Sunghyun Park

Sangyun Lee, Gyojung Gu, Sunghyun Park et al.

Image-based virtual try-on aims to synthesize an image of a person wearing a given clothing item. To solve the task, the existing methods warp the clothing item to fit the person's body and generate the segmentation map of the person wearing the item before fusing the item with the person. However, when the warping and the segmentation generation stages operate individually without information exchange, the misalignment between the warped clothes and the segmentation map occurs, which leads to the artifacts in the final image. The information disconnection also causes excessive warping near the clothing regions occluded by the body parts, so-called pixel-squeezing artifacts. To settle the issues, we propose a novel try-on condition generator as a unified module of the two stages (i.e., warping and segmentation generation stages). A newly proposed feature fusion block in the condition generator implements the information exchange, and the condition generator does not create any misalignment or pixel-squeezing artifacts. We also introduce discriminator rejection that filters out the incorrect segmentation map predictions and assures the performance of virtual try-on frameworks. Experiments on a high-resolution dataset demonstrate that our model successfully handles the misalignment and occlusion, and significantly outperforms the baselines. Code is available at https://github.com/sangyun884/HR-VITON.

Ruihang Lai, Junru Shao, Siyuan Feng et al. · openai, uw

Dynamic shape computations have become critical in modern machine learning workloads, especially in emerging large language models. The success of these models has driven the demand for their universal deployment across a diverse set of backend environments. In this paper, we present Relax, a compiler abstraction for optimizing end-to-end dynamic machine learning workloads. Relax introduces a cross-level abstraction that encapsulates computational graphs, loop-level tensor programs, and external library calls in a single representation. Relax also introduces first-class symbolic shape annotations to track dynamic shape computations globally across the program, enabling dynamic shape-aware cross-level optimizations. We build an end-to-end compilation framework using the proposed approach to optimize dynamic shape models. Experimental results on LLMs show that Relax delivers performance competitive with state-of-the-art systems across various GPUs and enables deployment of emerging models to a broader set of emerging environments, including mobile phones, embedded devices, and web browsers.

Chaeyeon Chung, Sunghyun Park, Jeongho Kim et al.

Hairstyle transfer is a challenging task in the image editing field that modifies the hairstyle of a given face image while preserving its other appearance and background features. The existing hairstyle transfer approaches heavily rely on StyleGAN, which is pre-trained on cropped and aligned face images. Hence, they struggle to generalize under challenging conditions such as extreme variations of head poses or focal lengths. To address this issue, we propose a one-stage hairstyle transfer diffusion model, HairFusion, that applies to real-world scenarios. Specifically, we carefully design a hair-agnostic representation as the input of the model, where the original hair information is thoroughly eliminated. Next, we introduce a hair align cross-attention (Align-CA) to accurately align the reference hairstyle with the face image while considering the difference in their head poses. To enhance the preservation of the face image's original features, we leverage adaptive hair blending during the inference, where the output's hair regions are estimated by the cross-attention map in Align-CA and blended with non-hair areas of the face image. Our experimental results show that our method achieves state-of-the-art performance compared to the existing methods in preserving the integrity of both the transferred hairstyle and the surrounding features. The codes are available at https://github.com/cychungg/HairFusion

Yeon Seonwoo, Guoyin Wang, Changmin Seo et al. · pku

Unsupervised sentence representation learning has progressed through contrastive learning and data augmentation methods such as dropout masking. Despite this progress, sentence encoders are still limited to using only an input sentence when predicting its semantic vector. In this work, we show that the semantic meaning of a sentence is also determined by nearest-neighbor sentences that are similar to the input sentence. Based on this finding, we propose a novel unsupervised sentence encoder, RankEncoder. RankEncoder predicts the semantic vector of an input sentence by leveraging its relationship with other sentences in an external corpus, as well as the input sentence itself. We evaluate RankEncoder on semantic textual benchmark datasets. From the experimental results, we verify that 1) RankEncoder achieves 80.07% Spearman's correlation, a 1.1% absolute improvement compared to the previous state-of-the-art performance, 2) RankEncoder is universally applicable to existing unsupervised sentence embedding methods, and 3) RankEncoder is specifically effective for predicting the similarity scores of similar sentence pairs.

Taewoo Kim, Chaeyeon Chung, Yoonseo Kim et al.

Editing hairstyle is unique and challenging due to the complexity and delicacy of hairstyle. Although recent approaches significantly improved the hair details, these models often produce undesirable outputs when a pose of a source image is considerably different from that of a target hair image, limiting their real-world applications. HairFIT, a pose-invariant hairstyle transfer model, alleviates this limitation yet still shows unsatisfactory quality in preserving delicate hair textures. To solve these limitations, we propose a high-performing pose-invariant hairstyle transfer model equipped with latent optimization and a newly presented local-style-matching loss. In the StyleGAN2 latent space, we first explore a pose-aligned latent code of a target hair with the detailed textures preserved based on local style matching. Then, our model inpaints the occlusions of the source considering the aligned target hair and blends both images to produce a final output. The experimental results demonstrate that our model has strengths in transferring a hairstyle under larger pose differences and preserving local hairstyle textures.

Sunghyun Park, Seunghan Yang, Jaegul Choo et al.

