Tae-Hyun Oh

Chenshuang Zhang, Kyeong Seon Kim, Chengxin Liu et al.

Despite the success of audio-visual large-language models (LLMs), they can produce plausible but ungrounded outputs, termed hallucination. Existing benchmarks focus on environmental sounds (e.g., dog barking) to indicate event occurrence. In contrast, human speech carries fundamentally different, rich semantics and temporal structures, yet it remains unexplored whether current models can accurately align speech content with corresponding visual signals. In this work, we show that speech content can induce hallucinations in audio-visual LLMs. To systematically study this, we introduce SVHalluc, the first comprehensive benchmark for evaluating speech-vision hallucination in audio-visual LLMs. Our benchmark diagnoses speech-vision hallucinations from two critical and complementary aspects: semantic and temporal. Experimental results demonstrate that state-of-the-art open-source audio-visual LLMs struggle with aligning speech content with corresponding visual signals, with a near-random accuracy on multiple tasks. In contrast, Gemini 2.5 Pro significantly outperforms the open-source models. Our analysis suggests that their failures stem from limited ability in cross-modality understanding, despite strong performance in single-modality perception. Our work uncovers a new and fundamental limitation of current audio-visual LLMs and highlights the need for speech-grounded video comprehension. Project page: https://chenshuang-zhang.github.io/projects/svhalluc/.

Baek Seong-Eun, Lee Jung-Mok, Kim Sung-Bin et al.

Fine-tuning Large Language Models (LLMs) with Low-Rank Adaptation (LoRA) offers a resource-efficient way to personalize or specialize. However, LoRA is highly sensitive to hyperparameter choices, and exhaustive hyperparameter search is computationally expensive. To address this, we propose a Bayesian Optimization (BO) framework that leverages the domain knowledge of pre-trained LLMs to efficiently search for LoRA hyperparameters. Our approach repurposes a pre-trained LLM as a discrete-to-continuous mapping module to link hyperparameters and their domain knowledge to a continuous vector space, where BO is conducted. We design and control the mapping via language prompting, providing a domain-aware textual prompt that describes the relationships among hyperparameters and their respective roles. This allows us to explicitly inject domain knowledge about LoRA into the LLM in natural language. We also introduce an additional learnable token to capture residual information that is difficult to describe linguistically in the prompt. This aids BO to sample more high-performing hyperparameters. In addition, by leveraging the strong correlation observed between the performance obtained from full and subset training datasets in LoRA training regimes, we introduce proxy training and evaluation using a data subset. This significantly improves the efficiency of our method. We demonstrate that our hyperparameter, discovered with only about 30 iterations, achieves more than 20% performance improvement over standard hyperparameters found from about 45,000 combinations. Project page: https://baekseongeun.github.io/lora-bo/

Lee Jung-Mok, Kim Sung-Bin, Joohyun Chang et al.

Laughter is a complex social signal that conveys communicative intent beyond amusement. While prior work has focused on isolated laughter analysis tasks, a comprehensive understanding of laughter in real-world scenarios remains underexplored. Therefore, we introduce SMILE-Next, a dataset for real-world laughter understanding with multimodal textual representations and question-answer annotations across three tasks: laughter detection, laughter type classification, and laughter reasoning. Building upon SMILE-Next, we aim to develop a laughter-specialized large language model capable of nuanced understanding of laughter in real-world contexts. To this end, we propose two key components: laughter-specific Self-Instruct and the Mixture-of-Laugh-Experts (MoLE) framework. Laughter-specific Self-Instruct enhances generalization across tasks and domains by automatically synthesizing diverse laughter-centric instructions. MoLE introduces a task-adaptive expert routing mechanism that dynamically selects specialized experts tailored to each laughter-related task, improving task-specific performance and efficiency. Experimental results show that the combination of our proposed components substantially outperforms multimodal LLM baselines, advancing robust real-world laughter understanding. Project page is at: https://mok0102.github.io/smile-next/.

Seongyeon Park, Myungseo Song, Bohyung Kim et al.

Neural text-to-speech (TTS) models can synthesize natural human speech when trained on large amounts of transcribed speech. However, collecting such large-scale transcribed data is expensive. This paper proposes an unsupervised pre-training method for a sequence-to-sequence TTS model by leveraging large untranscribed speech data. With our pre-training, we can remarkably reduce the amount of paired transcribed data required to train the model for the target downstream TTS task. The main idea is to pre-train the model to reconstruct de-warped mel-spectrograms from warped ones, which may allow the model to learn proper temporal assignment relation between input and output sequences. In addition, we propose a data augmentation method that further improves the data efficiency in fine-tuning. We empirically demonstrate the effectiveness of our proposed method in low-resource language scenarios, achieving outstanding performance compared to competing methods. The code and audio samples are available at: https://github.com/cnaigithub/SpeechDewarping

Junhyeong Cho, Kim Youwang, Tae-Hyun Oh

Transformer encoder architectures have recently achieved state-of-the-art results on monocular 3D human mesh reconstruction, but they require a substantial number of parameters and expensive computations. Due to the large memory overhead and slow inference speed, it is difficult to deploy such models for practical use. In this paper, we propose a novel transformer encoder-decoder architecture for 3D human mesh reconstruction from a single image, called FastMETRO. We identify the performance bottleneck in the encoder-based transformers is caused by the token design which introduces high complexity interactions among input tokens. We disentangle the interactions via an encoder-decoder architecture, which allows our model to demand much fewer parameters and shorter inference time. In addition, we impose the prior knowledge of human body's morphological relationship via attention masking and mesh upsampling operations, which leads to faster convergence with higher accuracy. Our FastMETRO improves the Pareto-front of accuracy and efficiency, and clearly outperforms image-based methods on Human3.6M and 3DPW. Furthermore, we validate its generalizability on FreiHAND.

Kim Sung-Bin, Arda Senocak, Hyunwoo Ha et al.

How does audio describe the world around us? In this paper, we propose a method for generating an image of a scene from sound. Our method addresses the challenges of dealing with the large gaps that often exist between sight and sound. We design a model that works by scheduling the learning procedure of each model component to associate audio-visual modalities despite their information gaps. The key idea is to enrich the audio features with visual information by learning to align audio to visual latent space. We translate the input audio to visual features, then use a pre-trained generator to produce an image. To further improve the quality of our generated images, we use sound source localization to select the audio-visual pairs that have strong cross-modal correlations. We obtain substantially better results on the VEGAS and VGGSound datasets than prior approaches. We also show that we can control our model's predictions by applying simple manipulations to the input waveform, or to the latent space.

