Jae-Pil Heo

WonJun Moon, Sangeek Hyun, SangUk Park et al.

Recently, video moment retrieval and highlight detection (MR/HD) are being spotlighted as the demand for video understanding is drastically increased. The key objective of MR/HD is to localize the moment and estimate clip-wise accordance level, i.e., saliency score, to the given text query. Although the recent transformer-based models brought some advances, we found that these methods do not fully exploit the information of a given query. For example, the relevance between text query and video contents is sometimes neglected when predicting the moment and its saliency. To tackle this issue, we introduce Query-Dependent DETR (QD-DETR), a detection transformer tailored for MR/HD. As we observe the insignificant role of a given query in transformer architectures, our encoding module starts with cross-attention layers to explicitly inject the context of text query into video representation. Then, to enhance the model's capability of exploiting the query information, we manipulate the video-query pairs to produce irrelevant pairs. Such negative (irrelevant) video-query pairs are trained to yield low saliency scores, which in turn, encourages the model to estimate precise accordance between query-video pairs. Lastly, we present an input-adaptive saliency predictor which adaptively defines the criterion of saliency scores for the given video-query pairs. Our extensive studies verify the importance of building the query-dependent representation for MR/HD. Specifically, QD-DETR outperforms state-of-the-art methods on QVHighlights, TVSum, and Charades-STA datasets. Codes are available at github.com/wjun0830/QD-DETR.

Hyun Seok Seong, WonJun Moon, SuBeen Lee et al.

Dramatic demand for manpower to label pixel-level annotations triggered the advent of unsupervised semantic segmentation. Although the recent work employing the vision transformer (ViT) backbone shows exceptional performance, there is still a lack of consideration for task-specific training guidance and local semantic consistency. To tackle these issues, we leverage contrastive learning by excavating hidden positives to learn rich semantic relationships and ensure semantic consistency in local regions. Specifically, we first discover two types of global hidden positives, task-agnostic and task-specific ones for each anchor based on the feature similarities defined by a fixed pre-trained backbone and a segmentation head-in-training, respectively. A gradual increase in the contribution of the latter induces the model to capture task-specific semantic features. In addition, we introduce a gradient propagation strategy to learn semantic consistency between adjacent patches, under the inherent premise that nearby patches are highly likely to possess the same semantics. Specifically, we add the loss propagating to local hidden positives, semantically similar nearby patches, in proportion to the predefined similarity scores. With these training schemes, our proposed method achieves new state-of-the-art (SOTA) results in COCO-stuff, Cityscapes, and Potsdam-3 datasets. Our code is available at: https://github.com/hynnsk/HP.

WonJun Moon, Junho Park, Hyun Seok Seong et al.

Open set recognition (OSR) assumes unknown instances appear out of the blue at the inference time. The main challenge of OSR is that the response of models for unknowns is totally unpredictable. Furthermore, the diversity of open set makes it harder since instances have different difficulty levels. Therefore, we present a novel framework, DIfficulty-Aware Simulator (DIAS), that generates fakes with diverse difficulty levels to simulate the real world. We first investigate fakes from generative adversarial network (GAN) in the classifier's viewpoint and observe that these are not severely challenging. This leads us to define the criteria for difficulty by regarding samples generated with GANs having moderate-difficulty. To produce hard-difficulty examples, we introduce Copycat, imitating the behavior of the classifier. Furthermore, moderate- and easy-difficulty samples are also yielded by our modified GAN and Copycat, respectively. As a result, DIAS outperforms state-of-the-art methods with both metrics of AUROC and F-score. Our code is available at https://github.com/wjun0830/Difficulty-Aware-Simulator.

WonJun Moon, Ji-Hwan Kim, Jae-Pil Heo

Recently, it is shown that deploying a proper self-supervision is a prospective way to enhance the performance of supervised learning. Yet, the benefits of self-supervision are not fully exploited as previous pretext tasks are specialized for unsupervised representation learning. To this end, we begin by presenting three desirable properties for such auxiliary tasks to assist the supervised objective. First, the tasks need to guide the model to learn rich features. Second, the transformations involved in the self-supervision should not significantly alter the training distribution. Third, the tasks are preferred to be light and generic for high applicability to prior arts. Subsequently, to show how existing pretext tasks can fulfill these and be tailored for supervised learning, we propose a simple auxiliary self-supervision task, predicting localizable rotation (LoRot). Our exhaustive experiments validate the merits of LoRot as a pretext task tailored for supervised learning in terms of robustness and generalization capability. Our code is available at https://github.com/wjun0830/Localizable-Rotation.

WonJun Moon, Sangeek Hyun, SuBeen Lee et al.

Temporal Grounding is to identify specific moments or highlights from a video corresponding to textual descriptions. Typical approaches in temporal grounding treat all video clips equally during the encoding process regardless of their semantic relevance with the text query. Therefore, we propose Correlation-Guided DEtection TRansformer (CG-DETR), exploring to provide clues for query-associated video clips within the cross-modal attention. First, we design an adaptive cross-attention with dummy tokens. Dummy tokens conditioned by text query take portions of the attention weights, preventing irrelevant video clips from being represented by the text query. Yet, not all words equally inherit the text query's correlation to video clips. Thus, we further guide the cross-attention map by inferring the fine-grained correlation between video clips and words. We enable this by learning a joint embedding space for high-level concepts, i.e., moment and sentence level, and inferring the clip-word correlation. Lastly, we exploit the moment-specific characteristics and combine them with the context of each video to form a moment-adaptive saliency detector. By exploiting the degrees of text engagement in each video clip, it precisely measures the highlightness of each clip. CG-DETR achieves state-of-the-art results on various benchmarks for temporal grounding. Codes are available at https://github.com/wjun0830/CGDETR.

Kangmin Seo, MinKyu Lee, Tae-Young Kim et al.

Recent advances in 3D Gaussian Splatting (3DGS) have enabled impressive real-time photorealistic rendering. However, conventional training pipelines inherently assume full multi-view consistency among input images, which makes them sensitive to distractors that violate this assumption and cause visual artifacts. In this work, we revisit an underexplored aspect of 3DGS: its inherent ability to suppress inconsistent signals. Building on this insight, we propose PDF-GS (Progressive Distractor Filtering for Robust 3D Gaussian Splatting), a framework that amplifies this self-filtering property through a progressive multi-phase optimization. The progressive filtering phases gradually remove distractors by exploiting discrepancy cues, while the following reconstruction phase restores fine-grained, view-consistent details from the purified Gaussian representation. Through this iterative refinement, PDF-GS achieves robust, high-fidelity, and distractor-free reconstructions, consistently outperforming baselines across diverse datasets and challenging real-world conditions. Moreover, our approach is lightweight and easily adaptable to existing 3DGS frameworks, requiring no architectural changes or additional inference overhead, leading to a new state-of-the-art performance. The code is publicly available at https://github.com/kangrnin/PDF-GS.

