Diffusion models accurately recover mixture weights from generated samples despite score function insensitivity.
problem Score-based generative models often fail to learn correct relative mode amplitudes (mixture weights) from generated samples.
method Relate diffusion score matching (DSM) loss to mixture weight estimation error, define diffusion score sensitivity index (DSSI), and prove its governing role in mixture weight recovery.
result Generated samples can accurately recover mixture weights from the DSM loss, even when the target score is insensitive to mixture weights.
Conservation laws improve diffusion model training by optimizing likelihood.
problem Training diffusion models with denoising objectives.
method Developed conservation laws based on GEXIT functions for memoryless noise processes.
result Unified characterization of diffusion model likelihood, reducing training to learning marginal posteriors.
Improved text-to-image and multimodal understanding through adaptive generation order optimization.
problem Determining optimal generation sequences in text-to-image synthesis and multimodal understanding.
method Introduced a learnable control module trained via Group Relative Policy Optimization (GRPO) to determine the generation order.
result Learning the control block substantially improves text-to-image alignment and multimodal understanding in DLMs.
DiPhon generates scalable graphs via diffusion on graphons.
problem Scaling diffusion models to large graphs.
method Formulated a continuous diffusion process on graphon space via Jacobi SDE, discretized for finite graphs.
result DiPhon matches the first moment of graphon dynamics and approximates the second moment.
New bounds close the score matching gap for diffusion models.
problem The difference between sample quality and score matching loss in diffusion models.
method Theoretical analysis of score matching gap, developing tighter bounds for KL divergence, reverse KL divergence, and Wasserstein distance.
result The quality of score approximation impacts closing the score matching gap for low noise scales.
Diffusion models don't overfit, contrary to expectations.
problem Understanding generalization in diffusion models.
method Fundamental impossibility results and analysis of score matching.
result Diffusion models exhibit classical U-shaped loss curve, not double descent.
Diffusion models adapt to low-dimensional data regardless of coefficient choices.
problem Understanding how diffusion models adapt to low-dimensional data structures.
method Analysis of diffusion models with flexible coefficient choices.
result Proven that O ~ ( k / ε ) \widetilde{O}(k/\varepsilon) O ( k / ε ) iterations suffice for accurate sampling in total variation distance. Replacing MSE with f-divergence in diffusion models improves robustness under data contamination.
problem Improving robustness of diffusion models under data contamination.
method Replacing MSE with f-divergence in diffusion models.
result Empirical improvement in performance under data contamination.
Introduce a variance-weighted batch distribution for diverse sampling in diffusion models.
problem Independent sampling in diffusion models.
method Introduce a variance-weighted batch distribution.
result Sampler with a transparent probabilistic target.
Combines deep state space models with diffusion models for better forecasting and capturing latent dynamics
problem Forecasting and capturing latent dynamics in time series
method DDSSM: Diffusion-driven state space model
result Empirically outperforms state-of-the-art deep SSM
Local data coverage governs memorization in diffusion models.
problem Memorization in diffusion models
method Derive a theoretical criterion based on local data coverage
result Predicts memorization based on density of training data in neighborhood and dataset size
Recursive training of generative models can lead to model collapse, and the recursion converges to a unique limiting distribution.
problem Model collapse in recursive training of generative models
method Recursive training on their own outputs
result Recursive training converges to a unique limiting distribution
Accelerating Speculative Diffusions via Block Verification
problem Adapting speculative decoding for continuous diffusion models
method Introducing a novel speculative sampling mechanism for diffusion models
result Improves acceptance rate and speeds up inference
Jeffrey guidance extends diffusion-model control to more complex applications.
problem Controlling diffusion models beyond simple cases like conditional sampling.
method Leveraging Jeffrey's rule of conditioning to update marginal distributions towards a target distribution.
result Significant reductions in FID on CIFAR-10 and FFHQ with Inception embeddings as the target.
Score matching errors are not sufficient for measuring diffusion model quality.
problem The L 2 L^2 L 2 score matching error is not a reliable measure of diffusion model performance. method Decomposed score errors into gradient and solenoidal components and analyzed their geometric properties.
result Only the gradient component of the score error affects the marginal distributional quality.
New geometric analysis shows L 2 L^2 L 2 score error is flawed for diffusion models.
problem Score matching errors in diffusion models do not fully capture distributional quality.
method Decomposed score errors into gradient and solenoidal components, focusing on gradient's role in Fokker-Planck dynamics.
result Only gradient component affects marginal distributional quality; solenoidal component is structurally invisible.
