arXiv:2608. 19504v1 Announce Type: new Abstract: We propose a conditioning mechanism for diffusion models based on multi-speed joint diffusion of the target and the condition.
arXiv:2609.15702v1 Announce Type: new Abstract: Pretrained score-based diffusion models provide strong unconditional priors, yet enforcing measurement or physics consistency in inverse problems is of...
Efficient Covariance-Aware Vicinal Latent Diffusion (CAViLD) is a latent diffusion framework designed for continuous conditional image generation. It builds a purified, condition-aware latent space using density-adaptive vicinal learning, auxiliary semantic regularization, and NIQE-guided negative data augmentation. CAViLD incorporates an EMA teacher to internalize conditional guidance, enabling single-branch inference, and employs covariance-aware progressive distillation to reduce sampling steps to four NFEs, achieving a balance between fidelity, condition consistency, and efficiency on datasets such as RC-49, UTKFace, and Steering Angle.
arXiv:2609.14596v1 Announce Type: new Abstract: Training-free diffusion inverse solvers typically choose between local measurement guidance and costly clean-space posterior updates. Independent poste...
Inference-time alignment steers diffusion models toward downstream objectives without costly fine-tuning, but existing methods typically optimize expected cost (or reward) without explicitly controlli...
Score‑based diffusion models excel at inverse problems but suffer from slow inference and task‑specific hyperparameter tuning. This paper introduces principled reparameterizations that create invariances, enabling the reuse of the same hyperparameters across different tasks. Building on the RED‑diff framework, the authors present OptDiff, a streamlined pipeline that integrates convex optimization to accelerate inference, achieving notable speedups and higher image quality in reconstruction, deblurring, and super‑resolution experiments.
arXiv:2608. 05600v1 Announce Type: cross Abstract: Flow-based generative models are typically sampled by solving a deterministic ordinary differential equation (ODE), whereas online reinforcement learning requires stochastic rollouts for policy exploration and optimization.
Flow-based generative models are typically sampled by solving a deterministic ordinary differential equation (ODE), whereas online reinforcement learning requires stochastic rollouts for policy exploration and optimization. Existing GRPO methods for flow models therefore replace the inference-time ODE with a stochastic differential equation (SDE) during training.
arXiv:2606. 13240v1 Announce Type: cross Abstract: A key strength of diffusion models lies in their flexibility, since their outputs can be controlled at sampling time through guidance.
The paper investigates training diffusion models to sample from distributions defined by unnormalized densities or energy functions. It benchmarks various diffusion-structured inference techniques, including simulation-based variational methods and off-policy approaches such as continuous generative flow networks, highlighting their relative strengths and challenging some prior claims. Additionally, the authors introduce a new exploration strategy for off-policy methods that employs local search in the target space with a replay buffer, demonstrating improved sample quality across multiple target distributions.
The paper investigates when to activate classifier‑free guidance (CFG) in diffusion models by treating the diffusion trajectory as a time‑dependent information bottleneck. It shows that the usefulness of conditional information increases after a critical stage defined by a Fisher‑information balance, and proposes a training‑free dynamic CFG strategy that activates guidance based on a stochastic process of score statistics. Experiments across various architectures and tasks confirm that this adaptive rule improves effectiveness, robustness, and general applicability while offering a principled explanation for dynamic guidance.
arXiv:2607. 23488v1 Announce Type: new Abstract: Text-to-image diffusion models expose many inference-time sampling parameters, including prompts, negative prompts, classifier-free guidance scales, and noise schedules.
The paper introduces a statistical framework for diffusion transfer in generative modeling across related distributions. It analyzes a shared pooled score and a target‑specific offset, deriving convergence rates that separate the learning costs of these components. The authors prove that under certain regularity assumptions the offset remains bounded and jointly regular as diffusion time approaches zero, enabling polynomially vanishing lower time cutoffs and total‑variation guarantees for diffusion models as density estimators, with synthetic simulations and image benchmarks illustrating the approach.