SuBeen Lee, WonJun Moon, Jae-Pil Heo

Recognizing discriminative details such as eyes and beaks is important for distinguishing fine-grained classes since they have similar overall appearances. In this regard, we introduce Task Discrepancy Maximization (TDM), a simple module for fine-grained few-shot classification. Our objective is to localize the class-wise discriminative regions by highlighting channels encoding distinct information of the class. Specifically, TDM learns task-specific channel weights based on two novel components: Support Attention Module (SAM) and Query Attention Module (QAM). SAM produces a support weight to represent channel-wise discriminative power for each class. Still, since the SAM is basically only based on the labeled support sets, it can be vulnerable to bias toward such support set. Therefore, we propose QAM which complements SAM by yielding a query weight that grants more weight to object-relevant channels for a given query image. By combining these two weights, a class-wise task-specific channel weight is defined. The weights are then applied to produce task-adaptive feature maps more focusing on the discriminative details. Our experiments validate the effectiveness of TDM and its complementary benefits with prior methods in fine-grained few-shot classification.

Jaejoon Yoo, SuBeen Lee, Yerim Jeon et al.

3D Single Object Tracking (3D-SOT) aims to localize a target object across a sequence of LiDAR point clouds, given its 3D bounding box in the first frame. Recent methods have adopted a memory-based approach to utilize previously observed features of the target object, but remain limited to only a few recent frames. This work reveals that their temporal capacity is fundamentally constrained to short-term context due to severe temporal feature inconsistency and excessive memory overhead. To this end, we propose a robust long-term 3D-SOT framework, ChronoTrack, which preserves the temporal feature consistency while efficiently aggregating the diverse target features via long-term memory. Based on a compact set of learnable memory tokens, ChronoTrack leverages long-term information through two complementary objectives: a temporal consistency loss and a memory cycle consistency loss. The former enforces feature alignment across frames, alleviating temporal drift and improving the reliability of proposed long-term memory. In parallel, the latter encourages each token to encode diverse and discriminative target representations observed throughout the sequence via memory-point-memory cyclic walks. As a result, ChronoTrack achieves new state-of-the-art performance on multiple 3D-SOT benchmarks, demonstrating its effectiveness in long-term target modeling with compact memory while running at real-time speed of 42 FPS on a single RTX 4090 GPU. The code is available at https://github.com/ujaejoon/ChronoTrack

Hyun Seok Seong, WonJun Moon, Jae-Pil Heo

Unsupervised object-centric learning models, particularly slot-based architectures, have shown great promise in decomposing complex scenes. However, their reliance on reconstruction-based training creates a fundamental conflict between the sharp, high-frequency attention maps of the encoder and the spatially consistent but blurry reconstruction maps of the decoder. We identify that this discrepancy gives rise to a vicious cycle: the noisy feature map from the encoder forces the decoder to average over possibilities and produce even blurrier outputs, while the gradient computed from blurry reconstruction maps lacks high-frequency details necessary to supervise encoder features. To break this cycle, we introduce Synergistic Representation Learning (SRL) that establishes a virtuous cycle where the encoder and decoder mutually refine one another. SRL leverages the encoder's sharpness to deblur the semantic boundary within the decoder output, while exploiting the decoder's spatial consistency to denoise the encoder's features. This mutual refinement process is stabilized by a warm-up phase with a slot regularization objective that initially allocates distinct entities per slot. By bridging the representational gap between the encoder and decoder, SRL achieves state-of-the-art results on video object-centric learning benchmarks. Codes are available at https://github.com/hynnsk/SRL.

ByeongCheol Lee, Hyun Seok Seong, Sangeek Hyun et al.

A sliding-window inference strategy is commonly adopted in recent training-free open-vocabulary semantic segmentation methods to overcome limitation of the CLIP in processing high-resolution images. However, this approach introduces a new challenge: each window is processed independently, leading to semantic discrepancy across windows. To address this issue, we propose Global-Local Aligned CLIP~(GLA-CLIP), a framework that facilitates comprehensive information exchange across windows. Rather than limiting attention to tokens within individual windows, GLA-CLIP extends key-value tokens to incorporate contextual cues from all windows. Nevertheless, we observe a window bias: outer-window tokens are less likely to be attended, since query features are produced through interactions within the inner window patches, thereby lacking semantic grounding beyond their local context. To mitigate this, we introduce a proxy anchor, constructed by aggregating tokens highly similar to the given query from all windows, which provides a unified semantic reference for measuring similarity across both inner- and outer-window patches. Furthermore, we propose a dynamic normalization scheme that adjusts attention strength according to object scale by dynamically scaling and thresholding the attention map to cope with small-object scenarios. Moreover, GLA-CLIP can be equipped on existing methods and broad their receptive field. Extensive experiments validate the effectiveness of GLA-CLIP in enhancing training-free open-vocabulary semantic segmentation performance. Code is available at https://github.com/2btlFe/GLA-CLIP.

WonJun Moon, Hyun Seok Seong, Jae-Pil Heo

Video Object-Centric Learning seeks to decompose raw videos into a small set of object slots, but existing slot-attention models often suffer from severe over-fragmentation. This is because the model is implicitly encouraged to occupy all slots to minimize the reconstruction objective, thereby representing a single object with multiple redundant slots. We tackle this limitation with a reconstruction-guided slot curriculum (SlotCurri). Training starts with only a few coarse slots and progressively allocates new slots where reconstruction error remains high, thus expanding capacity only where it is needed and preventing fragmentation from the outset. Yet, during slot expansion, meaningful sub-parts can emerge only if coarse-level semantics are already well separated; however, with a small initial slot budget and an MSE objective, semantic boundaries remain blurry. Therefore, we augment MSE with a structure-aware loss that preserves local contrast and edge information to encourage each slot to sharpen its semantic boundaries. Lastly, we propose a cyclic inference that rolls slots forward and then backward through the frame sequence, producing temporally consistent object representations even in the earliest frames. All combined, SlotCurri addresses object over-fragmentation by allocating representational capacity where reconstruction fails, further enhanced by structural cues and cyclic inference. Notable FG-ARI gains of +6.8 on YouTube-VIS and +8.3 on MOVi-C validate the effectiveness of SlotCurri. Our code is available at github.com/wjun0830/SlotCurri.