TabSODA improves imputation of surveys with skips and ordinal data.
problem Handling structural skips and ordinal responses in survey data.
method TabSODA uses an Elucidated Diffusion Model with skip pattern detection and ordinal awareness.
result TabSODA reduces ordinal missing-at-random (MACE) by up to 23.7% and improves categorical accuracy by up to 9%.
HyFAD improves time series imputation by combining time and frequency diffusion.
problem Improve time series imputation by handling frequency-sensitive denoising and balancing global and local dynamics.
method HyFAD is a hybrid time-frequency diffusion model with frequency-aware embedding, built on DDPM paradigm.
result HyFAD achieves state-of-the-art performance in time series imputation.
Develops a framework for distilling flow models from few steps.
problem Improving few-step sampling in diffusion models for better performance.
method Local approximation errors and dynamical amplification controlled through analytical tractability.
result Deep residual compositions efficiently approximate long-horizon transport with controlled global error.
AugMask trains diffusion models on incomplete tabular data by augmenting missing values and applying denoising supervision.
problem Training diffusion models on incomplete tabular data with missing values.
method AugMask uses stochastic augmentation and denoising supervision to adapt diffusion models to incomplete data.
result AugMask enables diffusion-based tabular generators to outperform specialized missing-aware baselines across various datasets and missingness regimes.
New risk bound for drift estimator in stochastic models.
problem Theoretical guarantees for drift estimation in stochastic differential equations.
method Derives an explicit risk bound using diffusion model theory.
result Explicit decomposition of risk into multiple sources of error.
Self-regulating annealing improves sampling from heavy-tailed datasets.
problem Sampling from heavy-tailed distributions using diffusion models.
method Proposed an SDE-based sampler with a state-dependent diffusion coefficient.
result State dependence induces a self-regulating annealing mechanism.
Unified framework for unlearning in diffusion models using KL divergence and likelihood constraints.
problem Removing undesirable data or concepts while preserving utility of pretrained models.
method Constrained optimization framework based on reverse and forward KL divergences, and likelihood constraints.
result Our KL-constrained approach achieves superior retention-unlearning tradeoffs compared to weight-based baselines.
Diffusion models learn multi-modal distributions with optimal efficiency.
problem Learning high-dimensional distributions with low-dimensional multi-modal structures.
method Score-based diffusion models, focusing on subgaussian distributions within subspaces.
result Diffusion models require O ~ ( ε − k ∨ 2 ) \widetilde{O}(\varepsilon^{-k \vee 2}) O ( ε − k ∨ 2 ) samples for 1-Wasserstein ε \varepsilon ε error, improving over prior guarantees. Latent diffusion improves robustness in missing data imputation.
problem Missing data imputation under MCAR corruption.
method Two-stage framework: VAE for latent feature learning, diffusion model in latent space.
result Latent diffusion maintains high quality and stability up to 50% missingness.
Modern Hopfield networks help prevent forgetting in generative models after task changes.
problem How to prevent forgetting in generative models after task changes.
method Introduce intrinsic forgetting as an increase in Hopfield energy after task change, analyze memory replay effectiveness, and validate predictions in experiments.
result High-energy, outlier-like samples are more forgettable than cluster-like samples, and energy-based selection of replay samples mitigates forgetting.
U-turn chains improve sampling from complex distributions.
problem Sampling from high-dimensional learned distributions.
method Iterative forward-backward diffusion steps with Metropolis-Hastings correction.
result Minimal U-turn dynamics exhibit phase transitions and layer-ordering inversion.
Develops a framework for multi-objective learning in diffusion models with limited labeled data.
problem Achieving good trade-offs in multi-objective learning with diffusion models requires a generalist model class with larger capacity than individual tasks.
method Proposes a two-stage training procedure: first fitting specialist models from limited paired data, then distilling them into a generalist model.
result Establishes generalization bounds showing the number of paired samples depends only on specialist model complexity.
Improves inference-time alignment for diffusion models without updating weights.
problem Aligning diffusion models without updating weights for high-reward outputs.
method Trust-Region Iterative Twisted Sequential Monte Carlo (TRI-TSMC) for variance reduction and efficiency.
result Improves primary alignment objectives on text generation tasks.
CARV reduces compute cost for downstream pipelines using diffusion models.
problem High variance in Monte Carlo estimators from diffusion models limits compute efficiency.
method CARV uses hierarchical MC estimation with amortized upstream computation and stratified-inverse-CDF.
result CARV delivers 2-3x effective compute multipliers without changing the objective.
