Bohyung Han

Minsoo Kang, Jaeyoo Park, Bohyung Han

We present a novel class incremental learning approach based on deep neural networks, which continually learns new tasks with limited memory for storing examples in the previous tasks. Our algorithm is based on knowledge distillation and provides a principled way to maintain the representations of old models while adjusting to new tasks effectively. The proposed method estimates the relationship between the representation changes and the resulting loss increases incurred by model updates. It minimizes the upper bound of the loss increases using the representations, which exploits the estimated importance of each feature map within a backbone model. Based on the importance, the model restricts updates of important features for robustness while allowing changes in less critical features for flexibility. This optimization strategy effectively alleviates the notorious catastrophic forgetting problem despite the limited accessibility of data in the previous tasks. The experimental results show significant accuracy improvement of the proposed algorithm over the existing methods on the standard datasets. Code is available.

Jinkyu Kim, Geeho Kim, Bohyung Han

A critical challenge of federated learning is data heterogeneity and imbalance across clients, which leads to inconsistency between local networks and unstable convergence of global models. To alleviate the limitations, we propose a novel architectural regularization technique that constructs multiple auxiliary branches in each local model by grafting local and global subnetworks at several different levels and that learns the representations of the main pathway in the local model congruent to the auxiliary hybrid pathways via online knowledge distillation. The proposed technique is effective to robustify the global model even in the non-iid setting and is applicable to various federated learning frameworks conveniently without incurring extra communication costs. We perform comprehensive empirical studies and demonstrate remarkable performance gains in terms of accuracy and efficiency compared to existing methods. The source code is available at our project page.

Zunzhi You, Daochang Liu, Bohyung Han et al.

Recent advancements in masked image modeling (MIM) have made it a prevailing framework for self-supervised visual representation learning. The MIM pretrained models, like most deep neural network methods, remain vulnerable to adversarial attacks, limiting their practical application, and this issue has received little research attention. In this paper, we investigate how this powerful self-supervised learning paradigm can provide adversarial robustness to downstream classifiers. During the exploration, we find that noisy image modeling (NIM), a simple variant of MIM that adopts denoising as the pre-text task, reconstructs noisy images surprisingly well despite severe corruption. Motivated by this observation, we propose an adversarial defense method, referred to as De^3, by exploiting the pretrained decoder for denoising. Through De^3, NIM is able to enhance adversarial robustness beyond providing pretrained features. Furthermore, we incorporate a simple modification, sampling the noise scale hyperparameter from random distributions, and enable the defense to achieve a better and tunable trade-off between accuracy and robustness. Experimental results demonstrate that, in terms of adversarial robustness, NIM is superior to MIM thanks to its effective denoising capability. Moreover, the defense provided by NIM achieves performance on par with adversarial training while offering the extra tunability advantage. Source code and models are available at https://github.com/youzunzhi/NIM-AdvDef.

Taehoon Kim, Pyunghwan Ahn, Sangyun Kim et al. · nvidia, utoronto

In this report, we introduce NICE (New frontiers for zero-shot Image Captioning Evaluation) project and share the results and outcomes of 2023 challenge. This project is designed to challenge the computer vision community to develop robust image captioning models that advance the state-of-the-art both in terms of accuracy and fairness. Through the challenge, the image captioning models were tested using a new evaluation dataset that includes a large variety of visual concepts from many domains. There was no specific training data provided for the challenge, and therefore the challenge entries were required to adapt to new types of image descriptions that had not been seen during training. This report includes information on the newly proposed NICE dataset, evaluation methods, challenge results, and technical details of top-ranking entries. We expect that the outcomes of the challenge will contribute to the improvement of AI models on various vision-language tasks.

George Ciubotariu, Sharif S M A, Abdur Rehman et al.

This paper presents a comprehensive review of the NTIRE 2026 Low Light Image Enhancement Challenge, highlighting the proposed solutions and final results. The objective of this challenge is to identify effective networks capable of producing clearer and visually compelling images in diverse and challenging conditions by learning representative visual cues with the purpose of restoring information loss due to low-contrast and noisy images. A total of 195 participants registered for the first track and 153 for the second track of the competition, and 22 teams ultimately submitted valid entries. This paper thoroughly evaluates the state-of-the-art advances in (joint denoising and) low-light image enhancement, showcasing the significant progress in the field, while leveraging samples of our novel dataset.

Dongwan Kim, Bohyung Han

A primary goal of class-incremental learning is to strike a balance between stability and plasticity, where models should be both stable enough to retain knowledge learned from previously seen classes, and plastic enough to learn concepts from new classes. While previous works demonstrate strong performance on class-incremental benchmarks, it is not clear whether their success comes from the models being stable, plastic, or a mixture of both. This paper aims to shed light on how effectively recent class-incremental learning algorithms address the stability-plasticity trade-off. We establish analytical tools that measure the stability and plasticity of feature representations, and employ such tools to investigate models trained with various algorithms on large-scale class-incremental benchmarks. Surprisingly, we find that the majority of class-incremental learning algorithms heavily favor stability over plasticity, to the extent that the feature extractor of a model trained on the initial set of classes is no less effective than that of the final incremental model. Our observations not only inspire two simple algorithms that highlight the importance of feature representation analysis, but also suggest that class-incremental learning approaches, in general, should strive for better feature representation learning.

Jaeyoo Park, Minsoo Kang, Bohyung Han

We tackle catastrophic forgetting problem in the context of class-incremental learning for video recognition, which has not been explored actively despite the popularity of continual learning. Our framework addresses this challenging task by introducing time-channel importance maps and exploiting the importance maps for learning the representations of incoming examples via knowledge distillation. We also incorporate a regularization scheme in our objective function, which encourages individual features obtained from different time steps in a video to be uncorrelated and eventually improves accuracy by alleviating catastrophic forgetting. We evaluate the proposed approach on brand-new splits of class-incremental action recognition benchmarks constructed upon the UCF101, HMDB51, and Something-Something V2 datasets, and demonstrate the effectiveness of our algorithm in comparison to the existing continual learning methods that are originally designed for image data.