Jihwan Kim, Miso Lee, Jae-Pil Heo

Temporal Action Detection (TAD) is challenging but fundamental for real-world video applications. Recently, DETR-based models have been devised for TAD but have not performed well yet. In this paper, we point out the problem in the self-attention of DETR for TAD; the attention modules focus on a few key elements, called temporal collapse problem. It degrades the capability of the encoder and decoder since their self-attention modules play no role. To solve the problem, we propose a novel framework, Self-DETR, which utilizes cross-attention maps of the decoder to reactivate self-attention modules. We recover the relationship between encoder features by simple matrix multiplication of the cross-attention map and its transpose. Likewise, we also get the information within decoder queries. By guiding collapsed self-attention maps with the guidance map calculated, we settle down the temporal collapse of self-attention modules in the encoder and decoder. Our extensive experiments demonstrate that Self-DETR resolves the temporal collapse problem by keeping high diversity of attention over all layers.

Sangeek Hyun, MinKyu Lee, Jae-Pil Heo

Modern GANs often introduce adversarial supervision on intermediate generator outputs and interpret the resulting multi-stage synthesis as coarse-to-fine hierarchical generation. In this work, we challenge this interpretation. We argue that standard scale-wise adversarial supervision does not construct a proper coarse-to-fine hierarchy: each intermediate image is independently pushed toward the real distribution at its own resolution, but this scale-wise realism does not ensure that outputs across stages represent the identical generated sample. Moreover, the scale-specific image produced at each stage is not used as an explicit refinement target for the subsequent stage. Therefore, its adversarial loss can improve a scale-specific output without constraining later stages to preserve the same sample trajectory, allowing them to move toward a different sample rather than refine the previous output. We refer to this problem as a cross-scale trajectory misalignment problem. To resolve it, we propose CAT, a Cross-scale Aligned Transformer for multi-scale adversarial generation. CAT keeps the discriminator scale-wise, so each intermediate output is evaluated at its own resolution, while adding a simple generator-side consistency regularization that aligns intermediate outputs with the final output. On class-conditional ImageNet-256, CAT-H/2 achieves an FID-50K of 1.56 with one-step inference after only 60 training epochs, outperforming strong one-step GAN and diffusion/flow baselines.

SuBeen Lee, WonJun Moon, Hyun Seok Seong et al.

The difficulty of the fine-grained image classification mainly comes from a shared overall appearance across classes. Thus, recognizing discriminative details, such as eyes and beaks for birds, is a key in the task. However, this is particularly challenging when training data is limited. To address this, we propose Task Discrepancy Maximization (TDM), a task-oriented channel attention method tailored for fine-grained few-shot classification with two novel modules Support Attention Module (SAM) and Query Attention Module (QAM). SAM highlights channels encoding class-wise discriminative features, while QAM assigns higher weights to object-relevant channels of the query. Based on these submodules, TDM produces task-adaptive features by focusing on channels encoding class-discriminative details and possessed by the query at the same time, for accurate class-sensitive similarity measure between support and query instances. While TDM influences high-level feature maps by task-adaptive calibration of channel-wise importance, we further introduce Instance Attention Module (IAM) operating in intermediate layers of feature extractors to instance-wisely highlight object-relevant channels, by extending QAM. The merits of TDM and IAM and their complementary benefits are experimentally validated in fine-grained few-shot classification tasks. Moreover, IAM is also shown to be effective in coarse-grained and cross-domain few-shot classifications.

WonJun Moon, Hyun Seok Seong, Jae-Pil Heo

A dramatic increase in real-world video volume with extremely diverse and emerging topics naturally forms a long-tailed video distribution in terms of their categories, and it spotlights the need for Video Long-Tailed Recognition (VLTR). In this work, we summarize the challenges in VLTR and explore how to overcome them. The challenges are: (1) it is impractical to re-train the whole model for high-quality features, (2) acquiring frame-wise labels requires extensive cost, and (3) long-tailed data triggers biased training. Yet, most existing works for VLTR unavoidably utilize image-level features extracted from pretrained models which are task-irrelevant, and learn by video-level labels. Therefore, to deal with such (1) task-irrelevant features and (2) video-level labels, we introduce two complementary learnable feature aggregators. Learnable layers in each aggregator are to produce task-relevant representations, and each aggregator is to assemble the snippet-wise knowledge into a video representative. Then, we propose Minority-Oriented Vicinity Expansion (MOVE) that explicitly leverages the class frequency into approximating the vicinity distributions to alleviate (3) biased training. By combining these solutions, our approach achieves state-of-the-art results on large-scale VideoLT and synthetically induced Imbalanced-MiniKinetics200. With VideoLT features from ResNet-50, it attains 18% and 58% relative improvements on head and tail classes over the previous state-of-the-art method, respectively.

Gilhan Park, WonJun Moon, SuBeen Lee et al.

Class-Incremental Semantic Segmentation(CISS) aims to learn new classes without forgetting the old ones, using only the labels of the new classes. To achieve this, two popular strategies are employed: 1) pseudo-labeling and knowledge distillation to preserve prior knowledge; and 2) background weight transfer, which leverages the broad coverage of background in learning new classes by transferring background weight to the new class classifier. However, the first strategy heavily relies on the old model in detecting old classes while undetected pixels are regarded as the background, thereby leading to the background shift towards the old classes(i.e., misclassification of old class as background). Additionally, in the case of the second approach, initializing the new class classifier with background knowledge triggers a similar background shift issue, but towards the new classes. To address these issues, we propose a background-class separation framework for CISS. To begin with, selective pseudo-labeling and adaptive feature distillation are to distill only trustworthy past knowledge. On the other hand, we encourage the separation between the background and new classes with a novel orthogonal objective along with label-guided output distillation. Our state-of-the-art results validate the effectiveness of these proposed methods.

