Jonghyun Kim

Suyeon Choi, Manu Gopakumar, Yifan et al.

Holographic near-eye displays offer unprecedented capabilities for virtual and augmented reality systems, including perceptually important focus cues. Although artificial intelligence--driven algorithms for computer-generated holography (CGH) have recently made much progress in improving the image quality and synthesis efficiency of holograms, these algorithms are not directly applicable to emerging phase-only spatial light modulators (SLM) that are extremely fast but offer phase control with very limited precision. The speed of these SLMs offers time multiplexing capabilities, essentially enabling partially-coherent holographic display modes. Here we report advances in camera-calibrated wave propagation models for these types of holographic near-eye displays and we develop a CGH framework that robustly optimizes the heavily quantized phase patterns of fast SLMs. Our framework is flexible in supporting runtime supervision with different types of content, including 2D and 2.5D RGBD images, 3D focal stacks, and 4D light fields. Using our framework, we demonstrate state-of-the-art results for all of these scenarios in simulation and experiment.

Geonung Kim, Kyoungkook Kang, Seongtae Kim et al.

For realistic and vivid colorization, generative priors have recently been exploited. However, such generative priors often fail for in-the-wild complex images due to their limited representation space. In this paper, we propose BigColor, a novel colorization approach that provides vivid colorization for diverse in-the-wild images with complex structures. While previous generative priors are trained to synthesize both image structures and colors, we learn a generative color prior to focus on color synthesis given the spatial structure of an image. In this way, we reduce the burden of synthesizing image structures from the generative prior and expand its representation space to cover diverse images. To this end, we propose a BigGAN-inspired encoder-generator network that uses a spatial feature map instead of a spatially-flattened BigGAN latent code, resulting in an enlarged representation space. Our method enables robust colorization for diverse inputs in a single forward pass, supports arbitrary input resolutions, and provides multi-modal colorization results. We demonstrate that BigColor significantly outperforms existing methods especially on in-the-wild images with complex structures.

Jonghyun Kim, Gen Li, Joongkyu Kim

Despite recent progress in semantic image synthesis, complete control over image style remains a challenging problem. Existing methods require reference images to feed style information into semantic layouts, which indicates that the style is constrained by the given image. In this paper, we propose a model named RUCGAN for user controllable semantic image synthesis, which utilizes a singular color to represent the style of a specific semantic region. The proposed network achieves reference-free semantic image synthesis by injecting color as user-desired styles into each semantic layout, and is able to synthesize semantic images with unusual colors. Extensive experimental results on various challenging datasets show that the proposed method outperforms existing methods, and we further provide an interactive UI to demonstrate the advantage of our approach for style controllability.

Bosang Kim, Jonghyun Kim, Hyotae Lee et al.

In general, hand pose estimation aims to improve the robustness of model performance in the real-world scenes. However, it is difficult to enhance the robustness since existing datasets are obtained in restricted environments to annotate 3D information. Although neural networks quantitatively achieve a high estimation accuracy, unsatisfied results can be observed in visual quality. This discrepancy between quantitative results and their visual qualities remains an open issue in the hand pose representation. To this end, we propose a mesh represented recycle learning strategy for 3D hand pose and mesh estimation which reinforces synthesized hand mesh representation in a training phase. To be specific, a hand pose and mesh estimation model first predicts parametric 3D hand annotations (i.e., 3D keypoint positions and vertices for hand mesh) with real-world hand images in the training phase. Second, synthetic hand images are generated with self-estimated hand mesh representations. After that, the synthetic hand images are fed into the same model again. Thus, the proposed learning strategy simultaneously improves quantitative results and visual qualities by reinforcing synthetic mesh representation. To encourage consistency between original model output and its recycled one, we propose self-correlation loss which maximizes the accuracy and reliability of our learning strategy. Consequently, the model effectively conducts self-refinement on hand pose estimation by learning mesh representation from its own output. To demonstrate the effectiveness of our learning strategy, we provide extensive experiments on FreiHAND dataset. Notably, our learning strategy improves the performance on hand pose and mesh estimation without any extra computational burden during the inference.

