PyTorchKR Papers

업데이트된 논문

총 9편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2607.02137v5cs.LG

Adaptive Reparametrized Time for Score-Based Diffusion Sampling

We study timestep allocation for score-based diffusion sampling, where the learned reverse-time dynamics is discretized on a finite grid. Uniform and hand-crafted schedules are standard, but their fixed, ad hoc prescriptions can be suboptimal. We propose Adaptive Reparameterized Time (ART), a continuous-time control formulation that learns a time change by treating sampling-clock speed as the control, so a uniform grid on the learned clock induces adaptive timesteps in diffusion time and a leading-order Euler error surrogate defines the allocation objective. To solve the resulting high-dimensional deterministic control problem, we introduce ART-RL, an auxiliary randomized Gaussian-policy formulation that turns schedule learning into continuous-time reinforcement learning. We prove ART-RL is equivalent to ART at optimality: the mean of the optimal Gaussian policy solves ART. We further establish policy evaluation and policy improvement characterizations and derive trajectory-based moment identities that yield implementable actor-critic updates. Experiments ranging from controlled low-dimensional settings to image generation show that ART-learned schedules, when plugged into existing samplers by changing only the timestep grid, consistently improve sample quality over strong baselines at matched budgets. The learned schedules also exhibit broad generalizability, transferring without retraining across sampling budgets, data sets, solvers, pipelines, and representation spaces.

arXiv:2605.11506v3cs.CV

역문제를 위한 확산 기반 최적화기의 원칙적 설계

Principled Design of Diffusion-based Optimizers for Inverse Problems

스코어 기반 확산 모델은 역문제에서 최첨단 성능을 달성하지만, 긴 추론 시간과 번거로운 하이퍼파라미터 튜닝으로 인해 실제 배포가 저해된다. 사전학습된 확산 모델은 재학습 없이 여러 태스크에 재사용할 수 있지만, 노이즈 스케줄과 사후 샘플링 가중치와 같은 추론 시점 하이퍼파라미터는 일반적으로 각 문제 설정마다 임시방편적인 조정이 필요하다. 우리는 불변성을 유도하는 원칙적 재매개화를 제안하여, 동일한 하이퍼파라미터를 재튜닝 없이 여러 문제에 걸쳐 재사용할 수 있게 한다. 또한 사후 샘플링을 최적화 문제로 재구성하는 RED-diff 프레임워크를 바탕으로, OptDiff 파이프라인을 추가로 개발한다. OptDiff는 추론을 가속하기 위한 볼록 최적화 도구의 통합을 용이하게 하는 간소화된 튜닝 프레임워크를 제공한다. 이미지 재구성, 디블러링, 초해상도에 대한 실험은 상당한 속도 향상과 향상된 이미지 품질을 보여준다.

arXiv:2609.36882v3cs.CV

적은 감독, 더 나은 일반화: 디퓨전으로 편집된 이미지에서의 약지도 가짜 영역 위치 추정

Less Supervision, Better Generalization: Weakly Supervised Fake Region Localization in Diffusion-Edited Images

Localizing AI-edited regions is essential for interpretable forensic analysis, but remains challenging due to subtle and spatially distributed artifacts that are misaligned with semantic or object boundaries. Existing approaches rely on pixel-level supervision from controlled editing pipelines, which is difficult to scale and can introduce misleading signals: artifacts frequently extend beyond annotated regions, while out-of-mask pixels are treated as authentic. This limits models' ability to capture transferable evidence and generalize across generators and datasets. To address these issues, we propose ReGFLoW, a Reconstruction-Guided Fake Localization framework under Weak supervision, which is the first weakly supervised approach for diffusion-edited fake region localization. ReGFLoW requires only real/fake labels at the image-level and uses diffusion reconstruction errors as dense spatial guidance to inject them into both feature and score spaces. Furthermore, by artifact-centric multiple instance learning, ReGFLoW utilizes localized diffusion evidence without relying on semantic-affinity or boundary-based pseudo-mask priors. Extensive experiments show competitive cross-generator localization, while ReGFLoW outperforms all evaluated fully supervised baselines when evaluation includes both partially edited and fully synthetic images and in cross-dataset tests, without target-domain adaptation.