Minji Kim, Seungkwan Lee, Jungseul Ok et al.

Despite the extensive adoption of machine learning on the task of visual object tracking, recent learning-based approaches have largely overlooked the fact that visual tracking is a sequence-level task in its nature; they rely heavily on frame-level training, which inevitably induces inconsistency between training and testing in terms of both data distributions and task objectives. This work introduces a sequence-level training strategy for visual tracking based on reinforcement learning and discusses how a sequence-level design of data sampling, learning objectives, and data augmentation can improve the accuracy and robustness of tracking algorithms. Our experiments on standard benchmarks including LaSOT, TrackingNet, and GOT-10k demonstrate that four representative tracking models, SiamRPN++, SiamAttn, TransT, and TrDiMP, consistently improve by incorporating the proposed methods in training without modifying architectures.

Dong-Hwan Jang, Sanghyeok Chu, Joonhyuk Kim et al.

The size and shape of the receptive field determine how the network aggregates local information and affect the overall performance of a model considerably. Many components in a neural network, such as kernel sizes and strides for convolution and pooling operations, influence the configuration of a receptive field. However, they still rely on hyperparameters, and the receptive fields of existing models result in suboptimal shapes and sizes. Hence, we propose a simple yet effective Dynamically Optimized Pooling operation, referred to as DynOPool, which optimizes the scale factors of feature maps end-to-end by learning the desirable size and shape of its receptive field in each layer. Any kind of resizing modules in a deep neural network can be replaced by the operations with DynOPool at a minimal cost. Also, DynOPool controls the complexity of a model by introducing an additional loss term that constrains computational cost. Our experiments show that the models equipped with the proposed learnable resizing module outperform the baseline networks on multiple datasets in image classification and semantic segmentation.

Taehoon Kim, Jaeyoo Park, Bohyung Han

We propose a novel class incremental learning approach by incorporating a feature augmentation technique motivated by adversarial attacks. We employ a classifier learned in the past to complement training examples rather than simply play a role as a teacher for knowledge distillation towards subsequent models. The proposed approach has a unique perspective to utilize the previous knowledge in class incremental learning since it augments features of arbitrary target classes using examples in other classes via adversarial attacks on a previously learned classifier. By allowing the cross-class feature augmentations, each class in the old tasks conveniently populates samples in the feature space, which alleviates the collapse of the decision boundaries caused by sample deficiency for the previous tasks, especially when the number of stored exemplars is small. This idea can be easily incorporated into existing class incremental learning algorithms without any architecture modification. Extensive experiments on the standard benchmarks show that our method consistently outperforms existing class incremental learning methods by significant margins in various scenarios, especially under an environment with an extremely limited memory budget.

Jaeyoo Park, Bohyung Han

We propose a visual-linguistic representation learning approach within a self-supervised learning framework by introducing a new operation, loss, and data augmentation strategy. First, we generate diverse features for the image-text matching (ITM) task via soft-masking the regions in an image, which are most relevant to a certain word in the corresponding caption, instead of completely removing them. Since our framework relies only on image-caption pairs with no fine-grained annotations, we identify the relevant regions to each word by computing the word-conditional visual attention using multi-modal encoder. Second, we encourage the model to focus more on hard but diverse examples by proposing a focal loss for the image-text contrastive learning (ITC) objective, which alleviates the inherent limitations of overfitting and bias issues. Last, we perform multi-modal data augmentations for self-supervised learning via mining various examples by masking texts and rendering distortions on images. We show that the combination of these three innovations is effective for learning a pretrained model, leading to outstanding performance on multiple vision-language downstream tasks.

Jaeyoo Park, Junha Kim, Bohyung Han

We present a meta-learning framework for weakly supervised anomaly detection in videos, where the detector learns to adapt to unseen types of abnormal activities effectively when only video-level annotations of binary labels are available. Our work is motivated by the fact that existing methods suffer from poor generalization to diverse unseen examples. We claim that an anomaly detector equipped with a meta-learning scheme alleviates the limitation by leading the model to an initialization point for better optimization. We evaluate the performance of our framework on two challenging datasets, UCF-Crime and ShanghaiTech. The experimental results demonstrate that our algorithm boosts the capability to localize unseen abnormal events in a weakly supervised setting. Besides the technical contributions, we perform the annotation of missing labels in the UCF-Crime dataset and make our task evaluated effectively.

Seonguk Seo, Mustafa Gokhan Uzunbas, Bohyung Han et al.

Backfilling is the process of re-extracting all gallery embeddings from upgraded models in image retrieval systems. It inevitably requires a prohibitively large amount of computational cost and even entails the downtime of the service. Although backward-compatible learning sidesteps this challenge by tackling query-side representations, this leads to suboptimal solutions in principle because gallery embeddings cannot benefit from model upgrades. We address this dilemma by introducing an online backfilling algorithm, which enables us to achieve a progressive performance improvement during the backfilling process while not sacrificing the final performance of new model after the completion of backfilling. To this end, we first propose a simple distance rank merge technique for online backfilling. Then, we incorporate a reverse transformation module for more effective and efficient merging, which is further enhanced by adopting a metric-compatible contrastive learning approach. These two components help to make the distances of old and new models compatible, resulting in desirable merge results during backfilling with no extra computational overhead. Extensive experiments show the effectiveness of our framework on four standard benchmarks in various settings.

