arXiv subjects

Sungrack Yun

Publications and source records attributed to Sungrack Yun.

At least 19 recordsLinked to original sources

arXiv2026

MoNe: Modular Neural Memory for Efficient Long Context Inference

We present MoNe, a lightweight modular neural memory that attaches to any frozen pretrained Transformer to enable long-context inference without retraining. MoNe reads context in fixed-size segments via test-time learning of fast-weight neural memory networks with layer-localized gradient updates; at inference, the memory generates keys and values from the query tokens alone, with no context tokens re-read. This two-phase design decouples inference cost from context length, achieving $O(N)$ preprocessing and $O(1)$ query cost with peak GPU memory that does not grow with $N$. At 128K tokens, MoNe reduces both compute and peak GPU memory by approximately 80% compared to ICL with only 6.4% parameter overhead. MoNe generalizes to context lengths far beyond the backbone's native window, achieving strong performance on needle-in-a-haystack and word extraction benchmarks from RULER, where ICL degrades sharply.

cs.AI↗

arXiv2026

Ar2Can: An Architect and an Artist Leveraging a Canvas for Multi-Human Generation

Despite recent advances in personalized image generation, existing models consistently fail to produce reliable multi-human scenes, often merging or losing facial identity. We present Ar2Can, a novel two-stage framework that disentangles spatial planning from identity rendering for multi-human generation. The Architect predicts structured layouts, specifying where each person should appear. The Artist then synthesizes photorealistic images, guided by a spatially-grounded face matching reward that combines Hungarian spatial alignment with identity similarity. This approach ensures faces are rendered at correct locations and faithfully preserve reference identities. We develop two Architect variants, seamlessly integrated with our diffusion-based Artist model. This is optimized via Group Relative Policy Optimization (GRPO) using compositional rewards for count accuracy, image quality, and identity matching. Evaluated on the MultiHuman-Testbench, Ar2Can achieves substantial improvements in both count accuracy and identity preservation, while maintaining high perceptual quality. Notably, our method achieves these results using primarily synthetic data, without requiring real multi-human images. Project page: https://qualcomm-ai-research.github.io/ar2can/.

cs.CV↗

arXiv2026

Memory-Efficient Fine-Tuning Diffusion Transformers via Dynamic Patch Sampling and Block Skipping

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

cs.CV↗

arXiv2026

MultiHuman-Testbench: Benchmarking Image Generation for Multiple Humans

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

cs.CV↗

arXiv2025

Memory-Efficient Personalization of Text-to-Image Diffusion Models via Selective Optimization Strategies

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

cs.CV↗

arXiv2025

Steering Guidance for Personalized Text-to-Image Diffusion Models

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

cs.CV↗

arXiv2025

From Wardrobe to Canvas: Wardrobe Polyptych LoRA for Part-level Controllable Human Image Generation

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

cs.CV↗

arXiv2025

ConsNoTrainLoRA: Data-driven Weight Initialization of Low-rank Adapters using Constraints

Foundation models are pre-trained on large-scale datasets and subsequently fine-tuned on small-scale datasets using parameter-efficient fine-tuning (PEFT) techniques like low-rank adapters (LoRA). In most previous works, LoRA weight matrices are randomly initialized with a fixed rank across all attachment points. In this paper, we improve convergence and final performance of LoRA fine-tuning, using our proposed data-driven weight initialization method, ConsNoTrainLoRA (CNTLoRA). We express LoRA initialization as a domain shift problem where we use multiple constraints relating the pre-training and fine-tuning activations. By reformulating these constraints, we obtain a closed-form estimate of LoRA weights that depends on pre-training weights and fine-tuning activation vectors and hence requires no training during initialization. This weight estimate is decomposed to initialize the up and down matrices with proposed flexibility of variable ranks. With the proposed initialization method, we fine-tune on downstream tasks such as image generation, image classification and image understanding. Both quantitative and qualitative results demonstrate that CNTLoRA outperforms standard and data-driven weight initialization methods. Extensive analyses and ablations further elucidate the design choices of our framework, providing an optimal recipe for faster convergence and enhanced performance.