Kim Youwang, Kim Ji-Yeon, Tae-Hyun Oh

We propose CLIP-Actor, a text-driven motion recommendation and neural mesh stylization system for human mesh animation. CLIP-Actor animates a 3D human mesh to conform to a text prompt by recommending a motion sequence and optimizing mesh style attributes. We build a text-driven human motion recommendation system by leveraging a large-scale human motion dataset with language labels. Given a natural language prompt, CLIP-Actor suggests a text-conforming human motion in a coarse-to-fine manner. Then, our novel zero-shot neural style optimization detailizes and texturizes the recommended mesh sequence to conform to the prompt in a temporally-consistent and pose-agnostic manner. This is distinctive in that prior work fails to generate plausible results when the pose of an artist-designed mesh does not conform to the text from the beginning. We further propose the spatio-temporal view augmentation and mask-weighted embedding attention, which stabilize the optimization process by leveraging multi-frame human motion and rejecting poorly rendered views. We demonstrate that CLIP-Actor produces plausible and human-recognizable style 3D human mesh in motion with detailed geometry and texture solely from a natural language prompt.

Kim Jun-Seong, Kim Yu-Ji, Moon Ye-Bin et al.

We propose high dynamic range (HDR) radiance fields, HDR-Plenoxels, that learn a plenoptic function of 3D HDR radiance fields, geometry information, and varying camera settings inherent in 2D low dynamic range (LDR) images. Our voxel-based volume rendering pipeline reconstructs HDR radiance fields with only multi-view LDR images taken from varying camera settings in an end-to-end manner and has a fast convergence speed. To deal with various cameras in real-world scenarios, we introduce a tone mapping module that models the digital in-camera imaging pipeline (ISP) and disentangles radiometric settings. Our tone mapping module allows us to render by controlling the radiometric settings of each novel view. Finally, we build a multi-view dataset with varying camera conditions, which fits our problem setting. Our experiments show that HDR-Plenoxels can express detail and high-quality HDR novel views from only LDR images with various cameras.

Nam Hyeon-Woo, Kim Yu-Ji, Byeongho Heo et al.

The favorable performance of Vision Transformers (ViTs) is often attributed to the multi-head self-attention (MSA). The MSA enables global interactions at each layer of a ViT model, which is a contrasting feature against Convolutional Neural Networks (CNNs) that gradually increase the range of interaction across multiple layers. We study the role of the density of the attention. Our preliminary analyses suggest that the spatial interactions of attention maps are close to dense interactions rather than sparse ones. This is a curious phenomenon, as dense attention maps are harder for the model to learn due to steeper softmax gradients around them. We interpret this as a strong preference for ViT models to include dense interaction. We thus manually insert the uniform attention to each layer of ViT models to supply the much needed dense interactions. We call this method Context Broadcasting, CB. We observe that the inclusion of CB reduces the degree of density in the original attention maps and increases both the capacity and generalizability of the ViT models. CB incurs negligible costs: 1 line in your model code, no additional parameters, and minimal extra operations.

Arda Senocak, Hyeonggon Ryu, Junsik Kim et al.

Humans can easily perceive the direction of sound sources in a visual scene, termed sound source localization. Recent studies on learning-based sound source localization have mainly explored the problem from a localization perspective. However, prior arts and existing benchmarks do not account for a more important aspect of the problem, cross-modal semantic understanding, which is essential for genuine sound source localization. Cross-modal semantic understanding is important in understanding semantically mismatched audio-visual events, e.g., silent objects, or off-screen sounds. To account for this, we propose a cross-modal alignment task as a joint task with sound source localization to better learn the interaction between audio and visual modalities. Thereby, we achieve high localization performance with strong cross-modal semantic understanding. Our method outperforms the state-of-the-art approaches in both sound source localization and cross-modal retrieval. Our work suggests that jointly tackling both tasks is necessary to conquer genuine sound source localization.

Chengxin Liu, Wonseok Choi, Chenshuang Zhang et al.

Vision-Language Models (VLMs) have demonstrated strong capability in a wide range of tasks such as visual recognition, document parsing, and visual grounding. Nevertheless, recent work shows that while VLMs often manage to capture the correct image region corresponding to the question, they do not necessarily produce the correct answers. In this work, we demonstrate that this misalignment could be attributed to suboptimal information flow within VLMs, where text tokens distribute too much attention to irrelevant visual tokens, leading to incorrect answers. Based on the observation, we show that modulating the information flow during inference can improve the perception capability of VLMs. The idea is that text tokens should only be associated with important visual tokens during decoding, eliminating the interference of irrelevant regions. To achieve this, we propose a token dynamics-based method to determine the importance of visual tokens, where visual tokens that exhibit distinct activation patterns during different decoding stages are viewed as important. We apply our approach to representative open-source VLMs and evaluate on various datasets, including visual question answering, visual grounding and counting, optical character recognition, and object hallucination. The results show that our approach significantly improves the performance of baselines. Project page: https://cxliu0.github.io/AIF/.

Minkyung Kim, Jongmin Yu, Junsik Kim et al.

Most deep anomaly detection models are based on learning normality from datasets due to the difficulty of defining abnormality by its diverse and inconsistent nature. Therefore, it has been a common practice to learn normality under the assumption that anomalous data are absent in a training dataset, which we call normality assumption. However, in practice, the normality assumption is often violated due to the nature of real data distributions that includes anomalous tails, i.e., a contaminated dataset. Thereby, the gap between the assumption and actual training data affects detrimentally in learning of an anomaly detection model. In this work, we propose a learning framework to reduce this gap and achieve better normality representation. Our key idea is to identify sample-wise normality and utilize it as an importance weight, which is updated iteratively during the training. Our framework is designed to be model-agnostic and hyperparameter insensitive so that it applies to a wide range of existing methods without careful parameter tuning. We apply our framework to three different representative approaches of deep anomaly detection that are classified into one-class classification-, probabilistic model-, and reconstruction-based approaches. In addition, we address the importance of a termination condition for iterative methods and propose a termination criterion inspired by the anomaly detection objective. We validate that our framework improves the robustness of the anomaly detection models under different levels of contamination ratios on five anomaly detection benchmark datasets and two image datasets. On various contaminated datasets, our framework improves the performance of three representative anomaly detection methods, measured by area under the ROC curve.

Moon Ye-Bin, Jisoo Kim, Hongyeob Kim et al.