Test-time adaptation (TTA) aims to adapt a pre-trained model to the target domain in a batch-by-batch manner during inference. While label distributions often exhibit imbalances in real-world scenarios, most previous TTA approaches typically assume that both source and target domain datasets have balanced label distribution. Due to the fact that certain classes appear more frequently in certain domains (e.g., buildings in cities, trees in forests), it is natural that the label distribution shifts as the domain changes. However, we discover that the majority of existing TTA methods fail to address the coexistence of covariate and label shifts. To tackle this challenge, we propose a novel label shift adapter that can be incorporated into existing TTA approaches to deal with label shifts during the TTA process effectively. Specifically, we estimate the label distribution of the target domain to feed it into the label shift adapter. Subsequently, the label shift adapter produces optimal parameters for the target label distribution. By predicting only the parameters for a part of the pre-trained source model, our approach is computationally efficient and can be easily applied, regardless of the model architectures. Through extensive experiments, we demonstrate that integrating our strategy with TTA approaches leads to substantial performance improvements under the joint presence of label and covariate shifts.

Kangyeol Kim, Sunghyun Park, Junsoo Lee et al.

Recent remarkable improvements in large-scale text-to-image generative models have shown promising results in generating high-fidelity images. To further enhance editability and enable fine-grained generation, we introduce a multi-input-conditioned image composition model that incorporates a sketch as a novel modal, alongside a reference image. Thanks to the edge-level controllability using sketches, our method enables a user to edit or complete an image sub-part with a desired structure (i.e., sketch) and content (i.e., reference image). Our framework fine-tunes a pre-trained diffusion model to complete missing regions using the reference image while maintaining sketch guidance. Albeit simple, this leads to wide opportunities to fulfill user needs for obtaining the in-demand images. Through extensive experiments, we demonstrate that our proposed method offers unique use cases for image manipulation, enabling user-driven modifications of arbitrary scenes.

Ke Bai, Guoyin Wang, Jiwei Li et al.

Open world classification is a task in natural language processing with key practical relevance and impact. Since the open or {\em unknown} category data only manifests in the inference phase, finding a model with a suitable decision boundary accommodating for the identification of known classes and discrimination of the open category is challenging. The performance of existing models is limited by the lack of effective open category data during the training stage or the lack of a good mechanism to learn appropriate decision boundaries. We propose an approach based on \underline{a}daptive \underline{n}egative \underline{s}amples (ANS) designed to generate effective synthetic open category samples in the training stage and without requiring any prior knowledge or external datasets. Empirically, we find a significant advantage in using auxiliary one-versus-rest binary classifiers, which effectively utilize the generated negative samples and avoid the complex threshold-seeking stage in previous works. Extensive experiments on three benchmark datasets show that ANS achieves significant improvements over state-of-the-art methods.

Chaeyeon Chung, Taewoo Kim, Hyelin Nam et al.

Hairstyle transfer is the task of modifying a source hairstyle to a target one. Although recent hairstyle transfer models can reflect the delicate features of hairstyles, they still have two major limitations. First, the existing methods fail to transfer hairstyles when a source and a target image have different poses (e.g., viewing direction or face size), which is prevalent in the real world. Also, the previous models generate unrealistic images when there is a non-trivial amount of regions in the source image occluded by its original hair. When modifying long hair to short hair, shoulders or backgrounds occluded by the long hair need to be inpainted. To address these issues, we propose a novel framework for pose-invariant hairstyle transfer, HairFIT. Our model consists of two stages: 1) flow-based hair alignment and 2) hair synthesis. In the hair alignment stage, we leverage a keypoint-based optical flow estimator to align a target hairstyle with a source pose. Then, we generate a final hairstyle-transferred image in the hair synthesis stage based on Semantic-region-aware Inpainting Mask (SIM) estimator. Our SIM estimator divides the occluded regions in the source image into different semantic regions to reflect their distinct features during the inpainting. To demonstrate the effectiveness of our model, we conduct quantitative and qualitative evaluations using multi-view datasets, K-hairstyle and VoxCeleb. The results indicate that HairFIT achieves a state-of-the-art performance by successfully transferring hairstyles between images of different poses, which has never been achieved before.

Sunghyun Park, Sunghyo Chung, Jungsoo Lee et al.

Despite the recent remarkable improvements in scene text recognition (STR), the majority of the studies focused mainly on the English language, which only includes few number of characters. However, STR models show a large performance degradation on languages with a numerous number of characters (e.g., Chinese and Korean), especially on characters that rarely appear due to the long-tailed distribution of characters in such languages. To address such an issue, we conducted an empirical analysis using synthetic datasets with different character-level distributions (e.g., balanced and long-tailed distributions). While increasing a substantial number of tail classes without considering the context helps the model to correctly recognize characters individually, training with such a synthetic dataset interferes the model with learning the contextual information (i.e., relation among characters), which is also important for predicting the whole word. Based on this motivation, we propose a novel Context-Aware and Free Experts Network (CAFE-Net) using two experts: 1) context-aware expert learns the contextual representation trained with a long-tailed dataset composed of common words used in everyday life and 2) context-free expert focuses on correctly predicting individual characters by utilizing a dataset with a balanced number of characters. By training two experts to focus on learning contextual and visual representations, respectively, we propose a novel confidence ensemble method to compensate the limitation of each expert. Through the experiments, we demonstrate that CAFE-Net improves the STR performance on languages containing numerous number of characters. Moreover, we show that CAFE-Net is easily applicable to various STR models.