cs.CV↗

arXiv2025

Tripartite Weight-Space Ensemble for Few-Shot Class-Incremental Learning

Few-shot class incremental learning (FSCIL) enables the continual learning of new concepts with only a few training examples. In FSCIL, the model undergoes substantial updates, making it prone to forgetting previous concepts and overfitting to the limited new examples. Most recent trend is typically to disentangle the learning of the representation from the classification head of the model. A well-generalized feature extractor on the base classes (many examples and many classes) is learned, and then fixed during incremental learning. Arguing that the fixed feature extractor restricts the model's adaptability to new classes, we introduce a novel FSCIL method to effectively address catastrophic forgetting and overfitting issues. Our method enables to seamlessly update the entire model with a few examples. We mainly propose a tripartite weight-space ensemble (Tri-WE). Tri-WE interpolates the base, immediately previous, and current models in weight-space, especially for the classification heads of the models. Then, it collaboratively maintains knowledge from the base and previous models. In addition, we recognize the challenges of distilling generalized representations from the previous model from scarce data. Hence, we suggest a regularization loss term using amplified data knowledge distillation. Simply intermixing the few-shot data, we can produce richer data enabling the distillation of critical knowledge from the previous model. Consequently, we attain state-of-the-art results on the miniImageNet, CUB200, and CIFAR100 datasets.

cs.LG↗

arXiv2024

Hollowed Net for On-Device Personalization of Text-to-Image Diffusion Models

Recent advancements in text-to-image diffusion models have enabled the personalization of these models to generate custom images from textual prompts. This paper presents an efficient LoRA-based personalization approach for on-device subject-driven generation, where pre-trained diffusion models are fine-tuned with user-specific data on resource-constrained devices. Our method, termed Hollowed Net, enhances memory efficiency during fine-tuning by modifying the architecture of a diffusion U-Net to temporarily remove a fraction of its deep layers, creating a hollowed structure. This approach directly addresses on-device memory constraints and substantially reduces GPU memory requirements for training, in contrast to previous methods that primarily focus on minimizing training steps and reducing the number of parameters to update. Additionally, the personalized Hollowed Net can be transferred back into the original U-Net, enabling inference without additional memory overhead. Quantitative and qualitative analyses demonstrate that our approach not only reduces training memory to levels as low as those required for inference but also maintains or improves personalization performance compared to existing methods.

cs.CV↗

arXiv2024

FedHide: Federated Learning by Hiding in the Neighbors

We propose a prototype-based federated learning method designed for embedding networks in classification or verification tasks. Our focus is on scenarios where each client has data from a single class. The main challenge is to develop an embedding network that can distinguish between different classes while adhering to privacy constraints. Sharing true class prototypes with the server or other clients could potentially compromise sensitive information. To tackle this issue, we propose a proxy class prototype that will be shared among clients instead of the true class prototype. Our approach generates proxy class prototypes by linearly combining them with their nearest neighbors. This technique conceals the true class prototype while enabling clients to learn discriminative embedding networks. We compare our method to alternative techniques, such as adding random Gaussian noise and using random selection with cosine similarity constraints. Furthermore, we evaluate the robustness of our approach against gradient inversion attacks and introduce a measure for prototype leakage. This measure quantifies the extent of private information revealed when sharing the proposed proxy class prototype. Moreover, we provide a theoretical analysis of the convergence properties of our approach. Our proposed method for federated learning from scratch demonstrates its effectiveness through empirical results on three benchmark datasets: CIFAR-100, VoxCeleb1, and VGGFace2.

cs.LG↗

arXiv2024

Feature Diversification and Adaptation for Federated Domain Generalization

Federated learning, a distributed learning paradigm, utilizes multiple clients to build a robust global model. In real-world applications, local clients often operate within their limited domains, leading to a `domain shift' across clients. Privacy concerns limit each client's learning to its own domain data, which increase the risk of overfitting. Moreover, the process of aggregating models trained on own limited domain can be potentially lead to a significant degradation in the global model performance. To deal with these challenges, we introduce the concept of federated feature diversification. Each client diversifies the own limited domain data by leveraging global feature statistics, i.e., the aggregated average statistics over all participating clients, shared through the global model's parameters. This data diversification helps local models to learn client-invariant representations while preserving privacy. Our resultant global model shows robust performance on unseen test domain data. To enhance performance further, we develop an instance-adaptive inference approach tailored for test domain data. Our proposed instance feature adapter dynamically adjusts feature statistics to align with the test input, thereby reducing the domain gap between the test and training domains. We show that our method achieves state-of-the-art performance on several domain generalization benchmarks within a federated learning setting.