We propose TextManiA, a text-driven manifold augmentation method that semantically enriches visual feature spaces, regardless of class distribution. TextManiA augments visual data with intra-class semantic perturbation by exploiting easy-to-understand visually mimetic words, i.e., attributes. This work is built on an interesting hypothesis that general language models, e.g., BERT and GPT, encompass visual information to some extent, even without training on visual training data. Given the hypothesis, TextManiA transfers pre-trained text representation obtained from a well-established large language encoder to a target visual feature space being learned. Our extensive analysis hints that the language encoder indeed encompasses visual information at least useful to augment visual representation. Our experiments demonstrate that TextManiA is particularly powerful in scarce samples with class imbalance as well as even distribution. We also show compatibility with the label mix-based approaches in evenly distributed scarce data.

Dong-Jin Kim, Tae-Hyun Oh, Jinsoo Choi et al.

We present a novel data-efficient semi-supervised framework to improve the generalization of image captioning models. Constructing a large-scale labeled image captioning dataset is an expensive task in terms of labor, time, and cost. In contrast to manually annotating all the training samples, separately collecting uni-modal datasets is immensely easier, e.g., a large-scale image dataset and a sentence dataset. We leverage such massive unpaired image and caption data upon standard paired data by learning to associate them. To this end, our proposed semi-supervised learning method assigns pseudo-labels to unpaired samples in an adversarial learning fashion, where the joint distribution of image and caption is learned. Our method trains a captioner to learn from a paired data and to progressively associate unpaired data. This approach shows noticeable performance improvement even in challenging scenarios including out-of-task data (i.e., relational captioning, where the target task is different from the unpaired data) and web-crawled data. We also show that our proposed method is theoretically well-motivated and has a favorable global optimal property. Our extensive and comprehensive empirical results both on (1) image-based and (2) dense region-based captioning datasets followed by comprehensive analysis on the scarcely-paired COCO dataset demonstrate the consistent effectiveness of our semisupervised learning method with unpaired data compared to competing methods.

Moon Ye-Bin, Dongmin Choi, Yongjin Kwon et al.

We address a weakly-supervised low-shot instance segmentation, an annotation-efficient training method to deal with novel classes effectively. Since it is an under-explored problem, we first investigate the difficulty of the problem and identify the performance bottleneck by conducting systematic analyses of model components and individual sub-tasks with a simple baseline model. Based on the analyses, we propose ENInst with sub-task enhancement methods: instance-wise mask refinement for enhancing pixel localization quality and novel classifier composition for improving classification accuracy. Our proposed method lifts the overall performance by enhancing the performance of each sub-task. We demonstrate that our ENInst is 7.5 times more efficient in achieving comparable performance to the existing fully-supervised few-shot models and even outperforms them at times.

Beichen Li, Bolei Deng, Wan Shou et al.

The conflict between stiffness and toughness is a fundamental problem in engineering materials design. However, the systematic discovery of microstructured composites with optimal stiffness-toughness trade-offs has never been demonstrated, hindered by the discrepancies between simulation and reality and the lack of data-efficient exploration of the entire Pareto front. We introduce a generalizable pipeline that integrates physical experiments, numerical simulations, and artificial neural networks to address both challenges. Without any prescribed expert knowledge of material design, our approach implements a nested-loop proposal-validation workflow to bridge the simulation-to-reality gap and discover microstructured composites that are stiff and tough with high sample efficiency. Further analysis of Pareto-optimal designs allows us to automatically identify existing toughness enhancement mechanisms, which were previously discovered through trial-and-error or biomimicry. On a broader scale, our method provides a blueprint for computational design in various research areas beyond solid mechanics, such as polymer chemistry, fluid dynamics, meteorology, and robotics.

Kwon Byung-Ki, Kim Sung-Bin, Tae-Hyun Oh

Recent work on dense optical flow has shown significant progress, primarily in a supervised learning manner requiring a large amount of labeled data. Due to the expensiveness of obtaining large scale real-world data, computer graphics are typically leveraged for constructing datasets. However, there is a common belief that synthetic-to-real domain gaps limit generalization to real scenes. In this paper, we show that the required characteristics in an optical flow dataset are rather simple and present a simpler synthetic data generation method that achieves a certain level of realism with compositions of elementary operations. With 2D motion-based datasets, we systematically analyze the simplest yet critical factors for generating synthetic datasets. Furthermore, we propose a novel method of utilizing occlusion masks in a supervised method and observe that suppressing gradients on occluded regions serves as a powerful initial state in the curriculum learning sense. The RAFT network initially trained on our dataset outperforms the original RAFT on the two most challenging online benchmarks, MPI Sintel and KITTI 2015.

Yohan Park, Hyunwoo Ha, Wonjun Jo et al.

Vision Language Models (VLMs) are increasingly adopted as central reasoning modules for embodied agents. Existing benchmarks evaluate their capabilities under ideal, well-lit conditions, yet robust 24/7 operation demands performance under a wide range of visual degradations, including low-light conditions at night or in dark environments--a core necessity that has been largely overlooked. To address this underexplored challenge, we present DarkEQA, an open-source benchmark for evaluating EQA-relevant perceptual primitives under multi-level low-light conditions. DarkEQA isolates the perception bottleneck by evaluating question answering from egocentric observations under controlled degradations, enabling attributable robustness analysis. A key design feature of DarkEQA is its physical fidelity: visual degradations are modeled in linear RAW space, simulating physics-based illumination drop and sensor noise followed by an ISP-inspired rendering pipeline. We demonstrate the utility of DarkEQA by evaluating a wide range of state-of-the-art VLMs and Low-Light Image Enhancement (LLIE) models. Our analysis systematically reveals VLMs' limitations when operating under these challenging visual conditions. Project website: https://darkeqa-benchmark.github.io/

Arda Senocak, Hyeonggon Ryu, Junsik Kim et al.

Recent studies on learning-based sound source localization have mainly focused on the localization performance perspective. However, prior work and existing benchmarks overlook a crucial aspect: cross-modal interaction, which is essential for interactive sound source localization. Cross-modal interaction is vital for understanding semantically matched or mismatched audio-visual events, such as silent objects or off-screen sounds. In this paper, we first comprehensively examine the cross-modal interaction of existing methods, benchmarks, evaluation metrics, and cross-modal understanding tasks. Then, we identify the limitations of previous studies and make several contributions to overcome the limitations. First, we introduce a new synthetic benchmark for interactive sound source localization. Second, we introduce new evaluation metrics to rigorously assess sound source localization methods, focusing on accurately evaluating both localization performance and cross-modal interaction ability. Third, we propose a learning framework with a cross-modal alignment strategy to enhance cross-modal interaction. Lastly, we evaluate both interactive sound source localization and auxiliary cross-modal retrieval tasks together to thoroughly assess cross-modal interaction capabilities and benchmark competing methods. Our new benchmarks and evaluation metrics reveal previously overlooked issues in sound source localization studies. Our proposed novel method, with enhanced cross-modal alignment, shows superior sound source localization performance. This work provides the most comprehensive analysis of sound source localization to date, with extensive validation of competing methods on both existing and new benchmarks using new and standard evaluation metrics.