Sang-Heon Shim, Jiwoo Chung, Jae-Pil Heo

In this paper, we first investigate a visual quality degradation problem observed in recent high-resolution virtual try-on approach. The tendency is empirically found that the textures of clothes are squeezed at the sleeve, as visualized in the upper row of Fig.1(a). A main reason for the issue arises from a gradient conflict between two popular losses, the Total Variation (TV) and adversarial losses. Specifically, the TV loss aims to disconnect boundaries between the sleeve and torso in a warped clothing mask, whereas the adversarial loss aims to combine between them. Such contrary objectives feedback the misaligned gradients to a cascaded appearance flow estimation, resulting in undesirable squeezing artifacts. To reduce this, we propose a Sequential Deformation (SD-VITON) that disentangles the appearance flow prediction layers into TV objective-dominant (TVOB) layers and a task-coexistence (TACO) layer. Specifically, we coarsely fit the clothes onto a human body via the TVOB layers, and then keep on refining via the TACO layer. In addition, the bottom row of Fig.1(a) shows a different type of squeezing artifacts around the waist. To address it, we further propose that we first warp the clothes into a tucked-out shirts style, and then partially erase the texture from the warped clothes without hurting the smoothness of the appearance flows. Experimental results show that our SD-VITON successfully resolves both types of artifacts and outperforms the baseline methods. Source code will be available at https://github.com/SHShim0513/SD-VITON.

Hyun Seok Seong, WonJun Moon, SuBeen Lee et al.

The labor-intensive labeling for semantic segmentation has spurred the emergence of Unsupervised Semantic Segmentation. Recent studies utilize patch-wise contrastive learning based on features from image-level self-supervised pretrained models. However, relying solely on similarity-based supervision from image-level pretrained models often leads to unreliable guidance due to insufficient patch-level semantic representations. To address this, we propose a Progressive Proxy Anchor Propagation (PPAP) strategy. This method gradually identifies more trustworthy positives for each anchor by relocating its proxy to regions densely populated with semantically similar samples. Specifically, we initially establish a tight boundary to gather a few reliable positive samples around each anchor. Then, considering the distribution of positive samples, we relocate the proxy anchor towards areas with a higher concentration of positives and adjust the positiveness boundary based on the propagation degree of the proxy anchor. Moreover, to account for ambiguous regions where positive and negative samples may coexist near the positiveness boundary, we introduce an instance-wise ambiguous zone. Samples within these zones are excluded from the negative set, further enhancing the reliability of the negative set. Our state-of-the-art performances on various datasets validate the effectiveness of the proposed method for Unsupervised Semantic Segmentation.

Suho Park, SuBeen Lee, Sangeek Hyun et al.

Few-shot segmentation aims to accurately segment novel target objects within query images using only a limited number of annotated support images. The recent works exploit support background as well as its foreground to precisely compute the dense correlations between query and support. However, they overlook the characteristics of the background that generally contains various types of objects. In this paper, we highlight this characteristic of background which can bring problematic cases as follows: (1) when the query and support backgrounds are dissimilar and (2) when objects in the support background are similar to the target object in the query. Without any consideration of the above cases, adopting the entire support background leads to a misprediction of the query foreground as background. To address this issue, we propose Task-disruptive Background Suppression (TBS), a module to suppress those disruptive support background features based on two spatial-wise scores: query-relevant and target-relevant scores. The former aims to mitigate the impact of unshared features solely existing in the support background, while the latter aims to reduce the influence of target-similar support background features. Based on these two scores, we define a query background relevant score that captures the similarity between the backgrounds of the query and the support, and utilize it to scale support background features to adaptively restrict the impact of disruptive support backgrounds. Our proposed method achieves state-of-the-art performance on PASCAL-5 and COCO-20 datasets on 1-shot segmentation. Our official code is available at github.com/SuhoPark0706/TBSNet.

Jihwan Kim, Miso Lee, Cheol-Ho Cho et al.

Temporal Action Detection (TAD) is fundamental yet challenging for real-world video applications. Leveraging the unique benefits of transformers, various DETR-based approaches have been adopted in TAD. However, it has recently been identified that the attention collapse in self-attention causes the performance degradation of DETR for TAD. Building upon previous research, this paper newly addresses the attention collapse problem in cross-attention within DETR-based TAD methods. Moreover, our findings reveal that cross-attention exhibits patterns distinct from predictions, indicating a short-cut phenomenon. To resolve this, we propose a new framework, Prediction-Feedback DETR (Pred-DETR), which utilizes predictions to restore the collapse and align the cross- and self-attention with predictions. Specifically, we devise novel prediction-feedback objectives using guidance from the relations of the predictions. As a result, Pred-DETR significantly alleviates the collapse and achieves state-of-the-art performance among DETR-based methods on various challenging benchmarks including THUMOS14, ActivityNet-v1.3, HACS, and FineAction.

MinKyu Lee, Jae-Pil Heo

Recent deep-learning-based single image super-resolution (SISR) methods have shown impressive performance whereas typical methods train their networks by minimizing the pixel-wise distance with respect to a given high-resolution (HR) image. However, despite the basic training scheme being the predominant choice, its use in the context of ill-posed inverse problems has not been thoroughly investigated. In this work, we aim to provide a better comprehension of the underlying constituent by decomposing target HR images into two subcomponents: (1) the optimal centroid which is the expectation over multiple potential HR images, and (2) the inherent noise defined as the residual between the HR image and the centroid. Our findings show that the current training scheme cannot capture the ill-posed nature of SISR and becomes vulnerable to the inherent noise term, especially during early training steps. To tackle this issue, we propose a novel optimization method that can effectively remove the inherent noise term in the early steps of vanilla training by estimating the optimal centroid and directly optimizing toward the estimation. Experimental results show that the proposed method can effectively enhance the stability of vanilla training, leading to overall performance gain. Codes are available at github.com/2minkyulee/ECO.

Yerim Jeon, Subeen Lee, Jihwan Kim et al.

Few-shot object counting has garnered significant attention for its practicality as it aims to count target objects in a query image based on given exemplars without additional training. However, the prevailing extract-and-match approach has a shortcoming: query and exemplar features lack interaction during feature extraction since they are extracted independently and later correlated based on similarity. This can lead to insufficient target awareness and confusion in identifying the actual target when multiple class objects coexist. To address this, we propose a novel framework, Mutually-Aware FEAture learning (MAFEA), which encodes query and exemplar features with mutual awareness from the outset. By encouraging interaction throughout the pipeline, we obtain target-aware features robust to a multi-category scenario. Furthermore, we introduce background token to effectively associate the query's target region with exemplars and decouple its background region. Our extensive experiments demonstrate that our model achieves state-of-the-art performance on FSCD-LVIS and FSC-147 benchmarks with remarkably reduced target confusion.