cs.LG↗

arXiv2023

Label Shift Adapter for Test-Time Adaptation under Covariate and Label Shifts

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

cs.CV↗

arXiv2023

Progressive Random Convolutions for Single Domain Generalization

Single domain generalization aims to train a generalizable model with only one source domain to perform well on arbitrary unseen target domains. Image augmentation based on Random Convolutions (RandConv), consisting of one convolution layer randomly initialized for each mini-batch, enables the model to learn generalizable visual representations by distorting local textures despite its simple and lightweight structure. However, RandConv has structural limitations in that the generated image easily loses semantics as the kernel size increases, and lacks the inherent diversity of a single convolution operation. To solve the problem, we propose a Progressive Random Convolution (Pro-RandConv) method that recursively stacks random convolution layers with a small kernel size instead of increasing the kernel size. This progressive approach can not only mitigate semantic distortions by reducing the influence of pixels away from the center in the theoretical receptive field, but also create more effective virtual domains by gradually increasing the style diversity. In addition, we develop a basic random convolution layer into a random convolution block including deformable offsets and affine transformation to support texture and contrast diversification, both of which are also randomly initialized. Without complex generators or adversarial learning, we demonstrate that our simple yet effective augmentation strategy outperforms state-of-the-art methods on single domain generalization benchmarks.

cs.CV↗

arXiv2022

Improving Test-Time Adaptation via Shift-agnostic Weight Regularization and Nearest Source Prototypes

This paper proposes a novel test-time adaptation strategy that adjusts the model pre-trained on the source domain using only unlabeled online data from the target domain to alleviate the performance degradation due to the distribution shift between the source and target domains. Adapting the entire model parameters using the unlabeled online data may be detrimental due to the erroneous signals from an unsupervised objective. To mitigate this problem, we propose a shift-agnostic weight regularization that encourages largely updating the model parameters sensitive to distribution shift while slightly updating those insensitive to the shift, during test-time adaptation. This regularization enables the model to quickly adapt to the target domain without performance degradation by utilizing the benefit of a high learning rate. In addition, we present an auxiliary task based on nearest source prototypes to align the source and target features, which helps reduce the distribution shift and leads to further performance improvement. We show that our method exhibits state-of-the-art performance on various standard benchmarks and even outperforms its supervised counterpart.

cs.CV↗

arXiv2022

Domain Agnostic Few-shot Learning for Speaker Verification

Deep learning models for verification systems often fail to generalize to new users and new environments, even though they learn highly discriminative features. To address this problem, we propose a few-shot domain generalization framework that learns to tackle distribution shift for new users and new domains. Our framework consists of domain-specific and domain-aggregation networks, which are the experts on specific and combined domains, respectively. By using these networks, we generate episodes that mimic the presence of both novel users and novel domains in the training phase to eventually produce better generalization. To save memory, we reduce the number of domain-specific networks by clustering similar domains together. Upon extensive evaluation on artificially generated noise domains, we can explicitly show generalization ability of our framework. In addition, we apply our proposed methods to the existing competitive architecture on the standard benchmark, which shows further performance improvements.

cs.SD↗

arXiv2021

Distribution Estimation to Automate Transformation Policies for Self-Supervision

In recent visual self-supervision works, an imitated classification objective, called pretext task, is established by assigning labels to transformed or augmented input images. The goal of pretext can be predicting what transformations are applied to the image. However, it is observed that image transformations already present in the dataset might be less effective in learning such self-supervised representations. Building on this observation, we propose a framework based on generative adversarial network to automatically find the transformations which are not present in the input dataset and thus effective for the self-supervised learning. This automated policy allows to estimate the transformation distribution of a dataset and also construct its complementary distribution from which training pairs are sampled for the pretext task. We evaluated our framework using several visual recognition datasets to show the efficacy of our automated transformation policy.

cs.CV↗

arXiv2021

Federated Learning of User Verification Models Without Sharing Embeddings

We consider the problem of training User Verification (UV) models in federated setting, where each user has access to the data of only one class and user embeddings cannot be shared with the server or other users. To address this problem, we propose Federated User Verification (FedUV), a framework in which users jointly learn a set of vectors and maximize the correlation of their instance embeddings with a secret linear combination of those vectors. We show that choosing the linear combinations from the codewords of an error-correcting code allows users to collaboratively train the model without revealing their embedding vectors. We present the experimental results for user verification with voice, face, and handwriting data and show that FedUV is on par with existing approaches, while not sharing the embeddings with other users or the server.

cs.LG↗