Chunsan Hong, Tae-Hyun Oh, Minhyuk Sung

Diffusion models have achieved remarkable success in Text-to-Image generation tasks, leading to the development of many commercial models. However, recent studies have reported that diffusion models often generate replicated images in train data when triggered by specific prompts, potentially raising social issues ranging from copyright to privacy concerns. To sidestep the memorization, there have been recent studies for developing memorization mitigation methods for diffusion models. Nevertheless, the lack of benchmarks impedes the assessment of the true effectiveness of these methods. In this work, we present MemBench, the first benchmark for evaluating image memorization mitigation methods. Our benchmark includes a large number of memorized image trigger prompts in various Text-to-Image diffusion models. Furthermore, in contrast to the prior work evaluating mitigation performance only on trigger prompts, we present metrics evaluating on both trigger prompts and general prompts, so that we can see whether mitigation methods address the memorization issue while maintaining performance for general prompts. This is an important development considering the practical applications which previous works have overlooked. Through evaluation on MemBench, we verify that the performance of existing image memorization mitigation methods is still insufficient for application to diffusion models. The code and datasets are available at https://github.com/chunsanHong/MemBench\_code.

Kim Sung-Bin, Lee Hyun, Da Hye Hong et al.

Laughter is a unique expression, essential to affirmative social interactions of humans. Although current 3D talking head generation methods produce convincing verbal articulations, they often fail to capture the vitality and subtleties of laughter and smiles despite their importance in social context. In this paper, we introduce a novel task to generate 3D talking heads capable of both articulate speech and authentic laughter. Our newly curated dataset comprises 2D laughing videos paired with pseudo-annotated and human-validated 3D FLAME parameters and vertices. Given our proposed dataset, we present a strong baseline with a two-stage training scheme: the model first learns to talk and then acquires the ability to express laughter. Extensive experiments demonstrate that our method performs favorably compared to existing approaches in both talking head generation and expressing laughter signals. We further explore potential applications on top of our proposed method for rigging realistic avatars.

Moon Ye-Bin, Nam Hyeon-Woo, Wonseok Choi et al.

Vision language models (VLMs) perceive the world through a combination of a visual encoder and a large language model (LLM). The visual encoder, pre-trained on large-scale vision-text datasets, provides zero-shot generalization to visual data, and the LLM endows its high reasoning ability to VLMs. It leads VLMs to achieve high performance on wide benchmarks without fine-tuning, exhibiting zero or few-shot capability. However, recent studies show that VLMs are vulnerable to hallucination. This undesirable behavior degrades reliability and credibility, thereby making users unable to fully trust the output from VLMs. To enhance trustworthiness and better tackle the hallucination of VLMs, we curate a new evaluation dataset, called the BEfore-AFter hallucination dataset (BEAF), and introduce new metrics: True Understanding (TU), IGnorance (IG), StuBbornness (SB), and InDecision (ID). Unlike prior works that focus only on constructing questions and answers, the key idea of our benchmark is to manipulate visual scene information by image editing models and to design the metrics based on scene changes. This allows us to clearly assess whether VLMs correctly understand a given scene by observing the ability to perceive changes. We also visualize image-wise object relationship by virtue of our two-axis view: vision and text. Upon evaluating VLMs with our dataset, we observed that our metrics reveal different aspects of VLM hallucination that have not been reported before. Project page: \url{https://beafbench.github.io/}

Moon Ye-Bin, Nam Hyeon-Woo, Wonseok Choi et al.

Data imbalance in training data often leads to biased predictions from trained models, which in turn causes ethical and social issues. A straightforward solution is to carefully curate training data, but given the enormous scale of modern neural networks, this is prohibitively labor-intensive and thus impractical. Inspired by recent developments in generative models, this paper explores the potential of synthetic data to address the data imbalance problem. To be specific, our method, dubbed SYNAuG, leverages synthetic data to equalize the unbalanced distribution of training data. Our experiments demonstrate that, although a domain gap between real and synthetic data exists, training with SYNAuG followed by fine-tuning with a few real samples allows to achieve impressive performance on diverse tasks with different data imbalance issues, surpassing existing task-specific methods for the same purpose.

Lee Hyun, Kim Sung-Bin, Seungju Han et al. · stanford

Despite the recent advances of the artificial intelligence, building social intelligence remains a challenge. Among social signals, laughter is one of the distinctive expressions that occurs during social interactions between humans. In this work, we tackle a new challenge for machines to understand the rationale behind laughter in video, Video Laugh Reasoning. We introduce this new task to explain why people laugh in a particular video and a dataset for this task. Our proposed dataset, SMILE, comprises video clips and language descriptions of why people laugh. We propose a baseline by leveraging the reasoning capacity of large language models (LLMs) with textual video representation. Experiments show that our baseline can generate plausible explanations for laughter. We further investigate the scalability of our baseline by probing other video understanding tasks and in-the-wild videos. We release our dataset, code, and model checkpoints on https://github.com/postech-ami/SMILE-Dataset.

Kim Sung-Bin, Oh Hyun-Bin, JungMok Lee et al.

Following the success of Large Language Models (LLMs), expanding their boundaries to new modalities represents a significant paradigm shift in multimodal understanding. Human perception is inherently multimodal, relying not only on text but also on auditory and visual cues for a complete understanding of the world. In recognition of this fact, audio-visual LLMs have recently emerged. Despite promising developments, the lack of dedicated benchmarks poses challenges for understanding and evaluating models. In this work, we show that audio-visual LLMs struggle to discern subtle relationships between audio and visual signals, leading to hallucinations and highlighting the need for reliable benchmarks. To address this, we introduce AVHBench, the first comprehensive benchmark specifically designed to evaluate the perception and comprehension capabilities of audio-visual LLMs. Our benchmark includes tests for assessing hallucinations, as well as the cross-modal matching and reasoning abilities of these models. Our results reveal that most existing audio-visual LLMs struggle with hallucinations caused by cross-interactions between modalities, due to their limited capacity to perceive complex multimodal signals and their relationships. Additionally, we demonstrate that simple training with our AVHBench improves robustness of audio-visual LLMs against hallucinations. Dataset: https://github.com/kaist-ami/AVHBench

Kim Sung-Bin, Joohyun Chang, David Harwath et al.