Taehoon Kim, Bohyung Han

We propose a novel domain generalization technique, referred to as Randomized Adversarial Style Perturbation (RASP), which is motivated by the observation that the characteristics of each domain are captured by the feature statistics corresponding to style. The proposed algorithm perturbs the style of a feature in an adversarial direction towards a randomly selected class, and makes the model learn against being misled by the unexpected styles observed in unseen target domains. While RASP is effective to handle domain shifts, its naive integration into the training procedure might degrade the capability of learning knowledge from source domains because it has no restriction on the perturbations of representations. This challenge is alleviated by Normalized Feature Mixup (NFM), which facilitates the learning of the original features while achieving robustness to perturbed representations via their mixup during training. We evaluate the proposed algorithm via extensive experiments on various benchmarks and show that our approach improves domain generalization performance, especially in large-scale benchmarks.

Fu-Yun Wang, Hao Zhou, Liangzhe Yuan et al.

The slow inference process of image diffusion models significantly degrades interactive user experiences. To address this, we introduce Diffusion Preview, a novel paradigm employing rapid, low-step sampling to generate preliminary outputs for user evaluation, deferring full-step refinement until the preview is deemed satisfactory. Existing acceleration methods, including training-free solvers and post-training distillation, struggle to deliver high-quality previews or ensure consistency between previews and final outputs. We propose ConsistencySolver derived from general linear multistep methods, a lightweight, trainable high-order solver optimized via Reinforcement Learning, that enhances preview quality and consistency. Experimental results demonstrate that ConsistencySolver significantly improves generation quality and consistency in low-step scenarios, making it ideal for efficient preview-and-refine workflows. Notably, it achieves FID scores on-par with Multistep DPM-Solver using 47% fewer steps, while outperforming distillation baselines. Furthermore, user studies indicate our approach reduces overall user interaction time by nearly 50% while maintaining generation quality. Code is available at https://github.com/G-U-N/consolver.

Minsoo Kang, Doyup Lee, Jiseob Kim et al.

We propose a text-to-image generation algorithm based on deep neural networks when text captions for images are unavailable during training. In this work, instead of simply generating pseudo-ground-truth sentences of training images using existing image captioning methods, we employ a pretrained CLIP model, which is capable of properly aligning embeddings of images and corresponding texts in a joint space and, consequently, works well on zero-shot recognition tasks. We optimize a text-to-image generation model by maximizing the data log-likelihood conditioned on pairs of image-text CLIP embeddings. To better align data in the two domains, we employ a principled way based on a variational inference, which efficiently estimates an approximate posterior of the hidden text embedding given an image and its CLIP feature. Experimental results validate that the proposed framework outperforms existing approaches by large margins under unsupervised and semi-supervised text-to-image generation settings.

Minsoo Kang, Hyewon Yoo, Eunhee Kang et al.

We propose an information-theoretic knowledge distillation approach for the compression of generative adversarial networks, which aims to maximize the mutual information between teacher and student networks via a variational optimization based on an energy-based model. Because the direct computation of the mutual information in continuous domains is intractable, our approach alternatively optimizes the student network by maximizing the variational lower bound of the mutual information. To achieve a tight lower bound, we introduce an energy-based model relying on a deep neural network to represent a flexible variational distribution that deals with high-dimensional images and consider spatial dependencies between pixels, effectively. Since the proposed method is a generic optimization algorithm, it can be conveniently incorporated into arbitrary generative adversarial networks and even dense prediction networks, e.g., image enhancement models. We demonstrate that the proposed algorithm achieves outstanding performance in model compression of generative adversarial networks consistently when combined with several existing models.

Ilchae Jung, Minji Kim, Eunhyeok Park et al.

This paper presents a novel hybrid representation learning framework for streaming data, where an image frame in a video is modeled by an ensemble of two distinct deep neural networks; one is a low-bit quantized network and the other is a lightweight full-precision network. The former learns coarse primary information with low cost while the latter conveys residual information for high fidelity to original representations. The proposed parallel architecture is effective to maintain complementary information since fixed-point arithmetic can be utilized in the quantized network and the lightweight model provides precise representations given by a compact channel-pruned network. We incorporate the hybrid representation technique into an online visual tracking task, where deep neural networks need to handle temporal variations of target appearances in real-time. Compared to the state-of-the-art real-time trackers based on conventional deep neural networks, our tracking algorithm demonstrates competitive accuracy on the standard benchmarks with a small fraction of computational cost and memory footprint.

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

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

Junoh Kang, Jinyoung Choi, Sungik Choi et al.

We propose a novel diffusion-based image generation method called the observation-guided diffusion probabilistic model (OGDM), which effectively addresses the tradeoff between quality control and fast sampling. Our approach reestablishes the training objective by integrating the guidance of the observation process with the Markov chain in a principled way. This is achieved by introducing an additional loss term derived from the observation based on a conditional discriminator on noise level, which employs a Bernoulli distribution indicating whether its input lies on the (noisy) real manifold or not. This strategy allows us to optimize the more accurate negative log-likelihood induced in the inference stage especially when the number of function evaluations is limited. The proposed training scheme is also advantageous even when incorporated only into the fine-tuning process, and it is compatible with various fast inference strategies since our method yields better denoising networks using the exactly the same inference procedure without incurring extra computational cost. We demonstrate the effectiveness of our training algorithm using diverse inference techniques on strong diffusion model baselines. Our implementation is available at https://github.com/Junoh-Kang/OGDM_edm.

