trade-off frontier. Next, we explore emerging efficient attention alternatives to evaluate their potential as candidate building operators. Using the identified promising operators, we construct an evolutionary search framework to automatically discover latency-optimal combinations of these operators within hybrid SLMs, thereby advancing the accuracy-latency frontier. In addition to architectural improvements, we further enhance SLM training using a weight normalization technique that enables more effective weight updates and improves final convergence. Combining these methods, we introduce a new family of hybrid SLMs, called Nemotron-Flash, which significantly advances the accuracy-efficiency frontier of state-of-the-art SLMs, e.g., achieving over +5.5% average accuracy, 1.3x/1.9x lower latency, and 18.7x/45.6x higher throughput compared to Qwen3-1.7B/0.6B, respectively.

Adjovi Sim, Zhengkui Wang, Aik Beng Ng et al.

Online continual learning for image classification is crucial for models to adapt to new data while retaining knowledge of previously learned tasks. This capability is essential to address real-world challenges involving dynamic environments and evolving data distributions. Traditional approaches predominantly employ Convolutional Neural Networks, which are limited to processing images as grids and primarily capture local patterns rather than relational information. Although the emergence of transformer architectures has improved the ability to capture relationships, these models often require significantly larger resources. In this paper, we present a novel online continual learning framework based on Graph Attention Networks (GATs), which effectively capture contextual relationships and dynamically update the task-specific representation via learned attention weights. Our approach utilizes a pre-trained feature extractor to convert images into graphs using hierarchical feature maps, representing information at varying levels of granularity. These graphs are then processed by a GAT and incorporate an enhanced global pooling strategy to improve classification performance for continual learning. In addition, we propose the rehearsal memory duplication technique that improves the representation of the previous tasks while maintaining the memory budget. Comprehensive evaluations on benchmark datasets, including SVHN, CIFAR10, CIFAR100, and MiniImageNet, demonstrate the superiority of our method compared to the state-of-the-art methods.

Roger Waleffe, Wonmin Byeon, Duncan Riach et al.

Selective state-space models (SSMs) like Mamba overcome some of the shortcomings of Transformers, such as quadratic computational complexity with sequence length and large inference-time memory requirements from the key-value cache. Moreover, recent studies have shown that SSMs can match or exceed the language modeling capabilities of Transformers, making them an attractive alternative. In a controlled setting (e.g., same data), however, studies so far have only presented small scale experiments comparing SSMs to Transformers. To understand the strengths and weaknesses of these architectures at larger scales, we present a direct comparison between 8B-parameter Mamba, Mamba-2, and Transformer models trained on the same datasets of up to 3.5T tokens. We also compare these models to a hybrid architecture consisting of 43% Mamba-2, 7% attention, and 50% MLP layers (Mamba-2-Hybrid). Using a diverse set of tasks, we answer the question of whether Mamba models can match Transformers at larger training budgets. Our results show that while pure SSMs match or exceed Transformers on many tasks, they lag behind Transformers on tasks which require strong copying or in-context learning abilities (e.g., 5-shot MMLU, Phonebook) or long-context reasoning. In contrast, we find that the 8B Mamba-2-Hybrid exceeds the 8B Transformer on all 12 standard tasks we evaluated (+2.65 points on average) and is predicted to be up to 8x faster when generating tokens at inference time. To validate long-context capabilities, we provide additional experiments evaluating variants of the Mamba-2-Hybrid and Transformer extended to support 16K, 32K, and 128K sequences. On an additional 23 long-context tasks, the hybrid model continues to closely match or exceed the Transformer on average. To enable further study, we release the checkpoints as well as the code used to train our models as part of NVIDIA's Megatron-LM project.

Benjamin Wu, Oliver Hennigh, Jan Kautz et al.