Talking face editing and face generation have often been studied as distinct problems. In this work, we propose viewing both not as separate tasks but as subtasks of a unifying formulation, speech-conditional facial motion infilling. We explore facial motion infilling as a self-supervised pretext task that also serves as a unifying formulation of dynamic talking face synthesis. To instantiate this idea, we propose FacEDiT, a speech-conditional Diffusion Transformer trained with flow matching. Inspired by masked autoencoders, FacEDiT learns to synthesize masked facial motions conditioned on surrounding motions and speech. This formulation enables both localized generation and edits, such as substitution, insertion, and deletion, while ensuring seamless transitions with unedited regions. In addition, biased attention and temporal smoothness constraints enhance boundary continuity and lip synchronization. To address the lack of a standard editing benchmark, we introduce FacEDiTBench, the first dataset for talking face editing, featuring diverse edit types and lengths, along with new evaluation metrics. Extensive experiments validate that talking face editing and generation emerge as subtasks of speech-conditional motion infilling; FacEDiT produces accurate, speech-aligned facial edits with strong identity preservation and smooth visual continuity while generalizing effectively to talking face generation.

Han EunGi, Oh Hyun-Bin, Kim Sung-Bin et al.

Speech-driven 3D facial animation has recently garnered attention due to its cost-effective usability in multimedia production. However, most current advances overlook the intelligibility of lip movements, limiting the realism of facial expressions. In this paper, we introduce a method for speech-driven 3D facial animation to generate accurate lip movements, proposing an audio-visual multimodal perceptual loss. This loss provides guidance to train the speech-driven 3D facial animators to generate plausible lip motions aligned with the spoken transcripts. Furthermore, to incorporate the proposed audio-visual perceptual loss, we devise an audio-visual lip reading expert leveraging its prior knowledge about correlations between speech and lip motions. We validate the effectiveness of our approach through broad experiments, showing noticeable improvements in lip synchronization and lip readability performance. Codes are available at https://3d-talking-head-avguide.github.io/.

Oh Hyun-Bin, Yuhta Takida, Toshimitsu Uesaka et al.

Recent advances in Video-to-Audio (V2A) generation have achieved impressive perceptual quality and temporal synchronization, yet most models remain appearance-driven, capturing visual-acoustic correlations without considering the physical factors that shape real-world sounds. We present Physics-Aware Video-to-Audio Synthesis (PAVAS), a method that incorporates physical reasoning into a latent diffusion-based V2A generation through the Physics-Driven Audio Adapter (Phy-Adapter). The adapter receives object-level physical parameters estimated by the Physical Parameter Estimator (PPE), which uses a Vision-Language Model (VLM) to infer the moving-object mass and a segmentation-based dynamic 3D reconstruction module to recover its motion trajectory for velocity computation. These physical cues enable the model to synthesize sounds that reflect underlying physical factors. To assess physical realism, we curate VGG-Impact, a benchmark focusing on object-object interactions, and introduce Audio-Physics Correlation Coefficient (APCC), an evaluation metric that measures consistency between physical and auditory attributes. Comprehensive experiments show that PAVAS produces physically plausible and perceptually coherent audio, outperforming existing V2A models in both quantitative and qualitative evaluations. Visit https://physics-aware-video-to-audio-synthesis.github.io for demo videos.

Nam Hyeon-Woo, Moon Ye-Bin, Wonseok Choi et al.

Vision language models (VLMs) have shown promising reasoning capabilities across various benchmarks; however, our understanding of their visual perception remains limited. In this work, we propose an eye examination process to investigate how a VLM perceives images, specifically focusing on key elements of visual recognition, from primitive color and shape to semantic levels. To this end, we introduce a dataset named LENS to guide a VLM to follow the examination and check its readiness. Once the model is ready, we conduct the examination. Through this examination, we quantify and visualize VLMs' sensitivities to color and shape, and semantic matching. Our findings reveal that VLMs have varying sensitivity to different colors while consistently showing insensitivity to green across different VLMs. Also, we found different shape sensitivity and semantic recognition depending on LLM's capacity despite using the same fixed visual encoder. Our analyses and findings have potential to inspire the design of VLMs and the pre-processing of visual input to VLMs for improving application performance.

Kim Youwang, Lee Hyun, Kim Sung-Bin et al.

We propose NeuFace, a 3D face mesh pseudo annotation method on videos via neural re-parameterized optimization. Despite the huge progress in 3D face reconstruction methods, generating reliable 3D face labels for in-the-wild dynamic videos remains challenging. Using NeuFace optimization, we annotate the per-view/-frame accurate and consistent face meshes on large-scale face videos, called the NeuFace-dataset. We investigate how neural re-parameterization helps to reconstruct image-aligned facial details on 3D meshes via gradient analysis. By exploiting the naturalness and diversity of 3D faces in our dataset, we demonstrate the usefulness of our dataset for 3D face-related tasks: improving the reconstruction accuracy of an existing 3D face reconstruction model and learning 3D facial motion prior. Code and datasets will be available at https://neuface-dataset.github.io.

Kyeong Seon Kim, Baek Seong-Eun, Lee Jung-Mok et al.

Scalable Vector Graphics (SVGs) function both as visual images and as structured code that encode rich geometric and layout information, yet most methods rasterize them and discard this symbolic organization. At the same time, recent sentence embedding methods produce strong text representations but do not naturally extend to visual or structured modalities. We propose a training-free, instruction-guided multimodal embedding framework that uses a Multimodal Large Language Model (MLLM) to map text, raster images, and SVG code into an aligned embedding space. We control the direction of embeddings through modality-specific instructions and structural SVG cues, eliminating the need for learned projection heads or contrastive training. Our method has two key components: (1) Multimodal Explicit One-word Limitation (mEOL), which instructs the MLLM to summarize any multimodal input into a single token whose hidden state serves as a compact semantic embedding. (2) A semantic SVG rewriting module that assigns meaningful identifiers and simplifies nested SVG elements through visual reasoning over the rendered image, exposing geometric and relational cues hidden in raw code. Using a repurposed VGBench, we build the first text-to-SVG retrieval benchmark and show that our training-free embeddings outperform encoder-based and training-based multimodal baselines. These results highlight prompt-level control as an effective alternative to parameter-level training for structure-aware multimodal retrieval. Project page: https://scene-the-ella.github.io/meol/

Kim Yu-Ji, Hyunwoo Ha, Kim Youwang et al.

