Organizations: Department of Electrical and Computer Engineering, Ohio State Univeristy · Department of Computer Science and Information Systems, BITS Pilani · Department of Computer Science, University of Central Florida · Edwardson School of Industrial Engineering, Purdue University
We study the finite-time training and sampling error of conditional diffusion models trained by online stochastic-gradient denoising score matching. Rather than assuming access to an exact empirical or population risk minimizer, we analyze the score network produced after a finite number of stochastic-gradient updates under an explicit localized Polyak--Łojasiewicz (PL) condition, stochastic-gradient regularity, and a trajectory-containment assumption. Our analysis separates the network approximation complexity from the online training-sample budget and propagates both errors through the reverse diffusion process. Under anisotropic smoothness, a joint intrinsic-dimensional structure for the data and conditioning variables, sub-exponential tails, and explicit approximation-class compatibility, we obtain a conditional score approximation error of order
O(M−ηapp),ηapp=2αx+dx′+(αx/αy)dy′2αx,
where M denotes the approximation complexity, dx′ and dy′ are the intrinsic response and conditioning dimensions, and αx,αy are their corresponding smoothness exponents. For N=RB online denoising samples used by SGD, the finite-time optimization error is controlled explicitly in terms of the PL, smoothness, gradient-noise, and initialization constants. Combining these ingredients with reverse-process stability yields a total-variation bound that decomposes into terminal-initialization, time-discretization, approximation/localization, and finite-time optimization errors.
Figures & tables
Reference
Conditional
Intrinsic dimension
Gradient training
Sampling bound
Fu et al. (2024)
Yes
No
No
Yes
Tang et al. (2025)
Yes
Yes
No
Yes
Hu et al. (2025)
Yes
Yes
No
Yes
Li (2025)
Yes
No
No
Yes
Gaur et al. (2026)
No
No
Finite-time SGD
Yes
Han et al. (2024)
No
No
GD with early stopping
No (training only)
Table 1: Structural comparison with representative adjacent theory. “Gradient training” records whether a concrete finite-run gradient method is analyzed. The rows are not rate comparisons.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 1: Randomized-time score MSE versus online denoising budget N=RB for a single time-conditioned score network trained with fresh online SGD samples.
Figure 2: Randomized-time score MSE versus online budget N , with network width used as an operational proxy for approximation complexity M .
Figure 3: Final randomized-time score MSE versus parameter count, used as a practical proxy for the complexity index M .
Figure 4: Joint M -versus- N diagnostic. The heat map summarizes randomized-time score MSE over width and online denoising budget.
Figure 5: Randomized-time reference-subtracted evaluation loss along the online SGD trajectory.
Figure 6: Parameter displacement ∥θr−θ0∥2 along the online SGD trajectory.
Figure 7: Randomized-time score MSE for the fiber-compatible intrinsic-dimension generator. The ambient dimension is fixed while the joint intrinsic dimension is varied.
Figure 8: Standardized excess kurtosis for the Gaussian-matched and sub-exponential regimes. Larger Laplace components produce visibly heavier tails.
Figure 9: Randomized-time score MSE under online SGD for Gaussian-matched and sub-exponential conditional laws.
Conditional diffusion models generate or infer data by steering iterative denoising with observations, labels, text, physical constraints, or other inputs.
We establish finite-sample guarantees for conditional diffusion models learned from data and implemented with finitely many denoising steps. Our approach aligns the score-training objective with the sampling grid, allowing statistical estimation and numerical discretization errors to be analyzed together. Under a smooth conditional density model, we derive high-probability bounds on the generated distribution's error and quantify the required number of score evaluations. For every smoothness index β>0, a suitable grid yields averaged conditional Kullback--Leibler (KL) divergence Op(n−2β/(dx+dy+2β)) relative to the original target with or without early stopping. Both the full and early-stopped samplers attain average target total-variation (TV) error Op(n−β/(dx+dy+2β)). Here n is the sample size, dx,dy are the data and conditioning dimensions, and errors are averaged over the conditioning variable. The total-variation bound matches the rate established in existing conditional distribution-estimation theory, up to logarithmic factors.
Qiuyun Zhu
Department of Statistics and Actuarial Science University of Iowa
Despite the empirical success of score-based diffusion models, a complete theoretical understanding of how finite-sample learning, network parameterization, and numerical discretization jointly dictate generative quality remains underdeveloped. Existing sampling analyses often evaluate the generative performance conditional on an oracle score or a pre-specified error threshold. In this work, we establish a unified convergence and generalization framework for score-based diffusion models parameterized by practical ResNet-type architectures. We analyze the generalization and convergence properties from the practical finite-sample, discrete-time learning problem of the score function to the ideal continuous-time, population-level objective. Based on the generalization result of the learning problem of score function, we analyze the sampling process induced by the learned score function and provide an end-to-end total variation distance estimate for the generated terminal distribution. This estimate explicitly decomposes the overall generative error into four interpretable components: the truncation error of the forward process, the reverse-time discretization error, the generalization error incorporating both finite data and forward-time discretization, and the training optimization gap. Our results quantitatively characterize how the training sample size, temporal discretization grids, and optimization accuracy jointly control the final fidelity of samples generated by diffusion models.
Jinshu Huang, Yiming Jiang, Chunlin Wu
School of Mathematical Sciences, Nankai University, Tianjin, China
Score-based diffusion models are typically trained by minimizing the L2 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 L2 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 L2 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 L2 error.
Naïl B. Khelifa, Richard E. Turner, Ramji Venkataramanan