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.

Jeongho Kim, Gyojung Gu, Minho Park et al.

Given a clothing image and a person image, an image-based virtual try-on aims to generate a customized image that appears natural and accurately reflects the characteristics of the clothing image. In this work, we aim to expand the applicability of the pre-trained diffusion model so that it can be utilized independently for the virtual try-on task.The main challenge is to preserve the clothing details while effectively utilizing the robust generative capability of the pre-trained model. In order to tackle these issues, we propose StableVITON, learning the semantic correspondence between the clothing and the human body within the latent space of the pre-trained diffusion model in an end-to-end manner. Our proposed zero cross-attention blocks not only preserve the clothing details by learning the semantic correspondence but also generate high-fidelity images by utilizing the inherent knowledge of the pre-trained model in the warping process. Through our proposed novel attention total variation loss and applying augmentation, we achieve the sharp attention map, resulting in a more precise representation of clothing details. StableVITON outperforms the baselines in qualitative and quantitative evaluation, showing promising quality in arbitrary person images. Our code is available at https://github.com/rlawjdghek/StableVITON.

Sunghyun Park, Jeongho Kim, Hyoungwoo Park et al.

Diffusion Transformers (DiTs) have significantly enhanced text-to-image (T2I) generation quality, enabling high-quality personalized content creation. However, fine-tuning these models requires substantial computational complexity and memory, limiting practical deployment under resource constraints. To tackle these challenges, we propose a memory-efficient fine-tuning framework called DiT-BlockSkip, integrating timestep-aware dynamic patch sampling and block skipping by precomputing residual features. Our dynamic patch sampling strategy adjusts patch sizes based on the diffusion timestep, then resizes the cropped patches to a fixed lower resolution. This approach reduces forward & backward memory usage while allowing the model to capture global structures at higher timesteps and fine-grained details at lower timesteps. The block skipping mechanism selectively fine-tunes essential transformer blocks and precomputes residual features for the skipped blocks, significantly reducing training memory. To identify vital blocks for personalization, we introduce a block selection strategy based on cross-attention masking. Evaluations demonstrate that our approach achieves competitive personalization performance qualitatively and quantitatively, while reducing memory usage substantially, moving toward on-device feasibility (e.g., smartphones, IoT devices) for large-scale diffusion transformers.

Saelyne Yang, Sunghyun Park, Yunseok Jang et al.

Instructional videos provide detailed how-to guides for various tasks, with viewers often posing questions regarding the content. Addressing these questions is vital for comprehending the content, yet receiving immediate answers is difficult. While numerous computational models have been developed for Video Question Answering (Video QA) tasks, they are primarily trained on questions generated based on video content, aiming to produce answers from within the content. However, in real-world situations, users may pose questions that go beyond the video's informational boundaries, highlighting the necessity to determine if a video can provide the answer. Discerning whether a question can be answered by video content is challenging due to the multi-modal nature of videos, where visual and verbal information are intertwined. To bridge this gap, we present the YTCommentQA dataset, which contains naturally-generated questions from YouTube, categorized by their answerability and required modality to answer -- visual, script, or both. Experiments with answerability classification tasks demonstrate the complexity of YTCommentQA and emphasize the need to comprehend the combined role of visual and script information in video reasoning. The dataset is available at https://github.com/lgresearch/YTCommentQA.

Jeongho Kim, Hoiyeong Jin, Sunghyun Park et al.

Recent virtual try-on approaches have advanced by finetuning pre-trained text-to-image diffusion models to leverage their powerful generative ability. However, the use of text prompts in virtual try-on remains underexplored. This paper tackles a text-editable virtual try-on task that modifies the clothing based on the provided clothing image while editing the wearing style (e.g., tucking style, fit) according to the text descriptions. In the text-editable virtual try-on, three key aspects exist: (i) designing rich text descriptions for paired person-clothing data to train the model, (ii) addressing the conflicts where textual information of the existing person's clothing interferes the generation of the new clothing, and (iii) adaptively adjust the inpainting mask aligned with the text descriptions, ensuring proper editing areas while preserving the original person's appearance irrelevant to the new clothing. To address these aspects, we propose PromptDresser, a text-editable virtual try-on model that leverages large multimodal model (LMM) assistance to enable high-quality and versatile manipulation based on generative text prompts. Our approach utilizes LMMs via in-context learning to generate detailed text descriptions for person and clothing images independently, including pose details and editing attributes using minimal human cost. Moreover, to ensure the editing areas, we adjust the inpainting mask depending on the text prompts adaptively. Our approach enhances text editability while effectively conveying clothing details that are difficult to capture through images alone, leading to improved image quality. Experiments show that PromptDresser significantly outperforms baselines, demonstrating superior text-driven control and versatile clothing manipulation. Our code is available at https://github.com/rlawjdghek/PromptDresser.

Shubhankar Borse, Seokeon Choi, Sunghyun Park et al.