Mijeong Kim, Hyunjoon Lee, Bohyung Han

Although 3D-aware GANs based on neural radiance fields have achieved competitive performance, their applicability is still limited to objects or scenes with the ground-truths or prediction models for clearly defined canonical camera poses. To extend the scope of applicable datasets, we propose a novel 3D-aware GAN optimization technique through contrastive learning with implicit pose embeddings. To this end, we first revise the discriminator design and remove dependency on ground-truth camera poses. Then, to capture complex and challenging 3D scene structures more effectively, we make the discriminator estimate a high-dimensional implicit pose embedding from a given image and perform contrastive learning on the pose embedding. The proposed approach can be employed for the dataset, where the canonical camera pose is ill-defined because it does not look up or estimate camera poses. Experimental results show that our algorithm outperforms existing methods by large margins on the datasets with multiple object categories and inconsistent canonical camera poses.

Jisoo Kim, Jungbin Cho, Sanghyeok Chu et al.

Humans learn not only how their bodies move, but also how the surrounding world responds to their actions. In contrast, while recent Vision-Language-Action (VLA) models exhibit impressive semantic understanding, they often fail to capture the spatiotemporal dynamics governing physical interaction. In this paper, we introduce Pri4R, a simple yet effective approach that endows VLA models with an implicit understanding of world dynamics by leveraging privileged 4D information during training. Specifically, Pri4R augments VLAs with a lightweight point track head that predicts 3D point tracks. By injecting VLA features into this head to jointly predict future 3D trajectories, the model learns to incorporate evolving scene geometry within its shared representation space, enabling more physically aware context for precise control. Due to its architectural simplicity, Pri4R is compatible with dominant VLA design patterns with minimal changes. During inference, we run the model using the original VLA architecture unchanged; Pri4R adds no extra inputs, outputs, or computational overhead. Across simulation and real-world evaluations, Pri4R significantly improves performance on challenging manipulation tasks, including a +10% gain on LIBERO-Long and a +40% gain on RoboCasa. We further show that 3D point track prediction is an effective supervision target for learning action-world dynamics, and validate our design choices through extensive ablations.

Junsung Lee, Minsoo Kang, Bohyung Han

We propose a simple but effective training-free approach tailored to diffusion-based image-to-image translation. Our approach revises the original noise prediction network of a pretrained diffusion model by introducing a noise correction term. We formulate the noise correction term as the difference between two noise predictions; one is computed from the denoising network with a progressive interpolation of the source and target prompt embeddings, while the other is the noise prediction with the source prompt embedding. The final noise prediction network is given by a linear combination of the standard denoising term and the noise correction term, where the former is designed to reconstruct must-be-preserved regions while the latter aims to effectively edit regions of interest relevant to the target prompt. Our approach can be easily incorporated into existing image-to-image translation methods based on diffusion models. Extensive experiments verify that the proposed technique achieves outstanding performance with low latency and consistently improves existing frameworks when combined with them.

Donghun Ryou, Inju Ha, Sanghyeok Chu et al.

Deep learning-based image restoration has achieved significant success. However, when addressing real-world degradations, model performance is limited by the quality of ground-truth images in datasets due to practical constraints in data acquisition. To address this limitation, we propose a novel framework that enhances existing ground truth images to provide higher-quality supervision for real-world restoration. Our framework generates perceptually enhanced ground truth images using super-resolution by incorporating adaptive frequency masks, which are learned by a conditional frequency mask generator. These masks guide the optimal fusion of frequency components from the original ground truth and its super-resolved variants, yielding enhanced ground truth images. This frequency-domain mixup preserves the semantic consistency of the original content while selectively enriching perceptual details, preventing hallucinated artifacts that could compromise fidelity. The enhanced ground truth images are used to train a lightweight output refinement network that can be seamlessly integrated with existing restoration models. Extensive experiments demonstrate that our approach consistently improves the quality of restored images. We further validate the effectiveness of both supervision enhancement and output refinement through user studies. Code is available at https://github.com/dhryougit/Beyond-the-Ground-Truth.

Tackgeun You, Mijeong Kim, Jungtaek Kim et al.

We propose a novel approach to learning the generative neural fields represented by linear combinations of implicit basis networks. Our algorithm learns basis networks in the form of implicit neural representations and their coefficients in a latent space by either conducting meta-learning or adopting auto-decoding paradigms. The proposed method easily enlarges the capacity of generative neural fields by increasing the number of basis networks while maintaining the size of a network for inference to be small through their weighted model averaging. Consequently, sampling instances using the model is efficient in terms of latency and memory footprint. Moreover, we customize denoising diffusion probabilistic model for a target task to sample latent mixture coefficients, which allows our final model to generate unseen data effectively. Experiments show that our approach achieves competitive generation performance on diverse benchmarks for images, voxel data, and NeRF scenes without sophisticated designs for specific modalities and domains.

Seonguk Seo, Dongwan Kim, Bohyung Han

Machine unlearning, an emerging research topic focusing on compliance with data privacy regulations, enables trained models to remove the information learned from specific data. While many existing methods indirectly address this issue by intentionally injecting incorrect supervisions, they can drastically and unpredictably alter the decision boundaries and feature spaces, leading to training instability and undesired side effects. To fundamentally approach this task, we first analyze the changes in latent feature spaces between original and retrained models, and observe that the feature representations of samples not involved in training are closely aligned with the feature manifolds of previously seen samples in training. Based on these findings, we introduce a novel evaluation metric for machine unlearning, coined dimensional alignment, which measures the alignment between the eigenspaces of the forget and retain set samples. We employ this metric as a regularizer loss to build a robust and stable unlearning framework, which is further enhanced by integrating a self-distillation loss and an alternating training scheme. Our framework effectively eliminates information from the forget set and preserves knowledge from the retain set. Lastly, we identify critical flaws in established evaluation metrics for machine unlearning, and introduce new evaluation tools that more accurately reflect the fundamental goals of machine unlearning.

Minji Kim, Dongyoon Han, Taekyung Kim et al.