arXiv:2606.17342v3cs.CV

디퓨전을 이용한 시각 텍스처 최대 엔트로피 모델 학습

Learning a Maximum Entropy Model for Visual Textures using Diffusion

Visual textures -- spatially homogeneous image regions containing repeated elements (e.g. a field of grass, the bark of a tree) -- are ubiquitous in visual scenes and provide important cues for recognizing and analyzing materials, surface geometry and objects. A number of existing texture models extract essential statistics from a single texture image, and can then generate high-quality samples that are visually similar to the original by matching these statistics. However, their statistics are either hand-designed or based on a network pretrained for another purpose (e.g., object recognition). Here, we develop the first principled method for unsupervised learning of a set of statistics that are used to constrain a maximum entropy probability model for texture. We leverage methods developed for generative diffusion models to derive training and sampling procedures, and compare these to the traditional method of sampling via matching the statistics. Despite the compactness of our trained model (512 statistics), it generates texture images whose quality is as good as or better than the current state-of-the-art model (~177k statistics), both visually and by quantitative perceptual scoring using DreamSim. Finally, we show that images generated from parameters lying along a line between the parameter vectors corresponding to two different reference textures are homogeneous textures with features that smoothly transition between those of the two references.

arXiv:2606.21036v2stat.ML

확산 기반 상태 공간 모델

Diffusion-Driven State Space Models

딥 상태 공간 모델(deep SSM)의 가우시안 잠재 전이 한계와 확산 모델(diffusion model)의 기저 역학 추론 부족을 해결하기 위해, 가우시안 전이를 확산 모델로 대체하는 DDSSM(Diffusion-Driven State Space Model)을 제안하고 순차 데이터에서 오토인코더와 확산 모델의 공동 학습(joint training) 방법을 제시한다. 다중모드 전이를 갖는 시뮬레이션 시계열에서 DDSSM이 최첨단 딥 SSM보다 적합과 예측에서 우수한 성능을 보였다.

arXiv:2606.18071v2cs.LG

Volterra Generative Models via Markovian Lifting

Diffusion-based generative models usually perturb data through Markov processes driven by Brownian motion. Volterra processes allow more flexible temporal dependence, but their non-Markovian structure complicates score learning and reverse-time simulation. We develop Volterra generative models and provide Markovian approximations. Gaussian conditioning then reduces the denoising problem to the original data dimension. We present the connections with variance-exploding and variance-preserving representations. Our framework also permits reuse of pretrained denoisers such as EDM2 networks. With one auxiliary factor, our models trained from scratch achieve FIDs of $0.547$ on MNIST and $1.860$ on conditional CIFAR-10. With autoguidance, ImageNet64 FID reaches $1.912$ after fine-tuning, while ImageNet512 FID reaches $1.903$ using frozen pretrained networks.

arXiv:2606.06179v3stat.ML

확산 모델은 그래디언트만 관측한다: 스코어 매칭 오류에 대한 기하학적 관점

Diffusion Models Observe Only Gradients: A Geometric Perspective on Score Matching Errors

Score-based diffusion models are typically trained by minimizing the $L^2$ score matching error, and standard theoretical analyses rely on this quantity to bound the sampling discrepancy between the learned and target distributions. Recent work has shown that the score admits an orthogonal decomposition into gradient and solenoidal components in which only the gradient part matters for exact sampling. We apply this decomposition to the score error of trained models and study its quantitative consequences. Since only the gradient component of the error enters the marginal Fokker-Planck dynamics, a model can incur arbitrarily large $L^2$ score error while perfectly matching the target. Starting from this observation, we present three results. First, we derive an upper bound on the Kullback-Leibler divergence that depends only on the observable gradient component of the error. It tightens the standard Girsanov bound for generic score networks and identifies that bound's looseness as the cost of working with path-space rather than marginal-space dynamics. Second, we prove that no non-trivial function of the $L^2$ score error can uniformly lower bound any divergence between the learned and target distributions. Third, we give an estimator of the gradient component, computable from denoising score matching residuals via a dual Sobolev identity, which empirically correlates substantially better with sample quality than the full $L^2$ error.