Generation of images containing multiple humans, performing complex actions, while preserving their facial identities, is a significant challenge. A major factor contributing to this is the lack of a dedicated benchmark. To address this, we introduce MultiHuman-Testbench, a novel benchmark for rigorously evaluating generative models for multi-human generation. The benchmark comprises 1,800 samples, including carefully curated text prompts, describing a range of simple to complex human actions. These prompts are matched with a total of 5,550 unique human face images, sampled uniformly to ensure diversity across age, ethnic background, and gender. Alongside captions, we provide human-selected pose conditioning images which accurately match the prompt. We propose a multi-faceted evaluation suite employing four key metrics to quantify face count, ID similarity, prompt alignment, and action detection. We conduct a thorough evaluation of a diverse set of models, including zero-shot approaches and training-based methods, with and without regional priors. We also propose novel techniques to incorporate image and region isolation using human segmentation and Hungarian matching, significantly improving ID similarity. Our proposed benchmark and key findings provide valuable insights and a standardized tool for advancing research in multi-human image generation. The dataset and evaluation codes will be available at https://github.com/Qualcomm-AI-research/MultiHuman-Testbench.

Minho Park, Sunghyun Park, Jooyeol Yun et al.

Recent advancements in text-to-image generation have inspired researchers to generate datasets tailored for perception models using generative models, which prove particularly valuable in scenarios where real-world data is limited. In this study, our goal is to address the challenges when fine-tuning vision-language models (e.g., CLIP) on generated datasets. Specifically, we aim to fine-tune vision-language models to a specific classification model without access to any real images, also known as name-only transfer. However, despite the high fidelity of generated images, we observed a significant performance degradation when fine-tuning the model using the generated datasets due to the domain gap between real and generated images. To overcome the domain gap, we provide two regularization methods for training and post-training, respectively. First, we leverage the domain-agnostic knowledge from the original pre-trained vision-language model by conducting the weight-space ensemble of the fine-tuned model on the generated dataset with the original pre-trained model at the post-training. Secondly, we reveal that fine-tuned models with high feature diversity score high performance in the real domain, which indicates that increasing feature diversity prevents learning the generated domain-specific knowledge. Thus, we encourage feature diversity by providing additional regularization at training time. Extensive experiments on various classification datasets and various text-to-image generation models demonstrated that our analysis and regularization techniques effectively mitigate the domain gap, which has long been overlooked, and enable us to achieve state-of-the-art performance by training with generated images. Code is available at https://github.com/pmh9960/regft-for-gen

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.

Byungsoo Jeon, Sunghyun Park, Peiyuan Liao et al.

The strong demand for efficient and performant deployment of Deep Learning (DL) applications prompts the rapid development of a rich DL ecosystem. To keep up with this fast advancement, it is crucial for modern DL frameworks to efficiently integrate a variety of optimized tensor algebra libraries and runtimes as their backends and generate the fastest possible executable using these backends. However, current DL frameworks require significant manual effort and expertise to integrate every new backend while failing to unleash its full potential. Given the fast-evolving nature of the DL ecosystem, this manual approach often slows down continuous innovations across different layers; it prevents hardware vendors from the fast deployment of their cutting-edge libraries, DL framework developers must repeatedly adjust their hand-coded rules to accommodate new versions of libraries, and machine learning practitioners need to wait for the integration of new technologies and often encounter unsatisfactory performance. In this paper, we propose Collage, a DL framework that offers seamless integration of DL backends. Collage provides an expressive backend registration interface that allows users to precisely specify the capability of various backends. By leveraging the specifications of available backends, Collage automatically searches for an optimized backend placement strategy for a given workload and execution environment. Our evaluation shows that Collage outperforms the best existing framework for each hardware by $1.26\times$, $1.43\times$, $1.40\times$ on average on NVIDIA's RTX 2070 GPU, V100 GPU, and Intel's Xeon 8259CL CPU, respectively. Collage has been open-sourced and deployed in Apache TVM.

Jungsoo Lee, Jooyeol Yun, Sunghyun Park et al.

Despite the unprecedented improvement of face recognition, existing face recognition models still show considerably low performances in determining whether a pair of child and adult images belong to the same identity. Previous approaches mainly focused on increasing the similarity between child and adult images of a given identity to overcome the discrepancy of facial appearances due to aging. However, we observe that reducing the similarity between child images of different identities is crucial for learning distinct features among children and thus improving face recognition performance in child-adult pairs. Based on this intuition, we propose a novel loss function called the Inter-Prototype loss which minimizes the similarity between child images. Unlike the previous studies, the Inter-Prototype loss does not require additional child images or training additional learnable parameters. Our extensive experiments and in-depth analyses show that our approach outperforms existing baselines in face recognition with child-adult pairs. Our code and newly-constructed test sets of child-adult pairs are available at https://github.com/leebebeto/Inter-Prototype.

Seunghwan Choi, Sunghyun Park, Minsoo Lee et al.

The task of image-based virtual try-on aims to transfer a target clothing item onto the corresponding region of a person, which is commonly tackled by fitting the item to the desired body part and fusing the warped item with the person. While an increasing number of studies have been conducted, the resolution of synthesized images is still limited to low (e.g., 256x192), which acts as the critical limitation against satisfying online consumers. We argue that the limitation stems from several challenges: as the resolution increases, the artifacts in the misaligned areas between the warped clothes and the desired clothing regions become noticeable in the final results; the architectures used in existing methods have low performance in generating high-quality body parts and maintaining the texture sharpness of the clothes. To address the challenges, we propose a novel virtual try-on method called VITON-HD that successfully synthesizes 1024x768 virtual try-on images. Specifically, we first prepare the segmentation map to guide our virtual try-on synthesis, and then roughly fit the target clothing item to a given person's body. Next, we propose ALIgnment-Aware Segment (ALIAS) normalization and ALIAS generator to handle the misaligned areas and preserve the details of 1024x768 inputs. Through rigorous comparison with existing methods, we demonstrate that VITON-HD highly surpasses the baselines in terms of synthesized image quality both qualitatively and quantitatively. Code is available at https://github.com/shadow2496/VITON-HD.

