Jaeseong Lee

Zheng Chen, Kai Liu, Jingkai Wang et al.

This paper presents the NTIRE 2026 image super-resolution ($\times$4) challenge, one of the associated competitions of the NTIRE 2026 Workshop at CVPR 2026. The challenge aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs generated through bicubic downsampling with a $\times$4 scaling factor. The objective is to develop effective super-resolution solutions and analyze recent advances in the field. To reflect the evolving objectives of image super-resolution, the challenge includes two tracks: (1) a restoration track, which emphasizes pixel-wise fidelity and ranks submissions based on PSNR; and (2) a perceptual track, which focuses on visual realism and evaluates results using a perceptual score. A total of 194 participants registered for the challenge, with 31 teams submitting valid entries. This report summarizes the challenge design, datasets, evaluation protocol, main results, and methods of participating teams. The challenge provides a unified benchmark and offers insights into current progress and future directions in image super-resolution.

Jaeseong Lee, Taewoo Kim, Sunghyun Park et al.

Face swapping aims at injecting a source image's identity (i.e., facial features) into a target image, while strictly preserving the target's attributes, which are irrelevant to identity. However, we observed that previous approaches still suffer from source attribute leakage, where the source image's attributes interfere with the target image's. In this paper, we analyze the latent space of StyleGAN and find the adequate combination of the latents geared for face swapping task. Based on the findings, we develop a simple yet robust face swapping model, RobustSwap, which is resistant to the potential source attribute leakage. Moreover, we exploit the coordination of 3DMM's implicit and explicit information as a guidance to incorporate the structure of the source image and the precise pose of the target image. Despite our method solely utilizing an image dataset without identity labels for training, our model has the capability to generate high-fidelity and temporally consistent videos. Through extensive qualitative and quantitative evaluations, we demonstrate that our method shows significant improvements compared with the previous face swapping models in synthesizing both images and videos. Project page is available at https://robustswap.github.io/

Taewoong Kang, Jeongsik Oh, Jaeseong Lee et al.

Despite the remarkable advancements in head reenactment, the existing methods face challenges in cross-domain head reenactment, which aims to transfer human motions to domains outside the human, including cartoon characters. It is still difficult to extract motion from out-of-domain images due to the distinct appearances, such as large eyes. Recently, previous work introduced a large-scale anime dataset called AnimeCeleb and a cross-domain head reenactment model, including an optimization-based mapping function to translate the human domain's expressions to the anime domain. However, we found that the mapping function, which relies on a subset of expressions, imposes limitations on the mapping of various expressions. To solve this challenge, we introduce a novel expression domain translation network that transforms human expressions into anime expressions. Specifically, to maintain the geometric consistency of expressions between the input and output of the expression domain translation network, we employ a 3D geometric-aware loss function that reduces the distances between the vertices in the 3D mesh of the human and anime. By doing so, it forces high-fidelity and one-to-one mapping with respect to two cross-expression domains. Our method outperforms existing methods in both qualitative and quantitative analysis, marking a significant advancement in the field of cross-domain head reenactment.

Gyumin Shim, Jaeseong Lee, Junha Hyung et al.

In this paper, we propose PixelHuman, a novel human rendering model that generates animatable human scenes from a few images of a person with unseen identity, views, and poses. Previous work have demonstrated reasonable performance in novel view and pose synthesis, but they rely on a large number of images to train and are trained per scene from videos, which requires significant amount of time to produce animatable scenes from unseen human images. Our method differs from existing methods in that it can generalize to any input image for animatable human synthesis. Given a random pose sequence, our method synthesizes each target scene using a neural radiance field that is conditioned on a canonical representation and pose-aware pixel-aligned features, both of which can be obtained through deformation fields learned in a data-driven manner. Our experiments show that our method achieves state-of-the-art performance in multiview and novel pose synthesis from few-shot images.

Jaeseong Lee, seung-won hwang, Aurick Qiao et al.