Reconstructing 3D from a single view image is a long-standing challenge. One of the popular approaches to tackle this problem is learning-based methods, but dealing with the test cases unfamiliar with training data (Out-of-distribution; OoD) introduces an additional challenge. To adapt for unseen samples in test time, we propose MeTTA, a test-time adaptation (TTA) exploiting generative prior. We design joint optimization of 3D geometry, appearance, and pose to handle OoD cases with only a single view image. However, the alignment between the reference image and the 3D shape via the estimated viewpoint could be erroneous, which leads to ambiguity. To address this ambiguity, we carefully design learnable virtual cameras and their self-calibration. In our experiments, we demonstrate that MeTTA effectively deals with OoD scenarios at failure cases of existing learning-based 3D reconstruction models and enables obtaining a realistic appearance with physically based rendering (PBR) textures.

Kwon Byung-Ki, Sohwi Lim, Nam Hyeon-Woo et al.

Text-to-video (T2V) diffusion models have rapidly advanced, yet generations still occasionally fail in practice, such as low text-video alignment or low perceptual quality. Since diffusion sampling is non-deterministic, it is difficult to know during inference whether a generation will succeed or fail, incurring high computational cost due to trial-and-error regeneration. To address this, we propose an early failure detection and diagnostic intervention pipeline for latent T2V diffusion models. For detection, we design a Real-time Inspection (RI) module that converts latents into intermediate video previews, enabling the use of established text-video alignment scorers for inspection in the RGB space. The RI module completes the conversion and inspection process in just 39.2ms. This is highly efficient considering that CogVideoX-5B requires 4.3s per denoising step when generating a 480p, 49-frame video on an NVIDIA A100 GPU. Subsequently, we trigger a hierarchical and early-exit intervention pipeline only when failure is predicted. Experiments on CogVideoX-5B and Wan2.1-1.3B demonstrate consistency gains on VBench with up to 2.64 times less time overhead compared to post-hoc regeneration. Our method also generalizes to a higher-capacity setting, remaining effective on Wan2.1-14B with 720p resolution and 81-frame generation. Furthermore, our pipeline is plug-and-play and orthogonal to existing techniques, showing seamless compatibility with prompt refinement and sampling guidance methods. We also provide evidence that failure signals emerge early in the denoising process and are detectable within intermediate video previews using standard vision-language evaluators.

Wonjun Jo, Hyunwoo Ha, Kim Ji-Yeon et al.

As deep learning continues to advance, self-supervised learning has made considerable strides. It allows 2D image encoders to extract useful features for various downstream tasks, including those related to vision-based systems. Nevertheless, pre-trained 2D image encoders fall short in conducting the task under noisy and adverse weather conditions beyond clear daytime scenes, which require for robust visual perception. To address these issues, we propose a novel self-supervised approach, \textbf{Collaborative Distillation}, which leverages 3D LiDAR as self-supervision to improve robustness to noisy and adverse weather conditions in 2D image encoders while retaining their original capabilities. Our method outperforms competing methods in various downstream tasks across diverse conditions and exhibits strong generalization ability. In addition, our method also improves 3D awareness stemming from LiDAR's characteristics. This advancement highlights our method's practicality and adaptability in real-world scenarios.

Kim Youwang, Lee Hyoseok, Subin Park et al.

We introduce ELITE, an Efficient Gaussian head avatar synthesis from a monocular video via Learned Initialization and TEst-time generative adaptation. Prior works rely either on a 3D data prior or a 2D generative prior to compensate for missing visual cues in monocular videos. However, 3D data prior methods often struggle to generalize in-the-wild, while 2D generative prior methods are computationally heavy and prone to identity hallucination. We identify a complementary synergy between these two priors and design an efficient system that achieves high-fidelity animatable avatar synthesis with strong in-the-wild generalization. Specifically, we introduce a feed-forward Mesh2Gaussian Prior Model (MGPM) that enables fast initialization of a Gaussian avatar. To further bridge the domain gap at test time, we design a test-time generative adaptation stage, leveraging both real and synthetic images as supervision. Unlike previous full diffusion denoising strategies that are slow and hallucination-prone, we propose a rendering-guided single-step diffusion enhancer that restores missing visual details, grounded on Gaussian avatar renderings. Our experiments demonstrate that ELITE produces visually superior avatars to prior works, even for challenging expressions, while achieving 60x faster synthesis than the 2D generative prior method.

Kim Jun-Seong, Mingyu Kim, GeonU Kim et al.

We introduce Fact-Hash, a novel parameter-encoding method for training on-device neural radiance fields. Neural Radiance Fields (NeRF) have proven pivotal in 3D representations, but their applications are limited due to large computational resources. On-device training can open large application fields, providing strength in communication limitations, privacy concerns, and fast adaptation to a frequently changing scene. However, challenges such as limited resources (GPU memory, storage, and power) impede their deployment. To handle this, we introduce Fact-Hash, a novel parameter-encoding merging Tensor Factorization and Hash-encoding techniques. This integration offers two benefits: the use of rich high-resolution features and the few-shot robustness. In Fact-Hash, we project 3D coordinates into multiple lower-dimensional forms (2D or 1D) before applying the hash function and then aggregate them into a single feature. Comparative evaluations against state-of-the-art methods demonstrate Fact-Hash's superior memory efficiency, preserving quality and rendering speed. Fact-Hash saves memory usage by over one-third while maintaining the PSNR values compared to previous encoding methods. The on-device experiment validates the superiority of Fact-Hash compared to alternative positional encoding methods in computational efficiency and energy consumption. These findings highlight Fact-Hash as a promising solution to improve feature grid representation, address memory constraints, and improve quality in various applications. Project page: https://facthash.github.io/

Sohwi Lim, Lee Hyoseok, Jungjoon Park et al.

Human perception of visual similarity is inherently adaptive and subjective, depending on the users' interests and focus. However, most image retrieval systems fail to reflect this flexibility, relying on a fixed, monolithic metric that cannot incorporate multiple conditions simultaneously. To address this, we propose CLAY, an adaptive similarity computation method that reframes the embedding space of pretrained Vision-Language Models (VLMs) as a text-conditional similarity space without additional training. This design separates the textual conditioning process and visual feature extraction, allowing highly efficient and multi-conditioned retrieval with fixed visual embeddings. We also construct a synthetic evaluation dataset CLAY-EVAL, for comprehensive assessment under diverse conditioned retrieval settings. Experiments on standard datasets and our proposed dataset show that CLAY achieves high retrieval accuracy and notable computational efficiency compared to previous works.