Dongmin Kim, Sunghyun Park, Jaegul Choo

Time-series anomaly detection deals with the problem of detecting anomalous timesteps by learning normality from the sequence of observations. However, the concept of normality evolves over time, leading to a "new normal problem", where the distribution of normality can be changed due to the distribution shifts between training and test data. This paper highlights the prevalence of the new normal problem in unsupervised time-series anomaly detection studies. To tackle this issue, we propose a simple yet effective test-time adaptation strategy based on trend estimation and a self-supervised approach to learning new normalities during inference. Extensive experiments on real-world benchmarks demonstrate that incorporating the proposed strategy into the anomaly detector consistently improves the model's performance compared to the baselines, leading to robustness to the distribution shifts.

Kyungjae Lee, Dasol Hwang, Sunghyun Park et al.

Despite the promise of RLHF in aligning LLMs with human preferences, it often leads to superficial alignment, prioritizing stylistic changes over improving downstream performance of LLMs. Underspecified preferences could obscure directions to align the models. Lacking exploration restricts identification of desirable outputs to improve the models. To overcome these challenges, we propose a novel framework: Reinforcement Learning from Reflective Feedback (RLRF), which leverages fine-grained feedback based on detailed criteria to improve the core capabilities of LLMs. RLRF employs a self-reflection mechanism to systematically explore and refine LLM responses, then fine-tuning the models via a RL algorithm along with promising responses. Our experiments across Just-Eval, Factuality, and Mathematical Reasoning demonstrate the efficacy and transformative potential of RLRF beyond superficial surface-level adjustment.

Seokeon Choi, Sunghyun Park, Hyoungwoo Park et al.

Memory-efficient personalization is critical for adapting text-to-image diffusion models while preserving user privacy and operating within the limited computational resources of edge devices. To this end, we propose a selective optimization framework that adaptively chooses between backpropagation on low-resolution images (BP-low) and zeroth-order optimization on high-resolution images (ZO-high), guided by the characteristics of the diffusion process. As observed in our experiments, BP-low efficiently adapts the model to target-specific features, but suffers from structural distortions due to resolution mismatch. Conversely, ZO-high refines high-resolution details with minimal memory overhead but faces slow convergence when applied without prior adaptation. By complementing both methods, our framework leverages BP-low for effective personalization while using ZO-high to maintain structural consistency, achieving memory-efficient and high-quality fine-tuning. To maximize the efficacy of both BP-low and ZO-high, we introduce a timestep-aware probabilistic function that dynamically selects the appropriate optimization strategy based on diffusion timesteps. This function mitigates the overfitting from BP-low at high timesteps, where structural information is critical, while ensuring ZO-high is applied more effectively as training progresses. Experimental results demonstrate that our method achieves competitive performance while significantly reducing memory consumption, enabling scalable, high-quality on-device personalization without increasing inference latency.

Sunghyun Park, Seokeon Choi, Hyoungwoo Park et al.

Personalizing text-to-image diffusion models is crucial for adapting the pre-trained models to specific target concepts, enabling diverse image generation. However, fine-tuning with few images introduces an inherent trade-off between aligning with the target distribution (e.g., subject fidelity) and preserving the broad knowledge of the original model (e.g., text editability). Existing sampling guidance methods, such as classifier-free guidance (CFG) and autoguidance (AG), fail to effectively guide the output toward well-balanced space: CFG restricts the adaptation to the target distribution, while AG compromises text alignment. To address these limitations, we propose personalization guidance, a simple yet effective method leveraging an unlearned weak model conditioned on a null text prompt. Moreover, our method dynamically controls the extent of unlearning in a weak model through weight interpolation between pre-trained and fine-tuned models during inference. Unlike existing guidance methods, which depend solely on guidance scales, our method explicitly steers the outputs toward a balanced latent space without additional computational overhead. Experimental results demonstrate that our proposed guidance can improve text alignment and target distribution fidelity, integrating seamlessly with various fine-tuning strategies.

Sunghyun Park, Jungsoo Lee, Shubhankar Borse et al.

While open-vocabulary semantic segmentation (OVSS) can segment an image into semantic regions based on arbitrarily given text descriptions even for classes unseen during training, it fails to understand personal texts (e.g., `my mug cup') for segmenting regions of specific interest to users. This paper addresses challenges like recognizing `my mug cup' among `multiple mug cups'. To overcome this challenge, we introduce a novel task termed \textit{personalized open-vocabulary semantic segmentation} and propose a text prompt tuning-based plug-in method designed to recognize personal visual concepts using a few pairs of images and masks, while maintaining the performance of the original OVSS. Based on the observation that reducing false predictions is essential when applying text prompt tuning to this task, our proposed method employs `negative mask proposal' that captures visual concepts other than the personalized concept. We further improve the performance by enriching the representation of text prompts by injecting visual embeddings of the personal concept into them. This approach enhances personalized OVSS without compromising the original OVSS performance. We demonstrate the superiority of our method on our newly established benchmarks for this task, including FSS$^\text{per}$, CUB$^\text{per}$, and ADE$^\text{per}$.