Mixture-of-experts (MoEs) have been adopted for reducing inference costs by sparsely activating experts in Large language models (LLMs). Despite this reduction, the massive number of experts in MoEs still makes them expensive to serve. In this paper, we study how to address this, by pruning MoEs. Among pruning methodologies, unstructured pruning has been known to achieve the highest performance for a given pruning ratio, compared to structured pruning, since the latter imposes constraints on the sparsification structure. This is intuitive, as the solution space of unstructured pruning subsumes that of structured pruning. However, our counterintuitive finding reveals that expert pruning, a form of structured pruning, can actually precede unstructured pruning to outperform unstructured-only pruning. As existing expert pruning, requiring $O(\frac{k^n}{\sqrt{n}})$ forward passes for $n$ experts, cannot scale for recent MoEs, we propose a scalable alternative with $O(1)$ complexity, yet outperforming the more expensive methods. The key idea is leveraging a latent structure between experts, based on behavior similarity, such that the greedy decision of whether to prune closely captures the joint pruning effect. Ours is highly effective -- for Snowflake Arctic, a 480B-sized MoE with 128 experts, our method needs only one H100 and two hours to achieve nearly no loss in performance with 40% sparsity, even in generative tasks such as GSM8K, where state-of-the-art unstructured pruning fails to. The code will be made publicly available.

Jaeseong Lee, Junyeong Ahn, Taewoong Kang et al.

Constructing drivable and photorealistic 3D head avatars has become a central task in AR/XR, enabling immersive and expressive user experiences. With the emergence of high-fidelity and efficient representations such as 3D Gaussians, recent works have pushed toward ultra-detailed head avatars. Existing approaches typically fall into two categories: rule-based analytic rigging or neural network-based deformation fields. While effective in constrained settings, both approaches often fail to generalize to unseen expressions and poses, particularly in extreme reenactment scenarios. Other methods constrain Gaussians to the global texel space of 3DMMs to reduce rendering complexity. However, these texel-based avatars tend to underutilize the underlying mesh structure. They apply minimal analytic deformation and rely heavily on neural regressors and heuristic regularization in UV space, which weakens geometric consistency and limits extrapolation to complex, out-of-distribution deformations. To address these limitations, we introduce TexAvatars, a hybrid avatar representation that combines the explicit geometric grounding of analytic rigging with the spatial continuity of texel space. Our approach predicts local geometric attributes in UV space via CNNs, but drives 3D deformation through mesh-aware Jacobians, enabling smooth and semantically meaningful transitions across triangle boundaries. This hybrid design separates semantic modeling from geometric control, resulting in improved generalization, interpretability, and stability. Furthermore, TexAvatars captures fine-grained expression effects, including muscle-induced wrinkles, glabellar lines, and realistic mouth cavity geometry, with high fidelity. Our method achieves state-of-the-art performance under extreme pose and expression variations, demonstrating strong generalization in challenging head reenactment settings.

Jaeseong Lee, Yeeun Choi, Heechan Choi et al.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in vision-language understanding, reasoning, and generation. However, they struggle with tasks requiring fine-grained localization and reasoning in high-resolution images. This constraint stems from the fact that MLLMs are fine-tuned with fixed image resolution to align with the pre-trained image encoder used in MLLM. Consequently, feeding high-resolution images directly into MLLMs leads to poor generalization due to a train-test resolution discrepancy, while downsampling these images-although ensuring consistency-compromises fine-grained visual details and ultimately degrades performance. To address this challenge, we propose Extract Candidate then Predict (ECP), a novel training-free, task-agnostic two-stage framework designed to enhance MLLM performance on high-resolution images. The key intuition behind ECP is that while MLLMs struggle with high-resolution images, their predictions on downsampled images still contain implicit localization cues. By first identifying candidate region using the coarse prediction and then predicting the final output based on candidate region, ECP effectively preserves fine-grained details while mitigating the challenges posed by high-resolution data. We validate our framework on 4K GUI grounding and 4K, 8K MLLM perception, achieving +21.3%, +5.8%, +5.2% absolute improvement compared to baseline respectively, demonstrating its effectiveness. Code is available at https://github.com/yenncye/ECP.