We propose a novel framework for video understanding, called Temporally Contextualized CLIP (TC-CLIP), which leverages essential temporal information through global interactions in a spatio-temporal domain within a video. To be specific, we introduce Temporal Contextualization (TC), a layer-wise temporal information infusion mechanism for videos, which 1) extracts core information from each frame, 2) connects relevant information across frames for the summarization into context tokens, and 3) leverages the context tokens for feature encoding. Furthermore, the Video-conditional Prompting (VP) module processes context tokens to generate informative prompts in the text modality. Extensive experiments in zero-shot, few-shot, base-to-novel, and fully-supervised action recognition validate the effectiveness of our model. Ablation studies for TC and VP support our design choices. Our project page with the source code is available at https://github.com/naver-ai/tc-clip

Jihwan Kim, Nikhil Parthasarathy, Danfeng Qin et al.

The fundamental challenge in scaling Video Large Language Models (Video LLMs) to long-form video lies in managing the explosion of visual-token context length. Existing strategies predominantly focus on "post-hoc" token reduction -- reducing visual tokens after feature extraction to alleviate the LLM's computational overhead. While these methods effectively reduce the number of visual tokens, we observe that the primary latency bottleneck then shifts from the LLM to the expensive per-frame processing of the vision encoder. To address this, we introduce LiteFrame, a strong, yet highly efficient video encoder backbone for Video LLMs. To train LiteFrame, we propose Compressed Token Distillation (CTD), a novel training framework that teaches a compact student vision encoder to directly predict information-dense, spatio-temporally compressed representations produced by a large teacher vision model, effectively bypassing redundant computation. When coupled with further Language Model Adaptation (LMA), this approach results in a new latency-accuracy Pareto frontier -- compared with InternVL3-8B, LiteFrame provides a 35% reduction in end-to-end latency while processing 8$\times$ more frames and improves average video understanding accuracy across multiple benchmarks. Our results demonstrate a new potential path to unlocking longer-form video understanding under fixed compute budgets.

Sanghyeok Chu, Pyunghwan Ahn, Gwangmo Song et al.

Sparse Upcycling provides an efficient way to initialize a Mixture-of-Experts (MoE) model from pretrained dense weights instead of training from scratch. However, since all experts start from identical weights and the router is randomly initialized, the model suffers from expert symmetry and limited early specialization. We propose Cluster-aware Upcycling, a strategy that incorporates semantic structure into MoE initialization. Our method first partitions the dense model's input activations into semantic clusters. Each expert is then initialized using the subspace representations of its corresponding cluster via truncated SVD, while setting the router's initial weights to the cluster centroids. This cluster-aware initialization breaks expert symmetry and encourages early specialization aligned with the data distribution. Furthermore, we introduce an expert-ensemble self-distillation loss that stabilizes training by providing reliable routing guidance using an ensemble teacher. When evaluated on CLIP ViT-B/32 and ViT-B/16, Cluster-aware Upcycling consistently outperforms existing methods across both zero-shot and few-shot benchmarks. The proposed method also produces more diverse and disentangled expert representations, reduces inter-expert similarity, and leads to more confident routing behavior. Project page: https://sanghyeokchu.github.io/cluster-aware-upcycling/

Kanghyun Baek, Sangyub Lee, Jin Young Choi et al.

Despite recent advances, diffusion-based text-to-image models still struggle with accurate text rendering. Several studies have proposed fine-tuning or training-free refinement methods for accurate text rendering. However, the critical issue of text omission, where the desired text is partially or entirely missing, remains largely overlooked. In this work, we propose TextGuider, a novel training-free method that encourages accurate and complete text appearance by aligning textual content tokens and text regions in the image. Specifically, we analyze attention patterns in MM-DiT models, particularly for text-related tokens intended to be rendered in the image. Leveraging this observation, we apply latent guidance during the early stage of denoising steps based on two loss functions that we introduce. Our method achieves state-of-the-art performance in test-time text rendering, with significant gains in recall and strong results in OCR accuracy and CLIP score.

Minji Kim, Taekyung Kim, Bohyung Han

Video Large Language Models (VideoLLMs) extend the capabilities of vision-language models to spatiotemporal inputs, enabling tasks such as video question answering (VideoQA). Despite recent advances in VideoLLMs, their internal mechanisms on where and how they extract and propagate video and textual information remain less explored. In this study, we investigate the internal information flow of VideoLLMs using mechanistic interpretability techniques. Our analysis reveals consistent patterns across diverse VideoQA tasks: (1) temporal reasoning in VideoLLMs initiates with active cross-frame interactions in early-to-middle layers, (2) followed by progressive video-language integration in middle layers. This is facilitated by alignment between video representations and linguistic embeddings containing temporal concepts. (3) Upon completion of this integration, the model is ready to generate correct answers in middle-to-late layers. (4) Based on our analysis, we show that VideoLLMs can retain their VideoQA performance by selecting these effective information pathways while suppressing a substantial amount of attention edges, e.g., 58% in LLaVA-NeXT-7B-Video-FT. These findings provide a blueprint on how VideoLLMs perform temporal reasoning and offer practical insights for improving model interpretability and downstream generalization. Our project page with the source code is available at https://map-the-flow.github.io

Geeho Kim, Jinkyu Kim, Bohyung Han

Federated learning often suffers from slow and unstable convergence due to the heterogeneous characteristics of participating client datasets. Such a tendency is aggravated when the client participation ratio is low since the information collected from the clients has large variations. To address this challenge, we propose a simple but effective federated learning framework, which improves the consistency across clients and facilitates the convergence of the server model. This is achieved by making the server broadcast a global model with a lookahead gradient. This strategy enables the proposed approach to convey the projected global update information to participants effectively without additional client memory and extra communication costs. We also regularize local updates by aligning each client with the overshot global model to reduce bias and improve the stability of our algorithm. We provide the theoretical convergence rate of our algorithm and demonstrate remarkable performance gains in terms of accuracy and communication efficiency compared to the state-of-the-art methods, especially with low client participation rates. The source code is available at our project page.