Jeongho Kim, Sunghyun Park, Hyoungwoo Park et al.

Recent diffusion models achieve personalization by learning specific subjects, allowing learned attributes to be integrated into generated images. However, personalized human image generation remains challenging due to the need for precise and consistent attribute preservation (e.g., identity, clothing details). Existing subject-driven image generation methods often require either (1) inference-time fine-tuning with few images for each new subject or (2) large-scale dataset training for generalization. Both approaches are computationally expensive and impractical for real-time applications. To address these limitations, we present Wardrobe Polyptych LoRA, a novel part-level controllable model for personalized human image generation. By training only LoRA layers, our method removes the computational burden at inference while ensuring high-fidelity synthesis of unseen subjects. Our key idea is to condition the generation on the subject's wardrobe and leverage spatial references to reduce information loss, thereby improving fidelity and consistency. Additionally, we introduce a selective subject region loss, which encourages the model to disregard some of reference images during training. Our loss ensures that generated images better align with text prompts while maintaining subject integrity. Notably, our Wardrobe Polyptych LoRA requires no additional parameters at the inference stage and performs generation using a single model trained on a few training samples. We construct a new dataset and benchmark tailored for personalized human image generation. Extensive experiments show that our approach significantly outperforms existing techniques in fidelity and consistency, enabling realistic and identity-preserving full-body synthesis.

Hyungjoo Chae, Dongjin Kang, Jihyuk Kim et al. · gatech

With the release of R1, a publicly available large reasoning model (LRM), researchers commonly train new LRMs by training language models on R1's long chain-of-thought (CoT) inferences. While prior works show that LRMs' capabilities can be reproduced through direct distillation, the continued reliance on the existing models (e.g., R1) remains a critical limitation in advancing the field. As a first step toward independent LRM development, this paper explores the possibility of constructing a long CoT dataset with LLMs that are not trained for inference-time scaling. To this end, we present the Long CoT Collection, a dataset of 100K CoT rationales annotated using existing short CoT LLMs. We develop a pipeline that induces o1's novel reasoning strategies into short CoT LLMs, enabling them to think longer and introducing controllability over the thought budget to better manage the overthinking problem. Our extensive analyses validate that our dataset achieves quality comparable to--or slightly below--R1. Furthermore, our experiments demonstrate that training on our dataset not only strengthens general reasoning skills, but also provides a strong foundation for reinforcement learning--models initialized on our data achieve 2-3x larger gains with RLVR.

Minho Park, Sunghyun Park, Jungsoo Lee et al.

This paper addresses the challenge of data scarcity in semantic segmentation by generating datasets through text-to-image (T2I) generation models, reducing image acquisition and labeling costs. Segmentation dataset generation faces two key challenges: 1) aligning generated samples with the target domain and 2) producing informative samples beyond the training data. Fine-tuning T2I models can help generate samples aligned with the target domain. However, it often overfits and memorizes training data, limiting their ability to generate diverse and well-aligned samples. To overcome these issues, we propose Concept-Aware LoRA (CA-LoRA), a novel fine-tuning approach that selectively identifies and updates only the weights associated with necessary concepts (e.g., style or viewpoint) for domain alignment while preserving the pretrained knowledge of the T2I model to produce informative samples. We demonstrate its effectiveness in generating datasets for urban-scene segmentation, outperforming baseline and state-of-the-art methods in in-domain (few-shot and fully-supervised) settings, as well as in domain generalization tasks, especially under challenging conditions such as adverse weather and varying illumination, further highlighting its superiority.

Byungsoo Jeon, Mengdi Wu, Shiyi Cao et al.

Deep neural networks (DNNs) continue to grow rapidly in size, making them infeasible to train on a single device. Pipeline parallelism is commonly used in existing DNN systems to support large-scale DNN training by partitioning a DNN into multiple stages, which concurrently perform DNN training for different micro-batches in a pipeline fashion. However, existing pipeline-parallel approaches only consider sequential pipeline stages and thus ignore the topology of a DNN, resulting in missed model-parallel opportunities. This paper presents graph pipeline parallelism (GPP), a new pipeline-parallel scheme that partitions a DNN into pipeline stages whose dependencies are identified by a directed acyclic graph. GPP generalizes existing sequential pipeline parallelism and preserves the inherent topology of a DNN to enable concurrent execution of computationally-independent operators, resulting in reduced memory requirement and improved GPU performance. In addition, we develop GraphPipe, a distributed system that exploits GPP strategies to enable performant and scalable DNN training. GraphPipe partitions a DNN into a graph of stages, optimizes micro-batch schedules for these stages, and parallelizes DNN training using the discovered GPP strategies. Evaluation on a variety of DNNs shows that GraphPipe outperforms existing pipeline-parallel systems such as PipeDream and Piper by up to 1.6X. GraphPipe also reduces the search time by 9-21X compared to PipeDream and Piper.

Kangyeol Kim, Sunghyun Park, Junsoo Lee et al.