Kim Youwang, Tae-Hyun Oh, Gerard Pons-Moll

We present Paint-it, a text-driven high-fidelity texture map synthesis method for 3D meshes via neural re-parameterized texture optimization. Paint-it synthesizes texture maps from a text description by synthesis-through-optimization, exploiting the Score-Distillation Sampling (SDS). We observe that directly applying SDS yields undesirable texture quality due to its noisy gradients. We reveal the importance of texture parameterization when using SDS. Specifically, we propose Deep Convolutional Physically-Based Rendering (DC-PBR) parameterization, which re-parameterizes the physically-based rendering (PBR) texture maps with randomly initialized convolution-based neural kernels, instead of a standard pixel-based parameterization. We show that DC-PBR inherently schedules the optimization curriculum according to texture frequency and naturally filters out the noisy signals from SDS. In experiments, Paint-it obtains remarkable quality PBR texture maps within 15 min., given only a text description. We demonstrate the generalizability and practicality of Paint-it by synthesizing high-quality texture maps for large-scale mesh datasets and showing test-time applications such as relighting and material control using a popular graphics engine. Project page: https://kim-youwang.github.io/paint-it

Hansam Cho, Jonghyun Lee, Seoung Bum Kim et al.

Text-guided diffusion models have become a popular tool in image synthesis, known for producing high-quality and diverse images. However, their application to editing real images often encounters hurdles primarily due to the text condition deteriorating the reconstruction quality and subsequently affecting editing fidelity. Null-text Inversion (NTI) has made strides in this area, but it fails to capture spatial context and requires computationally intensive per-timestep optimization. Addressing these challenges, we present Noise Map Guidance (NMG), an inversion method rich in a spatial context, tailored for real-image editing. Significantly, NMG achieves this without necessitating optimization, yet preserves the editing quality. Our empirical investigations highlight NMG's adaptability across various editing techniques and its robustness to variants of DDIM inversions.

Kim Jun-Seong, GeonU Kim, Kim Yu-Ji et al.

We introduce Dr. Splat, a novel approach for open-vocabulary 3D scene understanding leveraging 3D Gaussian Splatting. Unlike existing language-embedded 3DGS methods, which rely on a rendering process, our method directly associates language-aligned CLIP embeddings with 3D Gaussians for holistic 3D scene understanding. The key of our method is a language feature registration technique where CLIP embeddings are assigned to the dominant Gaussians intersected by each pixel-ray. Moreover, we integrate Product Quantization (PQ) trained on general large-scale image data to compactly represent embeddings without per-scene optimization. Experiments demonstrate that our approach significantly outperforms existing approaches in 3D perception benchmarks, such as open-vocabulary 3D semantic segmentation, 3D object localization, and 3D object selection tasks. For video results, please visit : https://drsplat.github.io/

GeonU Kim, Kim Youwang, Tae-Hyun Oh

We present FPRF, a feed-forward photorealistic style transfer method for large-scale 3D neural radiance fields. FPRF stylizes large-scale 3D scenes with arbitrary, multiple style reference images without additional optimization while preserving multi-view appearance consistency. Prior arts required tedious per-style/-scene optimization and were limited to small-scale 3D scenes. FPRF efficiently stylizes large-scale 3D scenes by introducing a style-decomposed 3D neural radiance field, which inherits AdaIN's feed-forward stylization machinery, supporting arbitrary style reference images. Furthermore, FPRF supports multi-reference stylization with the semantic correspondence matching and local AdaIN, which adds diverse user control for 3D scene styles. FPRF also preserves multi-view consistency by applying semantic matching and style transfer processes directly onto queried features in 3D space. In experiments, we demonstrate that FPRF achieves favorable photorealistic quality 3D scene stylization for large-scale scenes with diverse reference images. Project page: https://kim-geonu.github.io/FPRF/

Lee Hyoseok, Sohwi Lim, Eunju Cha et al.

While latent diffusion models (LDMs) have emerged as powerful priors for inverse problems, existing LDM-based solvers frequently suffer from instability. In this work, we first identify the instability as a discrepancy between the solver dynamics and stable reverse diffusion dynamics learned by the diffusion model, and show that reducing this gap stabilizes the solver. Building on this, we introduce \textit{Measurement-Consistent Langevin Corrector (MCLC)}, a theoretically grounded plug-and-play stabilization module that remedies the LDM-based inverse problem solvers through measurement-consistent Langevin updates. Compared to prior approaches that rely on linear manifold assumptions, which often fail to hold in latent space, MCLC provides a principled stabilization mechanism, leading to more stable and reliable behavior in latent space.

Lee Chae-Yeon, Oh Hyun-Bin, Han EunGi et al.

Recent advancements in speech-driven 3D talking head generation have made significant progress in lip synchronization. However, existing models still struggle to capture the perceptual alignment between varying speech characteristics and corresponding lip movements. In this work, we claim that three criteria -- Temporal Synchronization, Lip Readability, and Expressiveness -- are crucial for achieving perceptually accurate lip movements. Motivated by our hypothesis that a desirable representation space exists to meet these three criteria, we introduce a speech-mesh synchronized representation that captures intricate correspondences between speech signals and 3D face meshes. We found that our learned representation exhibits desirable characteristics, and we plug it into existing models as a perceptual loss to better align lip movements to the given speech. In addition, we utilize this representation as a perceptual metric and introduce two other physically grounded lip synchronization metrics to assess how well the generated 3D talking heads align with these three criteria. Experiments show that training 3D talking head generation models with our perceptual loss significantly improve all three aspects of perceptually accurate lip synchronization. Codes and datasets are available at https://perceptual-3d-talking-head.github.io/.

Jungbin Cho, Junwan Kim, Jisoo Kim et al.

