Seokju Yun

Dongheon Lee, Seokju Yun, Jaegyun Im et al.

Recent Super-Resolution~(SR) methods mainly adopt Transformers for their strong long-range modeling capability and exceptional representational capacity. However, most SR Transformers rely heavily on relative positional bias~(RPB), which prevents them from leveraging hardware-efficient attention kernels such as FlashAttention. This limitation imposes a prohibitive computational burden during both training and inference, severely restricting attempts to scale SR Transformers by enlarging the training patch size or the self-attention window. Consequently, unlike other domains that actively exploit the inherent scalability of Transformers, SR Transformers remain heavily focused on effectively utilizing limited receptive fields. In this paper, we propose Rank-factorized Implicit Neural Bias~(RIB), an alternative to RPB that enables FlashAttention in SR Transformers. Specifically, RIB approximates positional bias using low-rank implicit neural representations and concatenates them with pixel content tokens in a channel-wise manner, turning the element-wise bias addition in attention score computation into a dot-product operation. Further, we introduce a convolutional local attention and a cyclic window strategy to fully leverage the advantages of long-range interactions enabled by RIB and FlashAttention. We enlarge the window size up to \textbf{96$\times$96} while jointly scaling the training patch size and the dataset size, maximizing the benefits of Transformers in the SR task. As a result, our network achieves \textbf{35.63\,dB PSNR} on Urban100$\times$2, while reducing training and inference time by \textbf{2.1$\times$} and \textbf{2.9$\times$}, respectively, compared to the RPB-based SR Transformer~(PFT).

Seokju Yun, Youngmin Ro

We introduce Dynamic Mobile-Former(DMF), maximizes the capabilities of dynamic convolution by harmonizing it with efficient operators.Our Dynamic MobileFormer effectively utilizes the advantages of Dynamic MobileNet (MobileNet equipped with dynamic convolution) using global information from light-weight attention.A Transformer in Dynamic Mobile-Former only requires a few randomly initialized tokens to calculate global features, making it computationally efficient.And a bridge between Dynamic MobileNet and Transformer allows for bidirectional integration of local and global features.We also simplify the optimization process of vanilla dynamic convolution by splitting the convolution kernel into an input-agnostic kernel and an input-dependent kernel.This allows for optimization in a wider kernel space, resulting in enhanced capacity.By integrating lightweight attention and enhanced dynamic convolution, our Dynamic Mobile-Former achieves not only high efficiency, but also strong performance.We benchmark the Dynamic Mobile-Former on a series of vision tasks, and showcase that it achieves impressive performance on image classification, COCO detection, and instanace segmentation.For example, our DMF hits the top-1 accuracy of 79.4% on ImageNet-1K, much higher than PVT-Tiny by 4.3% with only 1/4 FLOPs.Additionally,our proposed DMF-S model performed well on challenging vision datasets such as COCO, achieving a 39.0% mAP,which is 1% higher than that of the Mobile-Former 508M model, despite using 3 GFLOPs less computations.Code and models are available at https://github.com/ysj9909/DMF

Seokju Yun, Dongheon Lee, Noori Bae et al.

As AI systems are being integrated more rapidly into diverse and complex real-world environments, the ability to perform holistic reasoning over an implicit query and an image to localize a target is becoming increasingly important. However, recent reasoning segmentation methods fail to sufficiently elicit the visual reasoning capabilities of the base mode. In this work, we present Segment Anything Reasoner (StAR), a comprehensive framework that refines the design space from multiple perspectives-including parameter-tuning scheme, reward functions, learning strategies and answer format-and achieves substantial improvements over recent baselines. In addition, for the first time, we successfully introduce parallel test-time scaling to the segmentation task, pushing the performance boundary even further. To extend the scope and depth of reasoning covered by existing benchmark, we also construct the ReasonSeg-X, which compactly defines reasoning types and includes samples that require deeper reasoning. Leveraging this dataset, we train StAR with a rollout-expanded selective-tuning approach to activate the base model's latent reasoning capabilities, and establish a rigorous benchmark for systematic, fine-grained evaluation of advanced methods. With only 5k training samples, StAR achieves significant gains over its base counterparts across extensive benchmarks, demonstrating that our method effectively brings dormant reasoning competence to the surface.

Chanseul Cho, Seokju Yun, Jeaseong Jeon et al.