Juseung Yun, Sunwoo Yu, Sumin Ha et al.

Cancer progression arises from interactions across multiple biological layers, especially beyond morphological and across molecular layers that remain invisible to image-only models. To capture this broader biological landscape, we present EXAONE Path 2.5, a pathology foundation model that jointly models histologic, genomic, epigenetic and transcriptomic modalities, producing an integrated patient representation that reflects tumor biology more comprehensively. Our approach incorporates three key components: (1) multimodal SigLIP loss enabling all-pairwise contrastive learning across heterogeneous modalities, (2) a fragment-aware rotary positional encoding (F-RoPE) module that preserves spatial structure and tissue-fragment topology in WSI, and (3) domain-specialized internal foundation models for both WSI and RNA-seq to provide biologically grounded embeddings for robust multimodal alignment. We evaluate EXAONE Path 2.5 against six leading pathology foundation models across two complementary benchmarks: an internal real-world clinical dataset and the Patho-Bench benchmark covering 80 tasks. Our framework demonstrates high data and parameter efficiency, achieving on-par performance with state-of-the-art foundation models on Patho-Bench while exhibiting the highest adaptability in the internal clinical setting. These results highlight the value of biologically informed multimodal design and underscore the potential of integrated genotype-to-phenotype modeling for next-generation precision oncology.

Kang Min Yoo, Jaegeun Han, Sookyo In et al.

We introduce HyperCLOVA X, a family of large language models (LLMs) tailored to the Korean language and culture, along with competitive capabilities in English, math, and coding. HyperCLOVA X was trained on a balanced mix of Korean, English, and code data, followed by instruction-tuning with high-quality human-annotated datasets while abiding by strict safety guidelines reflecting our commitment to responsible AI. The model is evaluated across various benchmarks, including comprehensive reasoning, knowledge, commonsense, factuality, coding, math, chatting, instruction-following, and harmlessness, in both Korean and English. HyperCLOVA X exhibits strong reasoning capabilities in Korean backed by a deep understanding of the language and cultural nuances. Further analysis of the inherent bilingual nature and its extension to multilingualism highlights the model's cross-lingual proficiency and strong generalization ability to untargeted languages, including machine translation between several language pairs and cross-lingual inference tasks. We believe that HyperCLOVA X can provide helpful guidance for regions or countries in developing their sovereign LLMs.

Minsoo Lee, Jonghyun Kim, Juseung Yun et al.

Pathology foundation models learn morphological representations through self-supervised pretraining on large-scale whole-slide images, yet they do not explicitly capture the underlying molecular state of the tissue. Spatial transcriptomics technologies bridge this gap by measuring gene expression in situ, offering a natural cross-modal supervisory signal. We propose MINT (Molecularly Informed Training), a fine-tuning framework that incorporates spatial transcriptomics supervision into pretrained pathology Vision Transformers. MINT appends a learnable ST token to the ViT input to encode transcriptomic information separately from the morphological CLS token, preventing catastrophic forgetting through DINO self-distillation and explicit feature anchoring to the frozen pretrained encoder. Gene expression regression at both spot-level (Visium) and patch-level (Xenium) resolutions provides complementary supervision across spatial scales. Trained on 577 publicly available HEST samples, MINT achieves the best overall performance on both HEST-Bench for gene expression prediction (mean Pearson r = 0.440) and EVA for general pathology tasks (0.803), demonstrating that spatial transcriptomics supervision complements morphology-centric self-supervised pretraining.

Myeongjang Pyeon, Janghyeon Lee, Minsoo Lee et al.