Physics-informed neural networks allow models to be trained by physical laws described by general nonlinear partial differential equations. However, traditional architectures struggle to solve more challenging time-dependent problems due to their architectural nature. In this work, we present a novel physics-informed framework for solving time-dependent partial differential equations. Using only the governing differential equations and problem initial and boundary conditions, we generate a latent representation of the problem's spatio-temporal dynamics. Our model utilizes discrete cosine transforms to encode spatial frequencies and recurrent neural networks to process the time evolution. This efficiently and flexibly produces a compressed representation which is used for additional conditioning of physics-informed models. We show experimental results on the Taylor-Green vortex solution to the Navier-Stokes equations. Our proposed model achieves state-of-the-art performance on the Taylor-Green vortex relative to other physics-informed baseline models.

Seung Hyun Lee, Wonseok Roh, Wonmin Byeon et al.

The recent success of the generative model shows that leveraging the multi-modal embedding space can manipulate an image using text information. However, manipulating an image with other sources rather than text, such as sound, is not easy due to the dynamic characteristics of the sources. Especially, sound can convey vivid emotions and dynamic expressions of the real world. Here, we propose a framework that directly encodes sound into the multi-modal (image-text) embedding space and manipulates an image from the space. Our audio encoder is trained to produce a latent representation from an audio input, which is forced to be aligned with image and text representations in the multi-modal embedding space. We use a direct latent optimization method based on aligned embeddings for sound-guided image manipulation. We also show that our method can mix text and audio modalities, which enrich the variety of the image modification. We verify the effectiveness of our sound-guided image manipulation quantitatively and qualitatively. We also show that our method can mix different modalities, i.e., text and audio, which enrich the variety of the image modification. The experiments on zero-shot audio classification and semantic-level image classification show that our proposed model outperforms other text and sound-guided state-of-the-art methods.

Jiahao Su, Wonmin Byeon, Furong Huang

Enforcing orthogonality in neural networks is an antidote for gradient vanishing/exploding problems, sensitivity by adversarial perturbation, and bounding generalization errors. However, many previous approaches are heuristic, and the orthogonality of convolutional layers is not systematically studied: some of these designs are not exactly orthogonal, while others only consider standard convolutional layers and propose specific classes of their realizations. To address this problem, we propose a theoretical framework for orthogonal convolutional layers, which establishes the equivalence between various orthogonal convolutional layers in the spatial domain and the paraunitary systems in the spectral domain. Since there exists a complete spectral factorization of paraunitary systems, any orthogonal convolution layer can be parameterized as convolutions of spatial filters. Our framework endows high expressive power to various convolutional layers while maintaining their exact orthogonality. Furthermore, our layers are memory and computationally efficient for deep networks compared to previous designs. Our versatile framework, for the first time, enables the study of architecture designs for deep orthogonal networks, such as choices of skip connection, initialization, stride, and dilation. Consequently, we scale up orthogonal networks to deep architectures, including ResNet, WideResNet, and ShuffleNet, substantially increasing the performance over the traditional shallow orthogonal networks.

Yiran Zhong, Charles Loop, Wonmin Byeon et al.

Although deep learning-based methods have dominated stereo matching leaderboards by yielding unprecedented disparity accuracy, their inference time is typically slow, on the order of seconds for a pair of 540p images. The main reason is that the leading methods employ time-consuming 3D convolutions applied to a 4D feature volume. A common way to speed up the computation is to downsample the feature volume, but this loses high-frequency details. To overcome these challenges, we propose a \emph{displacement-invariant cost computation module} to compute the matching costs without needing a 4D feature volume. Rather, costs are computed by applying the same 2D convolution network on each disparity-shifted feature map pair independently. Unlike previous 2D convolution-based methods that simply perform context mapping between inputs and disparity maps, our proposed approach learns to match features between the two images. We also propose an entropy-based refinement strategy to refine the computed disparity map, which further improves speed by avoiding the need to compute a second disparity map on the right image. Extensive experiments on standard datasets (SceneFlow, KITTI, ETH3D, and Middlebury) demonstrate that our method achieves competitive accuracy with much less inference time. On typical image sizes, our method processes over 100 FPS on a desktop GPU, making our method suitable for time-critical applications such as autonomous driving. We also show that our approach generalizes well to unseen datasets, outperforming 4D-volumetric methods.