Kangyeol Kim, Sunghyun Park, Jaeseong Lee et al.

We present a novel Animation CelebHeads dataset (AnimeCeleb) to address an animation head reenactment. Different from previous animation head datasets, we utilize 3D animation models as the controllable image samplers, which can provide a large amount of head images with their corresponding detailed pose annotations. To facilitate a data creation process, we build a semi-automatic pipeline leveraging an open 3D computer graphics software with a developed annotation system. After training with the AnimeCeleb, recent head reenactment models produce high-quality animation head reenactment results, which are not achievable with existing datasets. Furthermore, motivated by metaverse application, we propose a novel pose mapping method and architecture to tackle a cross-domain head reenactment task. During inference, a user can easily transfer one's motion to an arbitrary animation head. Experiments demonstrate the usefulness of the AnimeCeleb to train animation head reenactment models, and the superiority of our cross-domain head reenactment model compared to state-of-the-art methods. Our dataset and code are available at https://github.com/kangyeolk/AnimeCeleb.

Jaeseong Lee, Duseok Kang, Soonhoi Ha

As the application area of convolutional neural networks (CNN) is growing in embedded devices, it becomes popular to use a hardware CNN accelerator, called neural processing unit (NPU), to achieve higher performance per watt than CPUs or GPUs. Recently, automated neural architecture search (NAS) emerges as the default technique to find a state-of-the-art CNN architecture with higher accuracy than manually-designed architectures for image classification. In this paper, we present a fast NPU-aware NAS methodology, called S3NAS, to find a CNN architecture with higher accuracy than the existing ones under a given latency constraint. It consists of three steps: supernet design, Single-Path NAS for fast architecture exploration, and scaling. To widen the search space of the supernet structure that consists of stages, we allow stages to have a different number of blocks and blocks to have parallel layers of different kernel sizes. For a fast neural architecture search, we apply a modified Single-Path NAS technique to the proposed supernet structure. In this step, we assume a shorter latency constraint than the required to reduce the search space and the search time. The last step is to scale up the network maximally within the latency constraint. For accurate latency estimation, an analytical latency estimator is devised, based on a cycle-level NPU simulator that runs an entire CNN considering the memory access overhead accurately. With the proposed methodology, we are able to find a network in 3 hours using TPUv3, which shows 82.72% top-1 accuracy on ImageNet with 11.66 ms latency. Code are released at https://github.com/cap-lab/S3NAS

Jaeseong Lee, Taewoong Kang, Marcel C. Bühler et al.

Recent advancements in head avatar rendering using Gaussian primitives have achieved significantly high-fidelity results. Although precise head geometry is crucial for applications like mesh reconstruction and relighting, current methods struggle to capture intricate geometric details and render unseen poses due to their reliance on similarity transformations, which cannot handle stretch and shear transforms essential for detailed deformations of geometry. To address this, we propose SurFhead, a novel method that reconstructs riggable head geometry from RGB videos using 2D Gaussian surfels, which offer well-defined geometric properties, such as precise depth from fixed ray intersections and normals derived from their surface orientation, making them advantageous over 3D counterparts. SurFhead ensures high-fidelity rendering of both normals and images, even in extreme poses, by leveraging classical mesh-based deformation transfer and affine transformation interpolation. SurFhead introduces precise geometric deformation and blends surfels through polar decomposition of transformations, including those affecting normals. Our key contribution lies in bridging classical graphics techniques, such as mesh-based deformation, with modern Gaussian primitives, achieving state-of-the-art geometry reconstruction and rendering quality. Unlike previous avatar rendering approaches, SurFhead enables efficient reconstruction driven by Gaussian primitives while preserving high-fidelity geometry.

Jeonghoon Park, Juyoung Lee, Chaeyeon Chung et al.