Myungseo Song, Jinyoung Choi, Bohyung Han

We propose a versatile deep image compression network based on Spatial Feature Transform (SFT arXiv:1804.02815), which takes a source image and a corresponding quality map as inputs and produce a compressed image with variable rates. Our model covers a wide range of compression rates using a single model, which is controlled by arbitrary pixel-wise quality maps. In addition, the proposed framework allows us to perform task-aware image compressions for various tasks, e.g., classification, by efficiently estimating optimized quality maps specific to target tasks for our encoding network. This is even possible with a pretrained network without learning separate models for individual tasks. Our algorithm achieves outstanding rate-distortion trade-off compared to the approaches based on multiple models that are optimized separately for several different target rates. At the same level of compression, the proposed approach successfully improves performance on image classification and text region quality preservation via task-aware quality map estimation without additional model training. The code is available at the project website: https://github.com/micmic123/QmapCompression

Jonghwan Mun, Minsu Cho, Bohyung Han

This paper addresses the problem of text-to-video temporal grounding, which aims to identify the time interval in a video semantically relevant to a text query. We tackle this problem using a novel regression-based model that learns to extract a collection of mid-level features for semantic phrases in a text query, which corresponds to important semantic entities described in the query (e.g., actors, objects, and actions), and reflect bi-modal interactions between the linguistic features of the query and the visual features of the video in multiple levels. The proposed method effectively predicts the target time interval by exploiting contextual information from local to global during bi-modal interactions. Through in-depth ablation studies, we find out that incorporating both local and global context in video and text interactions is crucial to the accurate grounding. Our experiment shows that the proposed method outperforms the state of the arts on Charades-STA and ActivityNet Captions datasets by large margins, 7.44\% and 4.61\% points at Recall@tIoU=0.5 metric, respectively. Code is available in https://github.com/JonghwanMun/LGI4temporalgrounding.

Hyeonwoo Noh, Andre Araujo, Jack Sim et al.

We propose an attentive local feature descriptor suitable for large-scale image retrieval, referred to as DELF (DEep Local Feature). The new feature is based on convolutional neural networks, which are trained only with image-level annotations on a landmark image dataset. To identify semantically useful local features for image retrieval, we also propose an attention mechanism for keypoint selection, which shares most network layers with the descriptor. This framework can be used for image retrieval as a drop-in replacement for other keypoint detectors and descriptors, enabling more accurate feature matching and geometric verification. Our system produces reliable confidence scores to reject false positives---in particular, it is robust against queries that have no correct match in the database. To evaluate the proposed descriptor, we introduce a new large-scale dataset, referred to as Google-Landmarks dataset, which involves challenges in both database and query such as background clutter, partial occlusion, multiple landmarks, objects in variable scales, etc. We show that DELF outperforms the state-of-the-art global and local descriptors in the large-scale setting by significant margins. Code and dataset can be found at the project webpage: https://github.com/tensorflow/models/tree/master/research/delf .

Yeonsang Shin, Insoo Kim, Bongkeun Kim et al.

Transformer-based autoregressive models excel in data generation but are inherently constrained by their reliance on discretized tokens, which limits their ability to represent continuous values with high precision. We analyze the scalability limitations of existing discretization-based approaches for generating hybrid discrete-continuous sequences, particularly in high-precision domains such as semiconductor circuit designs, where precision loss can lead to functional failure. To address the challenge, we propose AGDC, a novel unified framework that jointly models discrete and continuous values for variable-length sequences. AGDC employs a hybrid approach that combines categorical prediction for discrete values with diffusion-based modeling for continuous values, incorporating two key technical components: an end-of-sequence (EOS) logit adjustment mechanism that uses an MLP to dynamically adjust EOS token logits based on sequence context, and a length regularization term integrated into the loss function. Additionally, we present ContLayNet, a large-scale benchmark comprising 334K high-precision semiconductor layout samples with specialized evaluation metrics that capture functional correctness where precision errors significantly impact performance. Experiments on semiconductor layouts (ContLayNet), graphic layouts, and SVGs demonstrate AGDC's superior performance in generating high-fidelity hybrid vector representations compared to discretization-based and fixed-schema baselines, achieving scalable high-precision generation across diverse domains.

Jihwan Kim, Junoh Kang, Jinyoung Choi et al.

We propose a novel inference technique based on a pretrained diffusion model for text-conditional video generation. Our approach, called FIFO-Diffusion, is conceptually capable of generating infinitely long videos without additional training. This is achieved by iteratively performing diagonal denoising, which simultaneously processes a series of consecutive frames with increasing noise levels in a queue; our method dequeues a fully denoised frame at the head while enqueuing a new random noise frame at the tail. However, diagonal denoising is a double-edged sword as the frames near the tail can take advantage of cleaner frames by forward reference but such a strategy induces the discrepancy between training and inference. Hence, we introduce latent partitioning to reduce the training-inference gap and lookahead denoising to leverage the benefit of forward referencing. Practically, FIFO-Diffusion consumes a constant amount of memory regardless of the target video length given a baseline model, while well-suited for parallel inference on multiple GPUs. We have demonstrated the promising results and effectiveness of the proposed methods on existing text-to-video generation baselines. Generated video examples and source codes are available at our project page.