Jiahao Su, Wonmin Byeon, Jean Kossaifi et al.

Learning from spatio-temporal data has numerous applications such as human-behavior analysis, object tracking, video compression, and physics simulation.However, existing methods still perform poorly on challenging video tasks such as long-term forecasting. This is because these kinds of challenging tasks require learning long-term spatio-temporal correlations in the video sequence. In this paper, we propose a higher-order convolutional LSTM model that can efficiently learn these correlations, along with a succinct representations of the history. This is accomplished through a novel tensor train module that performs prediction by combining convolutional features across time. To make this feasible in terms of computation and memory requirements, we propose a novel convolutional tensor-train decomposition of the higher-order model. This decomposition reduces the model complexity by jointly approximating a sequence of convolutional kernels asa low-rank tensor-train factorization. As a result, our model outperforms existing approaches, but uses only a fraction of parameters, including the baseline models.Our results achieve state-of-the-art performance in a wide range of applications and datasets, including the multi-steps video prediction on the Moving-MNIST-2and KTH action datasets as well as early activity recognition on the Something-Something V2 dataset.

Pantelis R. Vlachas, Wonmin Byeon, Zhong Y. Wan et al.

We introduce a data-driven forecasting method for high-dimensional chaotic systems using long short-term memory (LSTM) recurrent neural networks. The proposed LSTM neural networks perform inference of high-dimensional dynamical systems in their reduced order space and are shown to be an effective set of nonlinear approximators of their attractor. We demonstrate the forecasting performance of the LSTM and compare it with Gaussian processes (GPs) in time series obtained from the Lorenz 96 system, the Kuramoto-Sivashinsky equation and a prototype climate model. The LSTM networks outperform the GPs in short-term forecasting accuracy in all applications considered. A hybrid architecture, extending the LSTM with a mean stochastic model (MSM-LSTM), is proposed to ensure convergence to the invariant measure. This novel hybrid method is fully data-driven and extends the forecasting capabilities of LSTM networks.

Wonmin Byeon, Qin Wang, Rupesh Kumar Srivastava et al.

Video prediction models based on convolutional networks, recurrent networks, and their combinations often result in blurry predictions. We identify an important contributing factor for imprecise predictions that has not been studied adequately in the literature: blind spots, i.e., lack of access to all relevant past information for accurately predicting the future. To address this issue, we introduce a fully context-aware architecture that captures the entire available past context for each pixel using Parallel Multi-Dimensional LSTM units and aggregates it using blending units. Our model outperforms a strong baseline network of 20 recurrent convolutional layers and yields state-of-the-art performance for next step prediction on three challenging real-world video datasets: Human 3.6M, Caltech Pedestrian, and UCF-101. Moreover, it does so with fewer parameters than several recently proposed models, and does not rely on deep convolutional networks, multi-scale architectures, separation of background and foreground modeling, motion flow learning, or adversarial training. These results highlight that full awareness of past context is of crucial importance for video prediction.

Marijn F. Stollenga, Wonmin Byeon, Marcus Liwicki et al.

Convolutional Neural Networks (CNNs) can be shifted across 2D images or 3D videos to segment them. They have a fixed input size and typically perceive only small local contexts of the pixels to be classified as foreground or background. In contrast, Multi-Dimensional Recurrent NNs (MD-RNNs) can perceive the entire spatio-temporal context of each pixel in a few sweeps through all pixels, especially when the RNN is a Long Short-Term Memory (LSTM). Despite these theoretical advantages, however, unlike CNNs, previous MD-LSTM variants were hard to parallelize on GPUs. Here we re-arrange the traditional cuboid order of computations in MD-LSTM in pyramidal fashion. The resulting PyraMiD-LSTM is easy to parallelize, especially for 3D data such as stacks of brain slice images. PyraMiD-LSTM achieved best known pixel-wise brain image segmentation results on MRBrainS13 (and competitive results on EM-ISBI12).