Jihwan Kim, Miso Lee, Jae-Pil Heo

Temporal action detection (TAD) is challenging, yet fundamental for real-world video applications. Recently, DETR-based models for TAD have been prevailing thanks to their unique benefits. However, transformers demand a huge dataset, and unfortunately data scarcity in TAD causes a severe degeneration. In this paper, we identify two crucial problems from data scarcity: attention collapse and imbalanced performance. To this end, we propose a new pre-training strategy, Long-Term Pre-training (LTP), tailored for transformers. LTP has two main components: 1) class-wise synthesis, 2) long-term pretext tasks. Firstly, we synthesize long-form video features by merging video snippets of a target class and non-target classes. They are analogous to untrimmed data used in TAD, despite being created from trimmed data. In addition, we devise two types of long-term pretext tasks to learn long-term dependency. They impose long-term conditions such as finding second-to-fourth or short-duration actions. Our extensive experiments show state-of-the-art performances in DETR-based methods on ActivityNet-v1.3 and THUMOS14 by a large margin. Moreover, we demonstrate that LTP significantly relieves the data scarcity issues in TAD.

SuBeen Lee, WonJun Moon, Hyun Seok Seong et al.

Few-Shot Action Recognition (FSAR) aims to train a model with only a few labeled video instances. A key challenge in FSAR is handling divergent narrative trajectories for precise video matching. While the frame- and tuple-level alignment approaches have been promising, their methods heavily rely on pre-defined and length-dependent alignment units (e.g., frames or tuples), which limits flexibility for actions of varying lengths and speeds. In this work, we introduce a novel TEmporal Alignment-free Matching (TEAM) approach, which eliminates the need for temporal units in action representation and brute-force alignment during matching. Specifically, TEAM represents each video with a fixed set of pattern tokens that capture globally discriminative clues within the video instance regardless of action length or speed, ensuring its flexibility. Furthermore, TEAM is inherently efficient, using token-wise comparisons to measure similarity between videos, unlike existing methods that rely on pairwise comparisons for temporal alignment. Additionally, we propose an adaptation process that identifies and removes common information across classes, establishing clear boundaries even between novel categories. Extensive experiments demonstrate the effectiveness of TEAM. Codes are available at github.com/leesb7426/TEAM.

Suho Park, SuBeen Lee, Hyun Seok Seong et al.

We propose Foreground-Covering Prototype Generation and Matching to resolve Few-Shot Segmentation (FSS), which aims to segment target regions in unlabeled query images based on labeled support images. Unlike previous research, which typically estimates target regions in the query using support prototypes and query pixels, we utilize the relationship between support and query prototypes. To achieve this, we utilize two complementary features: SAM Image Encoder features for pixel aggregation and ResNet features for class consistency. Specifically, we construct support and query prototypes with SAM features and distinguish query prototypes of target regions based on ResNet features. For the query prototype construction, we begin by roughly guiding foreground regions within SAM features using the conventional pseudo-mask, then employ iterative cross-attention to aggregate foreground features into learnable tokens. Here, we discover that the cross-attention weights can effectively alternate the conventional pseudo-mask. Therefore, we use the attention-based pseudo-mask to guide ResNet features to focus on the foreground, then infuse the guided ResNet feature into the learnable tokens to generate class-consistent query prototypes. The generation of the support prototype is conducted symmetrically to that of the query one, with the pseudo-mask replaced by the ground-truth mask. Finally, we compare these query prototypes with support ones to generate prompts, which subsequently produce object masks through the SAM Mask Decoder. Our state-of-the-art performances on various datasets validate the effectiveness of the proposed method for FSS. Our official code is available at https://github.com/SuhoPark0706/FCP

Jiwoo Chung, Sangeek Hyun, MinKyu Lee et al.

Diffusion models are a strong backbone for visual generation, but their inherently sequential denoising process leads to slow inference. Previous methods accelerate sampling by caching and reusing intermediate outputs based on feature distances between adjacent timesteps. However, existing caching strategies typically rely on raw feature differences that entangle content and noise. This design overlooks spectral evolution, where low-frequency structure appears early and high-frequency detail is refined later. We introduce Spectral-Evolution-Aware Cache (SeaCache), a training-free cache schedule that bases reuse decisions on a spectrally aligned representation. Through theoretical and empirical analysis, we derive a Spectral-Evolution-Aware (SEA) filter that preserves content-relevant components while suppressing noise. Employing SEA-filtered input features to estimate redundancy leads to dynamic schedules that adapt to content while respecting the spectral priors underlying the diffusion model. Extensive experiments on diverse visual generative models and the baselines show that SeaCache achieves state-of-the-art latency-quality trade-offs.

SuBeen Lee, GilHan Park, WonJun Moon et al.

Despite the impressive zero-shot capabilities of Vision-Language Models (VLMs), they often struggle in downstream tasks with distribution shifts from the pre-training data. Few-Shot Adaptation (FSA-VLM) has emerged as a key solution, typically using Parameter-Efficient Fine-Tuning (PEFT) to adapt models with minimal data. However, these PEFT methods are constrained by their reliance on fixed, handcrafted prompts, which are often insufficient to understand the semantics of classes. While some studies have proposed leveraging image-induced prompts to provide additional clues for classification, they introduce prohibitive computational overhead at inference. Therefore, we introduce Auxiliary Descriptive Knowledge (ADK), a novel framework that efficiently enriches text representations without compromising efficiency. ADK first leverages a Large Language Model to generate a rich set of descriptive prompts for each class offline. These pre-computed features are then deployed in two ways: (1) as Compositional Knowledge, an averaged representation that provides rich semantics, especially beneficial when class names are ambiguous or unfamiliar to the VLM; and (2) as Instance-Specific Knowledge, where a lightweight, non-parametric attention mechanism dynamically selects the most relevant descriptions for a given image. This approach provides two additional types of knowledge alongside the handcrafted prompt, thereby facilitating category distinction across various domains. Also, ADK acts as a parameter-free, plug-and-play component that enhances existing PEFT methods. Extensive experiments demonstrate that ADK consistently boosts the performance of multiple PEFT baselines, setting a new state-of-the-art across various scenarios.

Yerim Jeon, Miso Lee, WonJun Moon et al.