In digital pathology, whole-slide images (WSIs) are often difficult to handle due to their gigapixel scale, so most approaches train patch encoders via self-supervised learning (SSL) and then aggregate the patch-level embeddings via multiple instance learning (MIL) or slide encoders for downstream tasks. However, patch-level SSL may overlook complex domain-specific features that are essential for biomarker prediction, such as mutation status and molecular characteristics, as SSL methods rely only on basic augmentations selected for natural image domains on small patch-level area. Moreover, SSL methods remain less data efficient than fully supervised approaches, requiring extensive computational resources and datasets to achieve competitive performance. To address these limitations, we present EXAONE Path 2.0, a pathology foundation model that learns patch-level representations under direct slide-level supervision. Using only 37k WSIs for training, EXAONE Path 2.0 achieves state-of-the-art average performance across 10 biomarker prediction tasks, demonstrating remarkable data efficiency.

Jonghyun Kim, Bosang Kim, Hyotae Lee et al.

In general, human pose estimation methods are categorized into two approaches according to their architectures: regression (i.e., heatmap-free) and heatmap-based methods. The former one directly estimates precise coordinates of each keypoint using convolutional and fully-connected layers. Although this approach is able to detect overlapped and dense keypoints, unexpected results can be obtained by non-existent keypoints in a scene. On the other hand, the latter one is able to filter the non-existent ones out by utilizing predicted heatmaps for each keypoint. Nevertheless, it suffers from quantization error when obtaining the keypoint coordinates from its heatmaps. In addition, unlike the regression one, it is difficult to distinguish densely placed keypoints in an image. To this end, we propose a hybrid model for single-stage multi-person pose estimation, named HybridPose, which mutually overcomes each drawback of both approaches by maximizing their strengths. Furthermore, we introduce self-correlation loss to inject spatial dependencies between keypoint coordinates and their visibility. Therefore, HybridPose is capable of not only detecting densely placed keypoints, but also filtering the non-existent keypoints in an image. Experimental results demonstrate that proposed HybridPose exhibits the keypoints visibility without performance degradation in terms of the pose estimation accuracy.

Jonghyun Kim, Gen Li, Cheolkon Jung et al.

Existing methods for image synthesis utilized a style encoder based on stacks of convolutions and pooling layers to generate style codes from input images. However, the encoded vectors do not necessarily contain local information of the corresponding images since small-scale objects are tended to "wash away" through such downscaling procedures. In this paper, we propose deep image synthesis with superpixel based style encoder, named as SuperStyleNet. First, we directly extract the style codes from the original image based on superpixels to consider local objects. Second, we recover spatial relationships in vectorized style codes based on graphical analysis. Thus, the proposed network achieves high-quality image synthesis by mapping the style codes into semantic labels. Experimental results show that the proposed method outperforms state-of-the-art ones in terms of visual quality and quantitative measurements. Furthermore, we achieve elaborate spatial style editing by adjusting style codes.

Gen Li, Varun Jampani, Laura Sevilla-Lara et al.

Prototype learning is extensively used for few-shot segmentation. Typically, a single prototype is obtained from the support feature by averaging the global object information. However, using one prototype to represent all the information may lead to ambiguities. In this paper, we propose two novel modules, named superpixel-guided clustering (SGC) and guided prototype allocation (GPA), for multiple prototype extraction and allocation. Specifically, SGC is a parameter-free and training-free approach, which extracts more representative prototypes by aggregating similar feature vectors, while GPA is able to select matched prototypes to provide more accurate guidance. By integrating the SGC and GPA together, we propose the Adaptive Superpixel-guided Network (ASGNet), which is a lightweight model and adapts to object scale and shape variation. In addition, our network can easily generalize to k-shot segmentation with substantial improvement and no additional computational cost. In particular, our evaluations on COCO demonstrate that ASGNet surpasses the state-of-the-art method by 5% in 5-shot segmentation.

Jonghyun Kim, Gen Li, Inyong Yun et al.