Taehoon Kim, Donghwan Jang, Bohyung Han

We present a novel training approach, named Merge-and-Bound (M&B) for Class Incremental Learning (CIL), which directly manipulates model weights in the parameter space for optimization. Our algorithm involves two types of weight merging: inter-task weight merging and intra-task weight merging. Inter-task weight merging unifies previous models by averaging the weights of models from all previous stages. On the other hand, intra-task weight merging facilitates the learning of current task by combining the model parameters within current stage. For reliable weight merging, we also propose a bounded update technique that aims to optimize the target model with minimal cumulative updates and preserve knowledge from previous tasks; this strategy reveals that it is possible to effectively obtain new models near old ones, reducing catastrophic forgetting. M&B is seamlessly integrated into existing CIL methods without modifying architecture components or revising learning objectives. We extensively evaluate our algorithm on standard CIL benchmarks and demonstrate superior performance compared to state-of-the-art methods.

Seonguk Seo, Jinkyu Kim, Geeho Kim et al.

We propose a novel contrastive learning framework to effectively address the challenges of data heterogeneity in federated learning. We first analyze the inconsistency of gradient updates across clients during local training and establish its dependence on the distribution of feature representations, leading to the derivation of the supervised contrastive learning (SCL) objective to mitigate local deviations. In addition, we show that a naïve adoption of SCL in federated learning leads to representation collapse, resulting in slow convergence and limited performance gains. To address this issue, we introduce a relaxed contrastive learning loss that imposes a divergence penalty on excessively similar sample pairs within each class. This strategy prevents collapsed representations and enhances feature transferability, facilitating collaborative training and leading to significant performance improvements. Our framework outperforms all existing federated learning approaches by huge margins on the standard benchmarks through extensive experimental results.

Lei Sun, Hang Guo, Bin Ren et al.

This paper presents an overview of the NTIRE 2025 Image Denoising Challenge (σ = 50), highlighting the proposed methodologies and corresponding results. The primary objective is to develop a network architecture capable of achieving high-quality denoising performance, quantitatively evaluated using PSNR, without constraints on computational complexity or model size. The task assumes independent additive white Gaussian noise (AWGN) with a fixed noise level of 50. A total of 290 participants registered for the challenge, with 20 teams successfully submitting valid results, providing insights into the current state-of-the-art in image denoising.

Mijeong Kim, Jongwoo Lim, Bohyung Han

Novel view synthesis of dynamic scenes is becoming important in various applications, including augmented and virtual reality. We propose a novel 4D Gaussian Splatting (4DGS) algorithm for dynamic scenes from casually recorded monocular videos. To overcome the overfitting problem of existing work for these real-world videos, we introduce an uncertainty-aware regularization that identifies uncertain regions with few observations and selectively imposes additional priors based on diffusion models and depth smoothness on such regions. This approach improves both the performance of novel view synthesis and the quality of training image reconstruction. We also identify the initialization problem of 4DGS in fast-moving dynamic regions, where the Structure from Motion (SfM) algorithm fails to provide reliable 3D landmarks. To initialize Gaussian primitives in such regions, we present a dynamic region densification method using the estimated depth maps and scene flow. Our experiments show that the proposed method improves the performance of 4DGS reconstruction from a video captured by a handheld monocular camera and also exhibits promising results in few-shot static scene reconstruction.

Jie Hu, Lingyun Chen, Geeho Kim et al.

History-dependent sampling can reduce long-run Monte Carlo variance by discouraging redundant revisits, but existing schemes typically encode history through empirical measure on finite state spaces, which is infeasible in high-dimensional discrete configuration spaces or ill-posed in continuous domains. We propose Score-Repellent Monte Carlo (SRMC) framework that summarizes trajectory history by a running average of score evaluations in $R^d$, where $d$ is the dimension of the score and state representation. This history is converted into a surrogate target through an exponential score tilt, indexed with $α$ that represents the strength of repellence in controlling the magnitude of the history-based repulsion. The surrogate family is normalization-free in the standard MCMC sense, yielding a generic wrapper: at each iteration, any base kernel targeting $π$ can instead be run on the current surrogate $π_{θ_n}$ while the history is updated online. We analyze the coupled evolution of the history recursion and Monte Carlo estimators using stochastic approximation with controlled Markovian noise, establishing almost sure convergence and a joint central limit theorem. We further identify regimes in which the asymptotic covariance decreases as $α$ increases, with scaling $O(1/α)$, extending the near-zero-variance effect of finite-state history-dependent samplers to general state spaces with constant memory. Experiments on continuous targets and discrete energy-based models demonstrate improved estimator variance and mode coverage, while retaining $O(d)$ memory usage and modest per-iteration overhead.

Jaeyoo Park, Jin Young Choi, Jeonghyung Park et al.

We present a novel OCR-free document understanding framework based on pretrained Multimodal Large Language Models (MLLMs). Our approach employs multi-scale visual features to effectively handle various font sizes within document images. To address the increasing costs of considering the multi-scale visual inputs for MLLMs, we propose the Hierarchical Visual Feature Aggregation (HVFA) module, designed to reduce the number of input tokens to LLMs. Leveraging a feature pyramid with cross-attentive pooling, our approach effectively manages the trade-off between information loss and efficiency without being affected by varying document image sizes. Furthermore, we introduce a novel instruction tuning task, which facilitates the model's text-reading capability by learning to predict the relative positions of input text, eventually minimizing the risk of truncated text caused by the limited capacity of LLMs. Comprehensive experiments validate the effectiveness of our approach, demonstrating superior performance in various document understanding tasks.

Myungseo Song, Jinyoung Choi, Bohyung Han

We study the robustness of learned image compression models against adversarial attacks and present a training-free defense technique based on simple image transform functions. Recent learned image compression models are vulnerable to adversarial attacks that result in poor compression rate, low reconstruction quality, or weird artifacts. To address the limitations, we propose a simple but effective two-way compression algorithm with random input transforms, which is conveniently applicable to existing image compression models. Unlike the naïve approaches, our approach preserves the original rate-distortion performance of the models on clean images. Moreover, the proposed algorithm requires no additional training or modification of existing models, making it more practical. We demonstrate the effectiveness of the proposed techniques through extensive experiments under multiple compression models, evaluation metrics, and attack scenarios.