Recent advances in 3D scene-language understanding have leveraged Large Language Models (LLMs) for 3D reasoning by transferring their general reasoning ability to 3D multi-modal contexts. However, existing methods typically adopt standard decoders from language modeling, which rely on a causal attention mask. This design introduces two fundamental conflicts in 3D scene understanding: sequential bias among order-agnostic 3D objects and restricted object-instruction attention, hindering task-specific reasoning. To overcome these limitations, we propose 3D Spatial Language Instruction Mask (3D-SLIM), an effective masking strategy that replaces the causal mask with an adaptive attention mask tailored to the spatial structure of 3D scenes. Our 3D-SLIM introduces two key components: a Geometry-adaptive Mask that constrains attention based on spatial density rather than token order, and an Instruction-aware Mask that enables object tokens to directly access instruction context. This design allows the model to process objects based on their spatial relationships while being guided by the user's task. 3D-SLIM is simple, requires no architectural modifications, and adds no extra parameters, yet it yields substantial performance improvements across diverse 3D scene-language tasks. Extensive experiments across multiple benchmarks and LLM baselines validate its effectiveness and underscore the critical role of decoder design in 3D multi-modal reasoning.

Jihwan Kim, Jaehyun Choi, Yerim Jeon et al.

Temporal action detection (TAD) is challenging, yet fundamental for real-world video applications. Large temporal scale variation of actions is one of the most primary difficulties in TAD. Naturally, multi-scale features have potential in localizing actions of diverse lengths as widely used in object detection. Nevertheless, unlike objects in images, actions have more ambiguity in their boundaries. That is, small neighboring objects are not considered as a large one while short adjoining actions can be misunderstood as a long one. In the coarse-to-fine feature pyramid via pooling, these vague action boundaries can fade out, which we call 'vanishing boundary problem'. To this end, we propose Boundary-Recovering Network (BRN) to address the vanishing boundary problem. BRN constructs scale-time features by introducing a new axis called scale dimension by interpolating multi-scale features to the same temporal length. On top of scale-time features, scale-time blocks learn to exchange features across scale levels, which can effectively settle down the issue. Our extensive experiments demonstrate that our model outperforms the state-of-the-art on the two challenging benchmarks, ActivityNet-v1.3 and THUMOS14, with remarkably reduced degree of the vanishing boundary problem.

WonJun Moon, MinSeok Jung, Gilhan Park et al.

Partially Relevant Video Retrieval (PRVR) seeks videos where only part of the content matches a text query. Existing methods treat every annotated text-video pair as a positive and all others as negatives, ignoring the rich semantic variation both within a single video and across different videos. Consequently, embeddings of both queries and their corresponding video-clip segments for distinct events within the same video collapse together, while embeddings of semantically similar queries and segments from different videos are driven apart. This limits retrieval performance when videos contain multiple, diverse events. This paper addresses the aforementioned problems, termed as semantic collapse, in both the text and video embedding spaces. We first introduce Text Correlation Preservation Learning, which preserves the semantic relationships encoded by the foundation model across text queries. To address collapse in video embeddings, we propose Cross-Branch Video Alignment (CBVA), a contrastive alignment method that disentangles hierarchical video representations across temporal scales. Subsequently, we introduce order-preserving token merging and adaptive CBVA to enhance alignment by producing video segments that are internally coherent yet mutually distinctive. Extensive experiments on PRVR benchmarks demonstrate that our framework effectively prevents semantic collapse and substantially improves retrieval accuracy.

Jiwoo Chung, Sangeek Hyun, Jae-Pil Heo

Despite the impressive generative capabilities of diffusion models, existing diffusion model-based style transfer methods require inference-stage optimization (e.g. fine-tuning or textual inversion of style) which is time-consuming, or fails to leverage the generative ability of large-scale diffusion models. To address these issues, we introduce a novel artistic style transfer method based on a pre-trained large-scale diffusion model without any optimization. Specifically, we manipulate the features of self-attention layers as the way the cross-attention mechanism works; in the generation process, substituting the key and value of content with those of style image. This approach provides several desirable characteristics for style transfer including 1) preservation of content by transferring similar styles into similar image patches and 2) transfer of style based on similarity of local texture (e.g. edge) between content and style images. Furthermore, we introduce query preservation and attention temperature scaling to mitigate the issue of disruption of original content, and initial latent Adaptive Instance Normalization (AdaIN) to deal with the disharmonious color (failure to transfer the colors of style). Our experimental results demonstrate that our proposed method surpasses state-of-the-art methods in both conventional and diffusion-based style transfer baselines.

WonJun Moon, Hyun Seok Seong, Jae-Pil Heo

Facilitating an entity's interaction with objects requires accurately identifying parts that afford specific actions. Weakly supervised affordance grounding (WSAG) seeks to imitate human learning from third-person demonstrations, where humans intuitively grasp functional parts without needing pixel-level annotations. To achieve this, grounding is typically learned using a shared classifier across images from different perspectives, along with distillation strategies incorporating part discovery process. However, since affordance-relevant parts are not always easily distinguishable, models primarily rely on classification, often focusing on common class-specific patterns that are unrelated to affordance. To address this limitation, we move beyond isolated part-level learning by introducing selective prototypical and pixel contrastive objectives that adaptively learn affordance-relevant cues at both the part and object levels, depending on the granularity of the available information. Initially, we find the action-associated objects in both egocentric (object-focused) and exocentric (third-person example) images by leveraging CLIP. Then, by cross-referencing the discovered objects of complementary views, we excavate the precise part-level affordance clues in each perspective. By consistently learning to distinguish affordance-relevant regions from affordance-irrelevant background context, our approach effectively shifts activation from irrelevant areas toward meaningful affordance cues. Experimental results demonstrate the effectiveness of our method. Codes are available at github.com/hynnsk/SelectiveCL.

Miso Lee, Sangeek Hyun, Yerim Jeon et al.

While Gaussian Splatting-based Feature Fields (GSFFs) have shown promise for visual localization, this paper highlights that photometrically optimized GSFFs are inherently ill-suited for 2D-3D matching. The volumetric extent of each Gaussian induces many-to-one pixel-to-point mappings that destabilize PnP-based pose estimation, while photometric optimization gives rise to superfluous Gaussians devoid of multi-view consistency. To address these issues, we propose SplitGS-Loc, a localization-specialized GSFFs construction framework that disambiguates 2D-3D correspondences by exploiting Gaussian attributes. Our key design, Mixture-of-Gaussians-based splitting, decomposes each Gaussian into smaller Gaussians, replacing ambiguous many-to-one with precise one-to-one correspondences. In parallel, we exploit composition weights from GS rasterization to select Gaussians that significantly and consistently contribute across multiple views and aggregate discriminative features through strong pixel-Gaussian associations, enforcing multi-view consistency. The resulting compact yet discriminative feature fields enable stable PnP convergence, achieving state-of-the-art performance on localization benchmarks. Extensive experiments validate that SplitGS-Loc extends the utility of photometric GSFFs to accurate and efficient localization by exploiting Gaussian attributes, without per-scene training or iterative pose refinement.