This work extends Tweedie's formulae to non-Gaussian processes for better diffusion model generation.
problem Limited exploration of non-Gaussian diffusion models and corresponding Tweedie's formulae.
method Extended Tweedie's formulae to geometric Brownian motion, squared Bessel, and Cox-Ingersoll-Ross processes.
result Demonstrated potential of non-Gaussian models in image and financial time series generation.
This paper examines how Higher-Order Langevin Dynamics reduces memorization in diffusion models.
problem Memorization of training samples in diffusion models, violating copyright and privacy.
method Introduces Higher-Order Langevin Dynamics (HOLD) to regularize diffusion model trajectories.
result The dynamics of the data variable in HOLD are governed by a low-pass-filtered version of the learned score function, with smoothness increasing with model order.
A new particle filter uses diffusion models to improve state estimation from noisy data.
problem Sequentially estimating the state of a dynamical system from noisy and incomplete observations.
method Uses a diffusion model to simulate and predict system dynamics, incorporating noisy observations to refine predicted states.
result An unbiased particle filtering method that rigorously fuses observational data with diffusion model simulations.
Paper proposes MUCS for more reliable TDA in diffusion models.
problem Current TDA approaches lack reliability and robustness.
method Mirrored unlearning and noise-consistent skew (MUCS).
result MUCS outperforms existing methods on three datasets.
URGE improves diffusion model quality without gradients or Hessian.
problem Improving sample quality in diffusion models without gradient evaluations.
method Path-wise importance reweighting via Girsanov change of measure.
result URGE achieves better generation quality than existing methods.
StAD predicts divergence of diffusion and flow models without Jacobian computation.
problem Computing likelihood from diffusion and flow models is computationally expensive.
method Introduces StAD, a distillation method to predict divergence using Langevin-Stein operator.
result StAD predicts divergence with competitive variance and speed compared to existing methods.
A new method for generating samples without training, using smoothed score matching.
problem Generating samples efficiently and without training.
method Moment-matched score-smoothed overdamped Langevin dynamics (MM-SOLD).
result The method enables fast, robust, training-free sampling with competitive sample fidelity and diversity.
Study on reliability of latent reuse in diffusion models under distribution shift.
problem When can latent spaces from a source dataset be reused for a target dataset with different distributions?
method Considered a source-target setting with approximately low-dimensional datasets near different subspaces. Analyzed the target-domain score error due to principal-angle misalignment and target ambient noise.
result Latent reuse is reliable only if the source and target subspaces are close and the target ambient noise is not too amplified.
Improved sampling for Diffusion Models by accounting for covariance.
problem Sampling quality degradation in few-step Diffusion Models.
method Covariance-aware sampler using Tweedie's formula and Fourier-space decomposition.
result Consistently superior samples compared to state-of-the-art samplers.
New method designs joint initial noises for diffusion models to improve diversity and alignment.
problem Independent initial noises limit diversity in generated images.
method Coupling of initial noises, maintaining Gaussian distribution while allowing dependence.
result Repulsive Gaussian coupling improves diversity without increasing sampling cost.
New method improves sampling from score-based models by correcting bias.
problem Bias in sampling from score-based diffusion models.
method Metropolis-Hastings or Barker's accept-reject steps to correct bias, using the score function.
result Improves sample quality on synthetic and image datasets, yielding consistent gains in FID.
Unified framework reduces NFEs for inverse problems.
problem High computational costs and degraded reconstruction quality in existing LDM-based inverse solvers.
method Consistency Regularised Gradient Flows for posterior sampling and prompt optimization.
result Significantly reduced computational cost with state-of-the-art performance.
Flow Matching for count data improves sample quality and efficiency.
problem Mapping between count distributions across batches or time points in high-dimensional count data.
method count-FM, a flow-matching framework based on a continuous-time birth-death process with local unit jumps.
result count-FM achieves better sample quality than representative baselines while using fewer parameters.
TRACE improves conformal prediction for multi-dimensional outputs.
problem Challenges in constructing valid and informative conformal prediction regions for multi-dimensional outputs.
method TRACE uses transport alignment in diffusion and flow matching models to define nonconformity scores.
result TRACE yields valid and adaptive conformal prediction regions for multimodal and non-convex distributions.
Analyzes how class imbalance and heterogeneity affect diffusion model learning dynamics.
problem Understanding how class imbalance and heterogeneity impact the learning dynamics of diffusion models.
method Developed a high-dimensional analytical framework to study class-dependent learning in score-based diffusion models.
result Class variance is the primary determinant of learning order, favoring higher-variance classes; centroid geometry plays a secondary role.
