Deterministic Denoising Algorithm Proposed for Discrete Diffusion Models

Global: Deterministic Denoising Algorithm Proposed for Discrete Diffusion Models

Researchers Hideyuki Suzuki and Hiroshi Yamashita introduced a deterministic denoising algorithm for discrete-state diffusion models, according to a paper posted on arXiv on September 25, 2025 and revised on December 26, 2025. The method replaces the stochastic reverse diffusion process with a deterministic approach that employs a variant of the herding algorithm exhibiting weakly chaotic dynamics.

Algorithm Overview

The proposed technique models the generative reverse process as a Markov chain, thereby eliminating the need for retraining or continuous state embeddings typically required in stochastic diffusion frameworks. By derandomizing state transitions, the algorithm ensures reproducible outputs while maintaining the theoretical foundations of diffusion modeling.

Performance Gains

Experimental results reported in the paper indicate consistent improvements in both computational efficiency and sample quality across text and image generation tasks. The authors attribute these gains to the reduced variance inherent in deterministic transitions, which streamlines inference without sacrificing fidelity.

Implications for Generative Modeling

The findings suggest that deterministic reverse processes, long established in continuous diffusion models, can be effectively extended to discrete state spaces. This extension may simplify the deployment of diffusion-based generative systems by lowering randomness-related overhead and easing reproducibility concerns.

Publication Details

The work is classified under Machine Learning (cs.LG) and Chaotic Dynamics (nlin.CD) and is accessible via the arXiv repository with DOI https://doi.org/10.48550/arXiv.2509.20896. The authors anticipate that their approach will stimulate further research into deterministic methods for discrete generative modeling.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.