Seunggu Kang, WonJun Moon, Euiyeon Kim et al.

Zero-Shot Object Counting (ZSOC) aims to count referred instances of arbitrary classes in a query image without human-annotated exemplars. To deal with ZSOC, preceding studies proposed a two-stage pipeline: discovering exemplars and counting. However, there remains a challenge of vulnerability to error propagation of the sequentially designed two-stage process. In this work, an one-stage baseline, Visual-Language Baseline (VLBase), exploring the implicit association of the semantic-patch embeddings of CLIP is proposed. Subsequently, the extension of VLBase to Visual-language Counter (VLCounter) is achieved by incorporating three modules devised to tailor VLBase for object counting. First, Semantic-conditioned Prompt Tuning (SPT) is introduced within the image encoder to acquire target-highlighted representations. Second, Learnable Affine Transformation (LAT) is employed to translate the semantic-patch similarity map to be appropriate for the counting task. Lastly, the layer-wisely encoded features are transferred to the decoder through Segment-aware Skip Connection (SaSC) to keep the generalization capability for unseen classes. Through extensive experiments on FSC147, CARPK, and PUCPR+, the benefits of the end-to-end framework, VLCounter, are demonstrated.

Jiwoo Chung, Sangeek Hyun, Hyunjun Kim et al.

Recent advances in text-to-image generative models have enabled numerous practical applications, including subject-driven generation, which fine-tunes pretrained models to capture subject semantics from only a few examples. While diffusion-based models produce high-quality images, their extensive denoising steps result in significant computational overhead, limiting real-world applicability. Visual autoregressive (VAR) models, which predict next-scale tokens rather than spatially adjacent ones, offer significantly faster inference suitable for practical deployment. In this paper, we propose the first VAR-based approach for subject-driven generation. However, naive fine-tuning VAR leads to computational overhead, language drift, and reduced diversity. To address these challenges, we introduce selective layer tuning to reduce complexity and prior distillation to mitigate language drift. Additionally, we found that the early stages have a greater influence on the generation of subject than the latter stages, which merely synthesize minor details. Based on this finding, we propose scale-wise weighted tuning, which prioritizes coarser resolutions for promoting the model to focus on the subject-relevant information instead of local details. Extensive experiments validate that our method significantly outperforms diffusion-based baselines across various metrics and demonstrates its practical usage.

Jiwoo Chung, Sangeek Hyun, Sang-Heon Shim et al.

StyleGAN has shown remarkable performance in unconditional image generation. However, its high computational cost poses a significant challenge for practical applications. Although recent efforts have been made to compress StyleGAN while preserving its performance, existing compressed models still lag behind the original model, particularly in terms of sample diversity. To overcome this, we propose a novel channel pruning method that leverages varying sensitivities of channels to latent vectors, which is a key factor in sample diversity. Specifically, by assessing channel importance based on their sensitivities to latent vector perturbations, our method enhances the diversity of samples in the compressed model. Since our method solely focuses on the channel pruning stage, it has complementary benefits with prior training schemes without additional training cost. Extensive experiments demonstrate that our method significantly enhances sample diversity across various datasets. Moreover, in terms of FID scores, our method not only surpasses state-of-the-art by a large margin but also achieves comparable scores with only half training iterations.

Miso Lee, Jihwan Kim, Jae-Pil Heo

Multi-scene absolute pose regression addresses the demand for fast and memory-efficient camera pose estimation across various real-world environments. Nowadays, transformer-based model has been devised to regress the camera pose directly in multi-scenes. Despite its potential, transformer encoders are underutilized due to the collapsed self-attention map, having low representation capacity. This work highlights the problem and investigates it from a new perspective: distortion of query-key embedding space. Based on the statistical analysis, we reveal that queries and keys are mapped in completely different spaces while only a few keys are blended into the query region. This leads to the collapse of the self-attention map as all queries are considered similar to those few keys. Therefore, we propose simple but effective solutions to activate self-attention. Concretely, we present an auxiliary loss that aligns queries and keys, preventing the distortion of query-key space and encouraging the model to find global relations by self-attention. In addition, the fixed sinusoidal positional encoding is adopted instead of undertrained learnable one to reflect appropriate positional clues into the inputs of self-attention. As a result, our approach resolves the aforementioned problem effectively, thus outperforming existing methods in both outdoor and indoor scenes.

WonJun Moon, Cheol-Ho Cho, Woojin Jun et al.

In a retrieval system, simultaneously achieving search accuracy and efficiency is inherently challenging. This challenge is particularly pronounced in partially relevant video retrieval (PRVR), where incorporating more diverse context representations at varying temporal scales for each video enhances accuracy but increases computational and memory costs. To address this dichotomy, we propose a prototypical PRVR framework that encodes diverse contexts within a video into a fixed number of prototypes. We then introduce several strategies to enhance text association and video understanding within the prototypes, along with an orthogonal objective to ensure that the prototypes capture a diverse range of content. To keep the prototypes searchable via text queries while accurately encoding video contexts, we implement cross- and uni-modal reconstruction tasks. The cross-modal reconstruction task aligns the prototypes with textual features within a shared space, while the uni-modal reconstruction task preserves all video contexts during encoding. Additionally, we employ a video mixing technique to provide weak guidance to further align prototypes and associated textual representations. Extensive evaluations on TVR, ActivityNet-Captions, and QVHighlights validate the effectiveness of our approach without sacrificing efficiency.

MinKyu Lee, Sangeek Hyun, Woojin Jun et al.

This work tackles the fidelity objective in the perceptual super-resolution~(SR). Specifically, we address the shortcomings of pixel-level $L_\text{p}$ loss ($\mathcal{L}_\text{pix}$) in the GAN-based SR framework. Since $L_\text{pix}$ is known to have a trade-off relationship against perceptual quality, prior methods often multiply a small scale factor or utilize low-pass filters. However, this work shows that these circumventions fail to address the fundamental factor that induces blurring. Accordingly, we focus on two points: 1) precisely discriminating the subcomponent of $L_\text{pix}$ that contributes to blurring, and 2) only guiding based on the factor that is free from this trade-off relationship. We show that they can be achieved in a surprisingly simple manner, with an Auto-Encoder (AE) pretrained with $L_\text{pix}$. Accordingly, we propose the Auto-Encoded Supervision for Optimal Penalization loss ($L_\text{AESOP}$), a novel loss function that measures distance in the AE space, instead of the raw pixel space. Note that the AE space indicates the space after the decoder, not the bottleneck. By simply substituting $L_\text{pix}$ with $L_\text{AESOP}$, we can provide effective reconstruction guidance without compromising perceptual quality. Designed for simplicity, our method enables easy integration into existing SR frameworks. Experimental results verify that AESOP can lead to favorable results in the perceptual SR task.