Bayesian approach improves rain field reconstruction using CMLs and DMs.
problem Challenges in accurately reconstructing ground-level rainfall from CML path-integrated measurements.
method Bayesian inverse problem with Diffusion Models as priors.
result Improved performance in rainfall estimation compared to existing methods.
Unified framework for training diffusion and flow models to sample from target distributions.
problem Training diffusion and flow models to sample from target distributions defined by exponential tilting.
method Unified framework combining stochastic optimal control and non-equilibrium thermodynamics perspectives.
result Unified bias-variance decompositions and theoretical support for adjoint-based methods.
Consistency distillation reduces memorization in diffusion models without harming sample quality.
problem Understanding how distillation affects memorization in diffusion models.
method Analysis of consistency distillation in diffusion models using a random feature neural network model.
result Consistency distillation reduces memorization in diffusion models without harming sample quality.
A new diffusion model encodes causal structures for better interventional sampling and edge inference.
problem Lack of causal analysis in standard diffusion models.
method Causality-encoded diffusion framework that trains conditional models consistent with a directed acyclic graph.
result The method enables accurate interventional sampling and edge inference, with theoretical guarantees and practical applications.
New model simplifies symmetry handling in generative AI.
problem Symmetry handling in generative models for scientific tasks.
method Quotient-space diffusion models, viewing symmetry as quotient space.
result Improves performance over existing methods for molecular structure generation.
Unified framework for constrained diffusion models on nonconvex sets with efficient landing mechanism.
problem Efficiently modeling generative models under nonconvex constraints.
method Unified framework with overdamped and underdamped dynamics, landing mechanism.
result Significantly reduces computational cost while maintaining sample quality.
The paper analyzes the score field of diffusion models using Burgers dynamics.
problem Understanding the evolution of score fields in diffusion models.
method Analyzes the score field through Burgers-type evolution law for diffusion models.
result Identifies a universal \( anh\) interfacial term in the score field.
Sharp Lipschitz bounds for flow-matching and diffusion models with optimal sampling rates.
problem Establishing optimal Lipschitz regularity for flow-matching and diffusion models.
method Sharp Lipschitz regularity theory for flow-matching vector fields and diffusion-model scores.
result Achieves optimal sampling rate of d / N \sqrt{d}/N d / N for Euler-type samplers in dimension d d d . Researchers develop a method to generate diffusion-based samples from a tilted distribution.
problem Generating samples from a distribution that has been tilted by a parameter.
method Developed a plug-in estimator and proved Wasserstein bounds and TV-accuracy under certain conditions.
result The method is minimax-optimal and can be applied in various domains like finance and climate modeling.
DDCD uses diffusion models to learn causal structures from noisy data.
problem Scalability and stability issues in high-dimensional causal structure learning.
method Adaptive k-hop acyclicity constraint and denoising score matching objective of diffusion models.
result DDCD achieves competitive performance on synthetic and real-world data.
A new diffusion model generates structured tensors for high-dimensional data.
problem Generating a structured tensor with a target distribution.
method Tucker diffusion model with Tucker-Unet architecture.
result Generated tensors converge to the true data distribution at a rate dependent on tensor mode dimensions.
ML-EM method speeds up diffusion model sampling.
problem Efficiently sampling from complex diffusion models.
method Multilevel Euler-Maruyama method with UNet approximations.
result Polynomial speedup in sampling from diffusion models.
Diffusion models can generalize well even with coarse scores, thanks to the manifold hypothesis.
problem Understanding why diffusion models generate novel samples with coarse scores.
method Exploring the manifold hypothesis to explain diffusion model behavior.
result Diffusion models trained with coarse scores can achieve near-parametric rates of generalization, faster than estimating the full data distribution.
Study shows how diffusion models learn on low-dimensional manifolds.
problem Learning efficiency of diffusion models on manifolds.
method Analyzes denoising score matching with random feature neural networks.
result Sample complexity scales linearly with intrinsic dimension, not ambient dimension.
Entropy-based decoding improves DLM sampling efficiency.
problem Decoding strategy challenges in flexible DLMs.
method Entropy sum-based confidence-based decoding.
result Entropy sum-based decoding achieves ε \varepsilon ε -accuracy with O ~ ( H ( X 0 ) / ε ) \widetilde O(H(X_0)/\varepsilon) O ( H ( X 0 ) / ε ) iterations.