Recent advancements in diffusion-based text-to-image (T2I) models have enabled the generation of high-quality and photorealistic images from text. However, they often exhibit societal biases related to gender, race, and socioeconomic status, thereby potentially reinforcing harmful stereotypes and shaping public perception in unintended ways. While existing bias mitigation methods demonstrate effectiveness, they often encounter attribute entanglement, where adjustments to attributes relevant to the bias (i.e., target attributes) unintentionally alter attributes unassociated with the bias (i.e., non-target attributes), causing undesirable distribution shifts. To address this challenge, we introduce Entanglement-Free Attention (EFA), a method that accurately incorporates target attributes (e.g., White, Black, and Asian) while preserving non-target attributes (e.g., background) during bias mitigation. At inference time, EFA randomly samples a target attribute with equal probability and adjusts the cross-attention in selected layers to incorporate the sampled attribute, achieving a fair distribution of target attributes. Extensive experiments demonstrate that EFA outperforms existing methods in mitigating bias while preserving non-target attributes, thereby maintaining the original model's output distribution and generative capacity.

Jaeseong Lee, Junha Hyung, Sohyun Jeong et al.

Face swapping has gained significant attention for its varied applications. Most previous face swapping approaches have relied on the seesaw game training scheme, also known as the target-oriented approach. However, this often leads to instability in model training and results in undesired samples with blended identities due to the target identity leakage problem. Source-oriented methods achieve more stable training with self-reconstruction objective but often fail to accurately reflect target image's skin color and illumination. This paper introduces the Shape Agnostic Masked AutoEncoder (SAMAE) training scheme, a novel self-supervised approach that combines the strengths of both target-oriented and source-oriented approaches. Our training scheme addresses the limitations of traditional training methods by circumventing the conventional seesaw game and introducing clear ground truth through its self-reconstruction training regime. Our model effectively mitigates identity leakage and reflects target albedo and illumination through learned disentangled identity and non-identity features. Additionally, we closely tackle the shape misalignment and volume discrepancy problems with new techniques, including perforation confusion and random mesh scaling. SAMAE establishes a new state-of-the-art, surpassing other baseline methods, preserving both identity and non-identity attributes without sacrificing on either aspect.

Jaeseong Lee, seung-won hwang, Aurick Qiao et al.

Speculative decoding promises faster inference for large language models (LLMs), yet existing methods fail to generalize to real-world settings. Benchmarks typically assume short contexts (e.g., 2K tokens), whereas practical workloads involve long contexts. We find current approaches degrade severely with long contexts; for instance, EAGLE3 even slows down the generation speed by 0.81x. We address these limitations by releasing a new long-context benchmark (LongSpecBench) and introducing a novel model (OWL). OWL achieves about 5x higher acceptance length than EAGLE3 on long-context inputs through three innovations: (1) an LSTM-based drafter conditioned only on the last-token state, making it generalize to various lengths, (2) a special token [SPEC] in the verifier that produces richer representation for drafter, and (3) a hybrid algorithm combining both tree and non-tree decoding methods. We release all code and datasets to advance future research.

Jaeseong Lee, Dayoung Kwon, seung-won hwang

Large Reasoning Models (LRMs) excel in structured tasks by emulating deliberate human reasoning but often suffer from overthinking, degrading performance and wasting resources. One possible baseline is to deploy both LLM and LRM, then route input by predicting whether it requires reasoning and may cause overthinking. However, deploying multiple models can be costly or impractical. We propose a superposed deployment strategy with a lightweight, training-free regulation to optimize inference by switching one model on and off. Instead of routing, we selectively unlearn from LRM at inference, scaling down computation while preserving reasoning. By analyzing the cumulative energy of singular values, we identify optimal low-rank projections to adjust reasoning just right.

Parth Padalkar, Jaeseong Lee, Shiyi Wei et al.