Sangeek Hyun, MinKyu Lee, Jae-Pil Heo

Scalability has driven recent advances in generative modeling, yet its principles remain underexplored for adversarial learning. We investigate the scalability of Generative Adversarial Networks (GANs) through two design choices that have proven to be effective in other types of generative models: training in a compact Variational Autoencoder latent space and adopting purely transformer-based generators and discriminators. Training in latent space enables efficient computation while preserving perceptual fidelity, and this efficiency pairs naturally with plain transformers, whose performance scales with computational budget. Building on these choices, we analyze failure modes that emerge when naively scaling GANs. Specifically, we find issues as underutilization of early layers in the generator and optimization instability as the network scales. Accordingly, we provide simple and scale-friendly solutions as lightweight intermediate supervision and width-aware learning-rate adjustment. Our experiments show that GAT, a purely transformer-based and latent-space GANs, can be easily trained reliably across a wide range of capacities (S through XL). Moreover, GAT-XL/2 achieves state-of-the-art single-step, class-conditional generation performance (FID of 2.96) on ImageNet-256 in just 40 epochs, 6x fewer epochs than strong baselines.

Jiyun Im, SuBeen Lee, Miso Lee et al.

Few-Shot 3D Point Cloud Segmentation (FS-PCS) aims to predict per-point labels for an unlabeled point cloud, given only a few labeled examples. To extract discriminative representations from the limited support set, existing methods have constructed prototypes using conventional algorithms such as farthest point sampling. However, we point out that its initial randomness significantly affects FS-PCS performance and that the prototype generation process remains underexplored despite its prevalence. This motivates us to investigate an advanced prototype generation method based on attention mechanism. Despite its potential, we found that vanilla module suffers from the distributional gap between learnable prototypical tokens and support features. To overcome this, we propose White Aggregation and Restoration Module (WARM), which resolves the misalignment by sandwiching cross-attention between whitening and coloring transformations. Specifically, whitening aligns the support features to prototypical tokens before attention process, and subsequently coloring restores the original distribution to the attended tokens. This simple yet effective design enables robust attention, thereby generating representative prototypes by capturing the semantic relationships among support features. Our method achieves state-of-the-art performance with a significant margin on multiple FS-PCS benchmarks, demonstrating its effectiveness through extensive experiments.

Eunseo Koh, Seunghoo Hong, Tae-Young Kim et al.

Text-to-Image (T2I) diffusion models have made significant progress in generating diverse high-quality images from textual prompts. However, these models still face challenges in suppressing content that is strongly entangled with specific words. For example, when generating an image of "Charlie Chaplin", a "mustache" consistently appears even if explicitly instructed not to include it, as the concept of "mustache" is strongly entangled with "Charlie Chaplin". To address this issue, we propose a novel approach to directly suppress such entangled content within the text embedding space of diffusion models. Our method introduces a delta vector that modifies the text embedding to weaken the influence of undesired content in the generated image, and we further demonstrate that this delta vector can be easily obtained through a zero-shot approach. Furthermore, we propose a Selective Suppression with Delta Vector (SSDV) method to adapt delta vector into the cross-attention mechanism, enabling more effective suppression of unwanted content in regions where it would otherwise be generated. Additionally, we enabled more precise suppression in personalized T2I models by optimizing delta vector, which previous baselines were unable to achieve. Extensive experimental results demonstrate that our approach significantly outperforms existing methods, both in terms of quantitative and qualitative metrics.

MinKyu Lee, Sangeek Hyun, Woojin Jun et al.

This work investigates the internal training dynamics of image restoration~(IR) Transformers and uncovers a critical yet overlooked issue: conventional LayerNorm leads feature magnitude divergence, up to a million scale, and collapses channel-wise entropy. We analyze this phenomenon from the perspective of networks attempting to bypass constraints imposed by conventional LayerNorm due to conflicts against requirements in IR tasks. Accordingly, we address two misalignments between LayerNorm and IR tasks, and later show that addressing these mismatches leads to both stabilized training dynamics and improved IR performance. Specifically, conventional LayerNorm works in a per-token manner, disrupting spatial correlations between tokens, essential in IR tasks. Also, it employs an input-independent normalization that restricts the flexibility of feature scales, required to preserve input-specific statistics. Together, these mismatches significantly hinder IR Transformer's ability to accurately preserve low-level features throughout the network. To this end, we introduce Image Restoration Transformer Tailored Layer Normalization~(i-LN), a surprisingly simple drop-in replacement for conventional LayerNorm. We propose to normalize features in a holistic manner across the entire spatio-channel dimension, preserving spatial relationships among individual tokens. Additionally, we introduce an input-adaptive rescaling strategy that maintains the feature range flexibility required by individual inputs. Together, these modifications effectively contribute to preserving low-level feature statistics of inputs throughout IR Transformers. Experimental results verify that this combined strategy enhances both the stability and performance of IR Transformers across various IR tasks.

Sangeek Hyun, Jae-Pil Heo

Most advances in 3D Generative Adversarial Networks (3D GANs) largely depend on ray casting-based volume rendering, which incurs demanding rendering costs. One promising alternative is rasterization-based 3D Gaussian Splatting (3D-GS), providing a much faster rendering speed and explicit 3D representation. In this paper, we exploit Gaussian as a 3D representation for 3D GANs by leveraging its efficient and explicit characteristics. However, in an adversarial framework, we observe that a naïve generator architecture suffers from training instability and lacks the capability to adjust the scale of Gaussians. This leads to model divergence and visual artifacts due to the absence of proper guidance for initialized positions of Gaussians and densification to manage their scales adaptively. To address these issues, we introduce a generator architecture with a hierarchical multi-scale Gaussian representation that effectively regularizes the position and scale of generated Gaussians. Specifically, we design a hierarchy of Gaussians where finer-level Gaussians are parameterized by their coarser-level counterparts; the position of finer-level Gaussians would be located near their coarser-level counterparts, and the scale would monotonically decrease as the level becomes finer, modeling both coarse and fine details of the 3D scene. Experimental results demonstrate that ours achieves a significantly faster rendering speed (x100) compared to state-of-the-art 3D consistent GANs with comparable 3D generation capability. Project page: https://hse1032.github.io/gsgan.