In order to perform unconditional video generation, we must learn the distribution of the real-world videos. In an effort to synthesize high-quality videos, various studies attempted to learn a mapping function between noise and videos, including recent efforts to separate motion distribution and appearance distribution. Previous methods, however, learn motion dynamics in discretized, fixed-interval timesteps, which is contrary to the continuous nature of motion of a physical body. In this paper, we propose a novel video generation approach that learns separate distributions for motion and appearance, the former modeled by neural ODE to learn natural motion dynamics. Specifically, we employ a two-stage approach where the first stage converts a noise vector to a sequence of keypoints in arbitrary frame rates, and the second stage synthesizes videos based on the given keypoints sequence and the appearance noise vector. Our model not only quantitatively outperforms recent baselines for video generation, but also demonstrates versatile functionality such as dynamic frame rate manipulation and motion transfer between two datasets, thus opening new doors to diverse video generation applications.

Taewon Kang, Sunghyun Park, Seunghwan Choi et al.

Image-based virtual try-on provides the capacity to transfer a clothing item onto a photo of a given person, which is usually accomplished by warping the item to a given human pose and adjusting the warped item to the person. However, the results of real-world synthetic images (e.g., selfies) from the previous method is not realistic because of the limitations which result in the neck being misrepresented and significant changes to the style of the garment. To address these challenges, we propose a novel method to solve this unique issue, called VITON-CROP. VITON-CROP synthesizes images more robustly when integrated with random crop augmentation compared to the existing state-of-the-art virtual try-on models. In the experiments, we demonstrate that VITON-CROP is superior to VITON-HD both qualitatively and quantitatively.

Jihun Yoon, Jiwon Lee, Sunghwan Heo et al.

Automated surgical instrument localization is an important technology to understand the surgical process and in order to analyze them to provide meaningful guidance during surgery or surgical index after surgery to the surgeon. We introduce a new dataset that reflects the kinematic characteristics of surgical instruments for automated surgical instrument localization of surgical videos. The hSDB(hutom Surgery DataBase)-instrument dataset consists of instrument localization information from 24 cases of laparoscopic cholecystecomy and 24 cases of robotic gastrectomy. Localization information for all instruments is provided in the form of a bounding box for object detection. To handle class imbalance problem between instruments, synthesized instruments modeled in Unity for 3D models are included as training data. Besides, for 3D instrument data, a polygon annotation is provided to enable instance segmentation of the tool. To reflect the kinematic characteristics of all instruments, they are annotated with head and body parts for laparoscopic instruments, and with head, wrist, and body parts for robotic instruments separately. Annotation data of assistive tools (specimen bag, needle, etc.) that are frequently used for surgery are also included. Moreover, we provide statistical information on the hSDB-instrument dataset and the baseline localization performances of the object detection networks trained by the MMDetection library and resulting analyses.

Seungbum Hong, Jiwon Lee, Bokyung Park et al.

As the area of application of deep neural networks expands to areas requiring expertise, e.g., in medicine and law, more exquisite annotation processes for expert knowledge training are required. In particular, it is difficult to guarantee generalization performance in the clinical field in the case of expert knowledge training where opinions may differ even among experts on annotations. To raise the issue of the annotation generation process for expertise training of CNNs, we verified the annotations for surgical phase recognition of laparoscopic cholecystectomy and subtotal gastrectomy for gastric cancer. We produce calibrated annotations for the seven phases of cholecystectomy by analyzing the discrepancies of previously annotated labels and by discussing the criteria of surgical phases. For gastrectomy for gastric cancer has more complex twenty-one surgical phases, we generate consensus annotation by the revision process with five specialists. By training the CNN-based surgical phase recognition networks with revised annotations, we achieved improved generalization performance over models trained with original annotation under the same cross-validation settings. We showed that the expertise data annotation pipeline for deep neural networks should be more rigorous based on the type of problem to apply clinical field.

Sangbeom Park, Yoonbyung Chai, Sunghyun Park et al.

In this paper, we present a semi-autonomous teleoperation framework for a pick-and-place task using an RGB-D sensor. In particular, we assume that the target object is located in a cluttered environment where both prehensile grasping and non-prehensile manipulation are combined for efficient teleoperation. A trajectory-based reinforcement learning is utilized for learning the non-prehensile manipulation to rearrange the objects for enabling direct grasping. From the depth image of the cluttered environment and the location of the goal object, the learned policy can provide multiple options of non-prehensile manipulation to the human operator. We carefully design a reward function for the rearranging task where the policy is trained in a simulational environment. Then, the trained policy is transferred to a real-world and evaluated in a number of real-world experiments with the varying number of objects where we show that the proposed method outperforms manual keyboard control in terms of the time duration for the grasping.

Boseop Kim, HyoungSeok Kim, Sang-Woo Lee et al.

GPT-3 shows remarkable in-context learning ability of large-scale language models (LMs) trained on hundreds of billion scale data. Here we address some remaining issues less reported by the GPT-3 paper, such as a non-English LM, the performances of different sized models, and the effect of recently introduced prompt optimization on in-context learning. To achieve this, we introduce HyperCLOVA, a Korean variant of 82B GPT-3 trained on a Korean-centric corpus of 560B tokens. Enhanced by our Korean-specific tokenization, HyperCLOVA with our training configuration shows state-of-the-art in-context zero-shot and few-shot learning performances on various downstream tasks in Korean. Also, we show the performance benefits of prompt-based learning and demonstrate how it can be integrated into the prompt engineering pipeline. Then we discuss the possibility of materializing the No Code AI paradigm by providing AI prototyping capabilities to non-experts of ML by introducing HyperCLOVA studio, an interactive prompt engineering interface. Lastly, we demonstrate the potential of our methods with three successful in-house applications.