There has been significant focus on creating neuro-symbolic models for interpretable image classification using Convolutional Neural Networks (CNNs). These methods aim to replace the CNN with a neuro-symbolic model consisting of the CNN, which is used as a feature extractor, and an interpretable rule-set extracted from the CNN itself. While these approaches provide interpretability through the extracted rule-set, they often compromise accuracy compared to the original CNN model. In this paper, we identify the root cause of this accuracy loss as the post-training binarization of filter activations to extract the rule-set. To address this, we propose a novel sparsity loss function that enables class-specific filter binarization during CNN training, thus minimizing information loss when extracting the rule-set. We evaluate several training strategies with our novel sparsity loss, analyzing their effectiveness and providing guidance on their appropriate use. Notably, we set a new benchmark, achieving a 9% improvement in accuracy and a 53% reduction in rule-set size on average, compared to the previous SOTA, while coming within 3% of the original CNN's accuracy. This highlights the significant potential of interpretable neuro-symbolic models as viable alternatives to black-box CNNs.

Yeonjoon Jung, Jaeseong Lee, Seungtaek Choi et al.

Recently, pre-trained language models (PLMs) have been increasingly adopted in spoken language understanding (SLU). However, automatic speech recognition (ASR) systems frequently produce inaccurate transcriptions, leading to noisy inputs for SLU models, which can significantly degrade their performance. To address this, our objective is to train SLU models to withstand ASR errors by exposing them to noises commonly observed in ASR systems, referred to as ASR-plausible noises. Speech noise injection (SNI) methods have pursued this objective by introducing ASR-plausible noises, but we argue that these methods are inherently biased towards specific ASR systems, or ASR-specific noises. In this work, we propose a novel and less biased augmentation method of introducing the noises that are plausible to any ASR system, by cutting off the non-causal effect of noises. Experimental results and analyses demonstrate the effectiveness of our proposed methods in enhancing the robustness and generalizability of SLU models against unseen ASR systems by introducing more diverse and plausible ASR noises in advance.

Junha Hyung, Susung Hong, Sungwon Hwang et al.

3D reconstruction from multi-view images is one of the fundamental challenges in computer vision and graphics. Recently, 3D Gaussian Splatting (3DGS) has emerged as a promising technique capable of real-time rendering with high-quality 3D reconstruction. This method utilizes 3D Gaussian representation and tile-based splatting techniques, bypassing the expensive neural field querying. Despite its potential, 3DGS encounters challenges such as needle-like artifacts, suboptimal geometries, and inaccurate normals caused by the Gaussians converging into anisotropic shapes with one dominant variance. We propose using the effective rank analysis to examine the shape statistics of 3D Gaussian primitives, and identify the Gaussians indeed converge into needle-like shapes with the effective rank 1. To address this, we introduce the effective rank as a regularization, which constrains the structure of the Gaussians. Our new regularization method enhances normal and geometry reconstruction while reducing needle-like artifacts. The approach can be integrated as an add-on module to other 3DGS variants, improving their quality without compromising visual fidelity. The project page is available at https://junhahyung.github.io/erankgs.github.io.

Sungsoo Kim, Jin Soo Park, Christos G. Bampis et al.

We propose a video compression framework using conditional Generative Adversarial Networks (GANs). We rely on two encoders: one that deploys a standard video codec and another which generates low-level maps via a pipeline of down-sampling, a newly devised soft edge detector, and a novel lossless compression scheme. For decoding, we use a standard video decoder as well as a neural network based one, which is trained using a conditional GAN. Recent "deep" approaches to video compression require multiple videos to pre-train generative networks to conduct interpolation. In contrast to this prior work, our scheme trains a generative decoder on pairs of a very limited number of key frames taken from a single video and corresponding low-level maps. The trained decoder produces reconstructed frames relying on a guidance of low-level maps, without any interpolation. Experiments on a diverse set of 131 videos demonstrate that our proposed GAN-based compression engine achieves much higher quality reconstructions at very low bitrates than prevailing standard codecs such as H.264 or HEVC.