Domain Generalized Semantic Segmentation (DGSS) aims to maintain robust performance across unseen target domains. Vision Foundation Models (VFMs) offer rich multi-domain knowledge that can enhance generalization. However, strategies for actively exploiting the rich subspace structures within VFMs remain under-explored, with many existing methods focusing primarily on preserving pre-trained knowledge. Furthermore, their LoRA components often suffer from limited representational diversity and inefficient parameter utilization. We propose RecycleLoRA, which addresses both challenges by employing Rank-Revealing QR Decomposition (RRQR) to systematically exploit VFM's subspace structures and enhance LoRA's representational richness. Our main adapter leverages minor subspace directions identified by RRQR to learn diverse and independent features, achieving competitive performance even when used alone. We further introduce a sub adapter that carefully refines major directions with minimal adjustments, providing complementary improvements to the main adapter's strong baseline performance. This design enables the dual adapters to learn distinct representations without requiring additional regularization losses. Our systematic exploitation of pre-trained subspace structures through RRQR-based initialization leads to superior domain generalization performance. RecycleLoRA achieves state-of-the-art performance on both synthetic-to-real generalization and real-to-real generalization tasks without complex architectures or additional inference latency.

Dongheon Lee, Seokju Yun, Youngmin Ro

For Image Super-Resolution (SR), it is common to train and evaluate scale-specific models composed of an encoder and upsampler for each targeted scale. Consequently, many SR studies encounter substantial training times and complex deployment requirements. In this paper, we address this limitation by training and evaluating multiple scales simultaneously. Notably, we observe that encoder features are similar across scales and that the Sub-Pixel Convolution (SPConv), widely-used scale-specific upsampler, exhibits strong inter-scale correlations in its functionality. Building on these insights, we propose a multi-scale framework that employs a single encoder in conjunction with Implicit Grid Convolution (IGConv), our novel upsampler, which unifies SPConv across all scales within a single module. Extensive experiments demonstrate that our framework achieves comparable performance to existing fixed-scale methods while reducing the training budget and stored parameters three-fold and maintaining the same latency. Additionally, we propose IGConv$^{+}$ to improve performance further by addressing spectral bias and allowing input-dependent upsampling and ensembled prediction. As a result, ATD-IGConv$^{+}$ achieves a notable 0.21dB improvement in PSNR on Urban100$\times$4, while also reducing the training budget, stored parameters, and inference cost compared to the existing ATD.

Seokju Yun, Youngmin Ro

Recently, efficient Vision Transformers have shown great performance with low latency on resource-constrained devices. Conventionally, they use 4x4 patch embeddings and a 4-stage structure at the macro level, while utilizing sophisticated attention with multi-head configuration at the micro level. This paper aims to address computational redundancy at all design levels in a memory-efficient manner. We discover that using larger-stride patchify stem not only reduces memory access costs but also achieves competitive performance by leveraging token representations with reduced spatial redundancy from the early stages. Furthermore, our preliminary analyses suggest that attention layers in the early stages can be substituted with convolutions, and several attention heads in the latter stages are computationally redundant. To handle this, we introduce a single-head attention module that inherently prevents head redundancy and simultaneously boosts accuracy by parallelly combining global and local information. Building upon our solutions, we introduce SHViT, a Single-Head Vision Transformer that obtains the state-of-the-art speed-accuracy tradeoff. For example, on ImageNet-1k, our SHViT-S4 is 3.3x, 8.1x, and 2.4x faster than MobileViTv2 x1.0 on GPU, CPU, and iPhone12 mobile device, respectively, while being 1.3% more accurate. For object detection and instance segmentation on MS COCO using Mask-RCNN head, our model achieves performance comparable to FastViT-SA12 while exhibiting 3.8x and 2.0x lower backbone latency on GPU and mobile device, respectively.

Seokju Yun, Seunghye Chae, Dongheon Lee et al.

Domain generalization (DG) aims to adapt a model using one or multiple source domains to ensure robust performance in unseen target domains. Recently, Parameter-Efficient Fine-Tuning (PEFT) of foundation models has shown promising results in the context of DG problem. Nevertheless, existing PEFT methods still struggle to strike a balance between preserving generalizable components of the pre-trained model and learning task-specific features. To gain insights into the distribution of generalizable components, we begin by analyzing the pre-trained weights through the lens of singular value decomposition. Building on these insights, we introduce Singular Value Decomposed Minor Components Adaptation (SoMA), an approach that selectively tunes minor singular components while keeping the residual parts frozen. SoMA effectively retains the generalization ability of the pre-trained model while efficiently acquiring task-specific skills. Moreover, we freeze domain-generalizable blocks and employ an annealing weight decay strategy, thereby achieving an optimal balance in the delicate trade-off between generalizability and discriminability. SoMA attains state-of-the-art results on multiple benchmarks that span both domain generalized semantic segmentation to domain generalized object detection. In addition, our methods introduce no additional inference overhead or regularization loss, maintain compatibility with any backbone or head, and are designed to be versatile, allowing easy integration into a wide range of tasks.

