Entity · paper

Score Accuracy Along the Forward Diffusion Does Not Certify Numerical Stability in Diffusion Sampling

paperactiveprovisionalscore-accuracy-along-the-forward-diffusion-does-not-certify-numerical-stability-in-diffusion-sampling-fa44bc59·1 events·first seen 8h ago

Aliases: Score Accuracy Along the Forward Diffusion Does Not Certify Numerical Stability in Diffusion Sampling

Co-occurring entities

DiT Euler-Maruyama discretization

More like this (12)

LESS: Mutual-Stability Sampling for Diffusion Language Models diffusion posterior sampling What Does a Discrete Diffusion Model Learn?Kolmogorov Regression for Robust Diffusion Policies Selective Timestep Weighting and Advantage-Based Replay for Sample-Efficient Diffusion RLHF A Diffusion Approximation for Temporal-Difference Learning with Linear Features under Markovian Noise Denoising Diffusion Probabilistic Models discrete diffusion models Gibbs-Accelerated Discrete Diffusion (GADD)Survival Diffusion Probabilistic Model (SDPM)Dynamic-Probabilistic Consistency Gap Self-Augmenting Retrieval for Diffusion Language Models

Recent events (1)

5arXiv · cs.LG·8h ago·source ↗

Score accuracy in forward diffusion does not guarantee numerical stability in reverse sampling

A new arXiv paper demonstrates that small forward-marginal L2 error in score matching does not certify numerical stability for discretized reverse-time diffusion samplers. The authors construct adversarial examples where Euler-Maruyama discretizations converge weakly yet all Wasserstein distances diverge, and show this failure can occur within a fixed finite neural architecture. A positive result is also provided: projecting the learned denoiser onto a bounded convex set containing the data support restores Wasserstein convergence under mild regularity. Experiments with a DiT-style network confirm the instability and its suppression via denoiser projection.

Evaluation and Benchmarking Score Accuracy Along the Forward Diffusion Does Not Certify Numerical Stability in Diffusion Sampling DiT Euler-Maruyama discretization