Face super-resolution (SR) has become an indispensable function in security solutions such as video surveillance and identification system, but the distortion in facial components is a great challenge in it. Most state-of-the-art methods have utilized facial priors with deep neural networks. These methods require extra labels, longer training time, and larger computation memory. In this paper, we propose a novel Edge and Identity Preserving Network for Face SR Network, named as EIPNet, to minimize the distortion by utilizing a lightweight edge block and identity information. We present an edge block to extract perceptual edge information, and concatenate it to the original feature maps in multiple scales. This structure progressively provides edge information in reconstruction to aggregate local and global structural information. Moreover, we define an identity loss function to preserve identification of SR images. The identity loss function compares feature distributions between SR images and their ground truth to recover identities in SR images. In addition, we provide a luminance-chrominance error (LCE) to separately infer brightness and color information in SR images. The LCE method not only reduces the dependency of color information by dividing brightness and color components but also enables our network to reflect differences between SR images and their ground truth in two color spaces of RGB and YUV. The proposed method facilitates the proposed SR network to elaborately restore facial components and generate high quality 8x scaled SR images with a lightweight network structure. Furthermore, our network is able to reconstruct an 128x128 SR image with 215 fps on a GTX 1080Ti GPU. Extensive experiments demonstrate that our network qualitatively and quantitatively outperforms state-of-the-art methods on two challenging datasets: CelebA and VGGFace2.

Makiya Nakashima, Alex Sim, Youngsoo Kim et al.

While variable selection is essential to optimize the learning complexity by prioritizing features, automating the selection process is preferred since it requires laborious efforts with intensive analysis otherwise. However, it is not an easy task to enable the automation due to several reasons. First, selection techniques often need a condition to terminate the reduction process, for example, by using a threshold or the number of features to stop, and searching an adequate stopping condition is highly challenging. Second, it is uncertain that the reduced variable set would work well; our preliminary experimental result shows that well-known selection techniques produce different sets of variables as a result of reduction (even with the same termination condition), and it is hard to estimate which of them would work the best in future testing. In this paper, we demonstrate the potential power of our approach to the automation of selection process that incorporates well-known selection methods identifying important variables. Our experimental results with two public network traffic data (UNSW-NB15 and IDS2017) show that our proposed method identifies a small number of core variables, with which it is possible to approximate the performance to the one with the entire variables.

Gen Li, Inyoung Yun, Jonghyun Kim et al.

As a pixel-level prediction task, semantic segmentation needs large computational cost with enormous parameters to obtain high performance. Recently, due to the increasing demand for autonomous systems and robots, it is significant to make a tradeoff between accuracy and inference speed. In this paper, we propose a novel Depthwise Asymmetric Bottleneck (DAB) module to address this dilemma, which efficiently adopts depth-wise asymmetric convolution and dilated convolution to build a bottleneck structure. Based on the DAB module, we design a Depth-wise Asymmetric Bottleneck Network (DABNet) especially for real-time semantic segmentation, which creates sufficient receptive field and densely utilizes the contextual information. Experiments on Cityscapes and CamVid datasets demonstrate that the proposed DABNet achieves a balance between speed and precision. Specifically, without any pretrained model and postprocessing, it achieves 70.1% Mean IoU on the Cityscapes test dataset with only 0.76 million parameters and a speed of 104 FPS on a single GTX 1080Ti card.

Josef Spjut, Ben Boudaoud, Jonghyun Kim et al.

Emergent in the field of head mounted display design is a desire to leverage the limitations of the human visual system to reduce the computation, communication, and display workload in power and form-factor constrained systems. Fundamental to this reduced workload is the ability to match display resolution to the acuity of the human visual system, along with a resulting need to follow the gaze of the eye as it moves, a process referred to as foveation. A display that moves its content along with the eye may be called a Foveated Display, though this term is also commonly used to describe displays with non-uniform resolution that attempt to mimic human visual acuity. We therefore recommend a definition for the term Foveated Display that accepts both of these interpretations. Furthermore, we include a simplified model for human visual Acuity Distribution Functions (ADFs) at various levels of visual acuity, across wide fields of view and propose comparison of this ADF with the Resolution Distribution Function of a foveated display for evaluation of its resolution at a particular gaze direction. We also provide a taxonomy to allow the field to meaningfully compare and contrast various aspects of foveated displays in a display and optical technology-agnostic manner.