arXiv:2607.12121v3cs.DC

FlashDiff: 확산 모델 서빙을 위한 효율적인 지역 실행 및 스케줄링

FlashDiff: Efficient Regional Execution and Scheduling for Diffusion Model Serving

확산 모델은 현대의 이미지, 비디오, 오디오 생성의 핵심 백본이 되었지만, 이를 효율적으로 서비스하는 것은 여전히 난제로 남아 있다. 자기 회귀(autoregressive) 디코딩과 달리, 확산 추론(inference)은 많은 노이즈 제거(denoising) 단계에 걸쳐 고차원 공간 또는 시간 잠재 변수(latent)를 반복적으로 갱신한다. 이 전체 영역(all-region) 실행 패턴은 생성 지연 시간을 높이고 서빙 처리량을 제한한다. 기존의 다중 GPU 병렬화 방법은 단계당 연산을 줄일 수 있지만, 종종 상당한 활성화(activation) 교환 오버헤드를 유발하여 통신 비용이 병렬 실행의 이점을 상쇄하거나 심지어 상회하기도 한다. 본 논문은 적응형 지역 실행 및 스케줄링을 통해 추론 효율을 향상시키는 확산 서빙 시스템 FlashDiff를 제안한다. FlashDiff는 확산 정제(refinement)가 잠재 변수 영역이나 노이즈 제거 단계 전반에 걸쳐 균일하지 않다는 관찰에 기반한다. 서로 다른 영역은 종종 서로 다른 속도로 안정화되며, 인접한 단계들은 강한 시간적 상관관계를 보인다. FlashDiff는 이러한 특성을 활용하여 추가 정제가 필요한 영역만 선택적으로 실행하고, 그로 인해 발생하는 연산 여유를 동시 서빙 요청들에 재할당한다. FlashDiff는 세 가지 메커니즘으로 구성된다. 첫째, 초기 단계 어텐션(attention) 신호를 사용하여 잠재 표현을 일관된 실행 영역으로 분해함으로써 의미 구조를 보존하면서 세밀한 병렬성을 드러낸다. 둘째, 경량 런타임 컨트롤러를 사용하여 영역 활성도를 추정하고, 추가 정제가 출력 품질에 영향을 줄 가능성이 낮을 때 영향이 적은 갱신을 우회한다. 셋째, 친화도 인식(affinity-aware) 온라인 스케줄러를 적용하여 의존적인 영역을 공동 배치(co-locate)하고, GPU 간 잔여 부하를 균형 있게 분배하며, 회수된 연산 용량을 재사용하여 서빙 효율을 향상시킨다. 실제 이미지, 비디오, 오디오 워크로드에서 FlashDiff는 종단 간 서빙 지연 시간을 30-97% 줄이고 처리량을 1.2-2.2배 향상시킨다.

arXiv:2607.10951v2cs.LG

Sticky Jump Diffusions: A Unifying View of Masked, Continuous, and Hybrid Diffusion

We introduce sticky jump diffusions (SJDs), continuous-time Markov processes on $\mathbb{R}^d$ whose discrete anchors are token embeddings. In forward time, anchors release their mass at a hazard rate and the released mass diffuses in the continuous ambient space; time reversal couples a score-driven SDE with a sticky jump kernel whose rate and destination are fixed by flux balance with the forward law. We estimate the score and the per-anchor reverse hazards from a single denoising classifier via denoising hazard matching, the hazard analogue of denoising score matching, with simulation-free cross-entropy training. SJD recovers masked diffusion, continuous diffusion, and hybrid diffusion as limits. Its reversal explains features that each family treats as given: the mask of masked diffusion carries no evidence about the source token because the unsticking kernel of every anchor collapses to the same absorbing point; the terminal projection of continuous diffusion is required due to the absence of atoms in its forward marginal, without which flux balance yields no reverse jumps; and the update rules of hybrid diffusion (commit rate, destination, and drift) all follow from flux balance rather than from separate design. Beyond these limits, the unsticking kernel becomes a design space: a cross-position blending corrupts each position toward a blend of its neighbors' clean values or embeddings, turning dependency structure such as spatial locality or a constraint graph into an inductive bias of the corruption itself, and improves over the identity-kernel hybrid on CIFAR-10, Text8, and Sudoku.