Hyunsoo Lee, Minsoo Kang, Bohyung Han

We present a simple but effective training-free approach for text-driven image-to-image translation based on a pretrained text-to-image diffusion model. Our goal is to generate an image that aligns with the target task while preserving the structure and background of a source image. To this end, we derive the representation guidance with a combination of two objectives: maximizing the similarity to the target prompt based on the CLIP score and minimizing the structural distance to the source latent variable. This guidance improves the fidelity of the generated target image to the given target prompt while maintaining the structure integrity of the source image. To incorporate the representation guidance component, we optimize the target latent variable of diffusion model's reverse process with the guidance. Experimental results demonstrate that our method achieves outstanding image-to-image translation performance on various tasks when combined with the pretrained Stable Diffusion model.

Inju Ha, Donghun Ryou, Seonguk Seo et al.

Deep learning-based image denoising techniques often struggle with poor generalization performance to out-of-distribution real-world noise. To tackle this challenge, we propose a novel noise translation framework that performs denoising on an image with translated noise rather than directly denoising an original noisy image. Specifically, our approach translates complex, unknown real-world noise into Gaussian noise, which is spatially uncorrelated and independent of image content, through a noise translation network. The translated noisy images are then processed by an image denoising network pretrained to effectively remove Gaussian noise, enabling robust and consistent denoising performance. We also design well-motivated loss functions and architectures for the noise translation network by leveraging the mathematical properties of Gaussian noise. Experimental results demonstrate that the proposed method substantially improves robustness and generalizability, outperforming state-of-the-art methods across diverse benchmarks. Visualized denoising results and the source code are available on our project page.

Sanghyeok Chu, Seonguk Seo, Bohyung Han

Recent advances in vision-language models have led to impressive progress in caption generation for images and short video clips. However, these models remain constrained by their limited temporal receptive fields, making it difficult to produce coherent and comprehensive captions for long videos. While several methods have been proposed to aggregate information across video segments, they often rely on supervised fine-tuning or incur significant computational overhead. To address these challenges, we introduce a novel framework for long video captioning based on graph consolidation. Our approach first generates segment-level captions, corresponding to individual frames or short video intervals, using off-the-shelf visual captioning models. These captions are then parsed into individual scene graphs, which are subsequently consolidated into a unified graph representation that preserves both holistic context and fine-grained details throughout the video. A lightweight graph-to-text decoder then produces the final video-level caption. This framework effectively extends the temporal understanding capabilities of existing models without requiring any additional fine-tuning on long video datasets. Experimental results show that our method significantly outperforms existing LLM-based consolidation approaches, achieving strong zero-shot performance while substantially reducing computational costs.

Heewon Kim, Hyun Sung Chang, Kiho Cho et al.

Labor-intensive labeling becomes a bottleneck in developing computer vision algorithms based on deep learning. For this reason, dealing with imperfect labels has increasingly gained attention and has become an active field of study. We address learning with noisy labels (LNL) problem, which is formalized as a task of finding a structured manifold in the midst of noisy data. In this framework, we provide a proper objective function and an optimization algorithm based on two expectation-maximization (EM) cycles. The separate networks associated with the two EM cycles collaborate to optimize the objective function, where one model is for distinguishing clean labels from corrupted ones while the other is for refurbishing the corrupted labels. This approach results in a non-collapsing LNL-flywheel model in the end. Experiments show that our algorithm achieves state-of-the-art performance in multiple standard benchmarks with substantial margins under various types of label noise.

Jinyoung Choi, Junoh Kang, Bohyung Han

Diffusion probabilistic models (DPMs), while effective in generating high-quality samples, often suffer from high computational costs due to their iterative sampling process. To address this, we propose an enhanced ODE-based sampling method for DPMs inspired by Richardson extrapolation, which reduces numerical error and improves convergence rates. Our method, RX-DPM, leverages multiple ODE solutions at intermediate time steps to extrapolate the denoised prediction in DPMs. This significantly enhances the accuracy of estimations for the final sample while maintaining the number of function evaluations (NFEs). Unlike standard Richardson extrapolation, which assumes uniform discretization of the time grid, we develop a more general formulation tailored to arbitrary time step scheduling, guided by local truncation error derived from a baseline sampling method. The simplicity of our approach facilitates accurate estimation of numerical solutions without significant computational overhead, and allows for seamless and convenient integration into various DPMs and solvers. Additionally, RX-DPM provides explicit error estimates, effectively demonstrating the faster convergence as the leading error term's order increases. Through a series of experiments, we show that the proposed method improves the quality of generated samples without requiring additional sampling iterations.

Mijeong Kim, Jungtaek Kim, Bohyung Han

We present GP-4DGS, a novel framework that integrates Gaussian Processes (GPs) into 4D Gaussian Splatting (4DGS) for principled probabilistic modeling of dynamic scenes. While existing 4DGS methods focus on deterministic reconstruction, they are inherently limited in capturing motion ambiguity and lack mechanisms to assess prediction reliability. By leveraging the kernel-based probabilistic nature of GPs, our approach introduces three key capabilities: (i) uncertainty quantification for motion predictions, (ii) motion estimation for unobserved or sparsely sampled regions, and (iii) temporal extrapolation beyond observed training frames. To scale GPs to the large number of Gaussian primitives in 4DGS, we design spatio-temporal kernels that capture the correlation structure of deformation fields and adopt variational Gaussian Processes with inducing points for tractable inference. Our experiments show that GP-4DGS enhances reconstruction quality while providing reliable uncertainty estimates that effectively identify regions of high motion ambiguity. By addressing these challenges, our work takes a meaningful step toward bridging probabilistic modeling and neural graphics.