Cheng Wang, Sungjin Lee, Sunghyun Park et al.

Real-world machine learning systems are achieving remarkable performance in terms of coarse-grained metrics like overall accuracy and F-1 score. However, model improvement and development often require fine-grained modeling on individual data subsets or slices, for instance, the data slices where the models have unsatisfactory results. In practice, it gives tangible values for developing such models that can pay extra attention to critical or interested slices while retaining the original overall performance. This work extends the recent slice-based learning (SBL)~\cite{chen2019slice} with a mixture of attentions (MoA) to learn slice-aware dual attentive representations. We empirically show that the MoA approach outperforms the baseline method as well as the original SBL approach on monitored slices with two natural language understanding (NLU) tasks.

Cheng Wang, Sun Kim, Taiwoo Park et al.

We have been witnessing the usefulness of conversational AI systems such as Siri and Alexa, directly impacting our daily lives. These systems normally rely on machine learning models evolving over time to provide quality user experience. However, the development and improvement of the models are challenging because they need to support both high (head) and low (tail) usage scenarios, requiring fine-grained modeling strategies for specific data subsets or slices. In this paper, we explore the recent concept of slice-based learning (SBL) (Chen et al., 2019) to improve our baseline conversational skill routing system on the tail yet critical query traffic. We first define a set of labeling functions to generate weak supervision data for the tail intents. We then extend the baseline model towards a slice-aware architecture, which monitors and improves the model performance on the selected tail intents. Applied to de-identified live traffic from a commercial conversational AI system, our experiments show that the slice-aware model is beneficial in improving model performance for the tail intents while maintaining the overall performance.

Han Li, Sunghyun Park, Aswarth Dara et al.

Current state-of-the-art large-scale conversational AI or intelligent digital assistant systems in industry comprises a set of components such as Automatic Speech Recognition (ASR) and Natural Language Understanding (NLU). For some of these systems that leverage a shared NLU ontology (e.g., a centralized intent/slot schema), there exists a separate skill routing component to correctly route a request to an appropriate skill, which is either a first-party or third-party application that actually executes on a user request. The skill routing component is needed as there are thousands of skills that can either subscribe to the same intent and/or subscribe to an intent under specific contextual conditions (e.g., device has a screen). Ensuring model robustness or resilience in the skill routing component is an important problem since skills may dynamically change their subscription in the ontology after the skill routing model has been deployed to production. We show how different modeling design choices impact the model robustness in the context of skill routing on a state-of-the-art commercial conversational AI system, specifically on the choices around data augmentation, model architecture, and optimization method. We show that applying data augmentation can be a very effective and practical way to drastically improve model robustness.

Taewoo Kim, Chaeyeon Chung, Sunghyun Park et al.

The hair and beauty industry is a fast-growing industry. This led to the development of various applications, such as virtual hair dyeing or hairstyle transfer, to satisfy the customer's needs. Although several hairstyle datasets are available for these applications, they often consist of a relatively small number of images with low resolution, thus limiting their performance on high-quality hair editing. In response, we introduce a novel large-scale Korean hairstyle dataset, K-hairstyle, containing 500,000 high-resolution images. In addition, K-hairstyle includes various hair attributes annotated by Korean expert hairstylists as well as hair segmentation masks. We validate the effectiveness of our dataset via several applications, such as hair dyeing, hairstyle transfer, and hairstyle classification. K-hairstyle is publicly available at https://psh01087.github.io/K-Hairstyle/.

Sunghyun Park, Han Li, Ameen Patel et al.

Natural Language Understanding (NLU) is an established component within a conversational AI or digital assistant system, and it is responsible for producing semantic understanding of a user request. We propose a scalable and automatic approach for improving NLU in a large-scale conversational AI system by leveraging implicit user feedback, with an insight that user interaction data and dialog context have rich information embedded from which user satisfaction and intention can be inferred. In particular, we propose a general domain-agnostic framework for curating new supervision data for improving NLU from live production traffic. With an extensive set of experiments, we show the results of applying the framework and improving NLU for a large-scale production system and show its impact across 10 domains.

Sunghyun Park, Kangyeol Kim, Junsoo Lee et al.

Video generation models often operate under the assumption of fixed frame rates, which leads to suboptimal performance when it comes to handling flexible frame rates (e.g., increasing the frame rate of the more dynamic portion of the video as well as handling missing video frames). To resolve the restricted nature of existing video generation models' ability to handle arbitrary timesteps, we propose continuous-time video generation by combining neural ODE (Vid-ODE) with pixel-level video processing techniques. Using ODE-ConvGRU as an encoder, a convolutional version of the recently proposed neural ODE, which enables us to learn continuous-time dynamics, Vid-ODE can learn the spatio-temporal dynamics of input videos of flexible frame rates. The decoder integrates the learned dynamics function to synthesize video frames at any given timesteps, where the pixel-level composition technique is used to maintain the sharpness of individual frames. With extensive experiments on four real-world video datasets, we verify that the proposed Vid-ODE outperforms state-of-the-art approaches under various video generation settings, both within the trained time range (interpolation) and beyond the range (extrapolation). To the best of our knowledge, Vid-ODE is the first work successfully performing continuous-time video generation using real-world videos.