Dongheon Lee, Seokju Yun, Youngmin Ro

Recently, in the super-resolution (SR) domain, transformers have outperformed CNNs with fewer FLOPs and fewer parameters since they can deal with long-range dependency and adaptively adjust weights based on instance. In this paper, we demonstrate that CNNs, although less focused on in the current SR domain, surpass Transformers in direct efficiency measures. By incorporating the advantages of Transformers into CNNs, we aim to achieve both computational efficiency and enhanced performance. However, using a large kernel in the SR domain, which mainly processes large images, incurs a large computational overhead. To overcome this, we propose novel approaches to employing the large kernel, which can reduce latency by 86\% compared to the naive large kernel, and leverage an Element-wise Attention module to imitate instance-dependent weights. As a result, we introduce Partial Large Kernel CNNs for Efficient Super-Resolution (PLKSR), which achieves state-of-the-art performance on four datasets at a scale of $\times$4, with reductions of 68.1\% in latency and 80.2\% in maximum GPU memory occupancy compared to SRFormer-light.

Min-Jung Kim, Dongjin Kim, Seokju Yun et al.

Video editing has garnered increasing attention alongside the rapid progress of diffusion-based video generation models. As part of these advancements, there is a growing demand for more accessible and controllable forms of video editing, such as prompt-based editing. Previous studies have primarily focused on tasks such as style transfer, background replacement, object substitution, and attribute modification, while maintaining the content structure of the source video. However, more complex tasks, including the addition of novel objects and nonrigid transformations, remain relatively unexplored. In this paper, we present TV-LiVE, a Training-free and text-guided Video editing framework via Layerinformed Vitality Exploitation. We empirically identify vital layers within the video generation model that significantly influence the quality of generated outputs. Notably, these layers are closely associated with Rotary Position Embeddings (RoPE). Based on this observation, our method enables both object addition and non-rigid video editing by selectively injecting key and value features from the source model into the corresponding layers of the target model guided by the layer vitality. For object addition, we further identify prominent layers to extract the mask regions corresponding to the newly added target prompt. We found that the extracted masks from the prominent layers faithfully indicate the region to be edited. Experimental results demonstrate that TV-LiVE outperforms existing approaches for both object addition and non-rigid video editing. Project Page: https://emjay73.github.io/TV_LiVE/

Dongheon Lee, Seokju Yun, Youngmin Ro

In this paper, we tackle the high computational overhead of Transformers for efficient image super-resolution~(SR). Motivated by the observations of self-attention's inter-layer repetition, we introduce a convolutionized self-attention module named Convolutional Attention~(ConvAttn) that emulates self-attention's long-range modeling capability and instance-dependent weighting with a single shared large kernel and dynamic kernels. By utilizing the ConvAttn module, we significantly reduce the reliance on self-attention and its involved memory-bound operations while maintaining the representational capability of Transformers. Furthermore, we overcome the challenge of integrating flash attention into the lightweight SR regime, effectively mitigating self-attention's inherent memory bottleneck. We scale up the window size to 32$\times$32 with flash attention rather than proposing an intricate self-attention module, significantly improving PSNR by 0.31dB on Urban100$\times$2 while reducing latency and memory usage by 16$\times$ and 12.2$\times$. Building on these approaches, our proposed network, termed Emulating Self-attention with Convolution~(ESC), notably improves PSNR by 0.27 dB on Urban100$\times$4 compared to HiT-SRF, reducing the latency and memory usage by 3.7$\times$ and 6.2$\times$, respectively. Extensive experiments demonstrate that our ESC maintains the ability for long-range modeling, data scalability, and the representational power of Transformers despite most self-attention being replaced by the ConvAttn module.

Seokju Yun, Dongheon Lee, Youngmin Ro

Transformer, composed of self-attention and Feed-Forward Network, has revolutionized the landscape of network design across various vision tasks. While self-attention is extensively explored as a key factor in performance, FFN has received little attention. FFN is a versatile operator seamlessly integrated into nearly all AI models to effectively harness rich representations. Recent works also show that FFN functions like key-value memories. Thus, akin to the query-key-value mechanism within self-attention, FFN can be viewed as a memory network, where the input serves as query and the two projection weights operate as keys and values, respectively. Based on these observations, we hypothesize that the importance lies in query-key-value framework itself for competitive performance. To verify this, we propose converting self-attention into a more FFN-like efficient token mixer with only convolutions while retaining query-key-value framework, namely FFNification. Specifically, FFNification replaces query-key-value interactions with large kernel convolutions and adopts GELU activation function instead of softmax. The derived token mixer, FFNified attention, serves as key-value memories for detecting locally distributed spatial patterns, and operates in the opposite dimension to the ConvNeXt block within each corresponding sub-operation of the query-key-value framework. Building upon the above two modules, we present a family of Fast-Forward Networks (FFNet). Despite being composed of only simple operators, FFNet outperforms sophisticated and highly specialized methods in each domain, with notable efficiency gains. These results validate our hypothesis, leading us to propose MetaMixer, a general mixer architecture that does not specify sub-operations within the query-key-value framework.