Human motion is inherently continuous and dynamic, posing significant challenges for generative models. While discrete generation methods are widely used, they suffer from limited expressiveness and frame-wise noise artifacts. In contrast, continuous approaches produce smoother, more natural motion but often struggle to adhere to conditioning signals due to high-dimensional complexity and limited training data. To resolve this 'discord' between discrete and continuous representations we introduce DisCoRD: Discrete Tokens to Continuous Motion via Rectified Flow Decoding, a novel method that leverages rectified flow to decode discrete motion tokens in the continuous, raw motion space. Our core idea is to frame token decoding as a conditional generation task, ensuring that DisCoRD captures fine-grained dynamics and achieves smoother, more natural motions. Compatible with any discrete-based framework, our method enhances naturalness without compromising faithfulness to the conditioning signals on diverse settings. Extensive evaluations demonstrate that DisCoRD achieves state-of-the-art performance, with FID of 0.032 on HumanML3D and 0.169 on KIT-ML. These results establish DisCoRD as a robust solution for bridging the divide between discrete efficiency and continuous realism. Project website: https://whwjdqls.github.io/discord-motion/

Junhyeong Cho, Kim Youwang, Hunmin Yang et al.

Recent monocular 3D shape reconstruction methods have shown promising zero-shot results on object-segmented images without any occlusions. However, their effectiveness is significantly compromised in real-world conditions, due to imperfect object segmentation by off-the-shelf models and the prevalence of occlusions. To effectively address these issues, we propose a unified regression model that integrates segmentation and reconstruction, specifically designed for occlusion-aware 3D shape reconstruction. To facilitate its reconstruction in the wild, we also introduce a scalable data synthesis pipeline that simulates a wide range of variations in objects, occluders, and backgrounds. Training on our synthetic data enables the proposed model to achieve state-of-the-art zero-shot results on real-world images, using significantly fewer parameters than competing approaches.

Lee Hyoseok, Kyeong Seon Kim, Kwon Byung-Ki et al.

Depth completion, predicting dense depth maps from sparse depth measurements, is an ill-posed problem requiring prior knowledge. Recent methods adopt learning-based approaches to implicitly capture priors, but the priors primarily fit in-domain data and do not generalize well to out-of-domain scenarios. To address this, we propose a zero-shot depth completion method composed of an affine-invariant depth diffusion model and test-time alignment. We use pre-trained depth diffusion models as depth prior knowledge, which implicitly understand how to fill in depth for scenes. Our approach aligns the affine-invariant depth prior with metric-scale sparse measurements, enforcing them as hard constraints via an optimization loop at test-time. Our zero-shot depth completion method demonstrates generalization across various domain datasets, achieving up to a 21\% average performance improvement over the previous state-of-the-art methods while enhancing spatial understanding by sharpening scene details. We demonstrate that aligning a monocular affine-invariant depth prior with sparse metric measurements is a proven strategy to achieve domain-generalizable depth completion without relying on extensive training data. Project page: https://hyoseok1223.github.io/zero-shot-depth-completion/.

Kim Sung-Bin, Arda Senocak, Hyunwoo Ha et al.

How does audio describe the world around us? In this work, we propose a method for generating images of visual scenes from diverse in-the-wild sounds. This cross-modal generation task is challenging due to the significant information gap between auditory and visual signals. We address this challenge by designing a model that aligns audio-visual modalities by enriching audio features with visual information and translating them into the visual latent space. These features are then fed into the pre-trained image generator to produce images. To enhance image quality, we use sound source localization to select audio-visual pairs with strong cross-modal correlations. Our method achieves substantially better results on the VEGAS and VGGSound datasets compared to previous work and demonstrates control over the generation process through simple manipulations to the input waveform or latent space. Furthermore, we analyze the geometric properties of the learned embedding space and demonstrate that our learning approach effectively aligns audio-visual signals for cross-modal generation. Based on this analysis, we show that our method is agnostic to specific design choices, showing its generalizability by integrating various model architectures and different types of audio-visual data.

Hyunwoo Ha, Oh Hyun-Bin, Kim Jun-Seong et al.

Video motion magnification is a technique to capture and amplify subtle motion in a video that is invisible to the naked eye. The deep learning-based prior work successfully demonstrates the modelling of the motion magnification problem with outstanding quality compared to conventional signal processing-based ones. However, it still lags behind real-time performance, which prevents it from being extended to various online applications. In this paper, we investigate an efficient deep learning-based motion magnification model that runs in real time for full-HD resolution videos. Due to the specified network design of the prior art, i.e. inhomogeneous architecture, the direct application of existing neural architecture search methods is complicated. Instead of automatic search, we carefully investigate the architecture module by module for its role and importance in the motion magnification task. Two key findings are 1) Reducing the spatial resolution of the latent motion representation in the decoder provides a good trade-off between computational efficiency and task quality, and 2) surprisingly, only a single linear layer and a single branch in the encoder are sufficient for the motion magnification task. Based on these findings, we introduce a real-time deep learning-based motion magnification model with4.2X fewer FLOPs and is 2.7X faster than the prior art while maintaining comparable quality.

Kim Sung-Bin, Jeongsoo Choi, Puyuan Peng et al.

We present VoiceCraft-Dub, a novel approach for automated video dubbing that synthesizes high-quality speech from text and facial cues. This task has broad applications in filmmaking, multimedia creation, and assisting voice-impaired individuals. Building on the success of Neural Codec Language Models (NCLMs) for speech synthesis, our method extends their capabilities by incorporating video features, ensuring that synthesized speech is time-synchronized and expressively aligned with facial movements while preserving natural prosody. To inject visual cues, we design adapters to align facial features with the NCLM token space and introduce audio-visual fusion layers to merge audio-visual information within the NCLM framework. Additionally, we curate CelebV-Dub, a new dataset of expressive, real-world videos specifically designed for automated video dubbing. Extensive experiments show that our model achieves high-quality, intelligible, and natural speech synthesis with accurate lip synchronization, outperforming existing methods in human perception and performing favorably in objective evaluations. We also adapt VoiceCraft-Dub for the video-to-speech task, demonstrating its versatility for various applications.

Kim Yu-Ji, Dahye Lee, Kim Jun-Seong et al.

We present GaussExplorer, a framework for embodied exploration and reasoning built on 3D Gaussian Splatting (3DGS). While prior approaches to language-embedded 3DGS have made meaningful progress in aligning simple text queries with Gaussian embeddings, they are generally optimized for relatively simple queries and struggle to interpret more complex, compositional language queries. Alternative studies based on object-centric RGB-D structured memories provide spatial grounding but are constrained by pre-fixed viewpoints. To address these issues, GaussExplorer introduces Vision-Language Models (VLMs) on top of 3DGS to enable question-driven exploration and reasoning within 3D scenes. We first identify pre-captured images that are most correlated with the query question, and subsequently adjust them into novel viewpoints to more accurately capture visual information for better reasoning by VLMs. Experiments show that ours outperforms existing methods on several benchmarks, demonstrating the effectiveness of integrating VLM-based reasoning with 3DGS for embodied tasks.