AI Diffusion Model Based on Stochastic Differential Equations: Noise Scheduling Variable Discrete Approximation

Recently, diffusion probabilistic models have shown great success in image synthesis, but their performance relies heavily on the pre-defined noise schedules, which have a direct impact on the convergence speed, stability, and quality of the generated images. Moreover, the traditional score-based SDE model is very sensitive to the choice of the step size in the numerical integration process. To overcome the above limitations, a variable discrete noise scheduling method combined with a Fully Connected Neural Network (FC-NN) denoiser is proposed in the SDE diffusion model. Experimental analysis on the CIFAR-10 dataset shows that the proposed Variable SDE model results in a final FID value of 8.7, which is better than DDPM (11.5), DDIM (10.9), and Score SDE (10.2), and an Inception Score of 9.4, which is better than 8.6 for the fixed scheduling approach. The proposed approach results in a reduction of convergence time by 20-25% and peak discretization error by 30-35%. The t-SNE analysis shows that the overlap between the real and generated distributions is good, and there is no mode collapse. The proposed approach provides faster convergence, better numerical stability, and improved generative quality compared to existing standard diffusion models.

Authors

Institutions

Publication Details

Journal
ICST Transactions on Scalable Information Systems
Published
2026-11-13
DOI
https://doi.org/10.4108/eetsis.12089
Primary Topic
Generative Adversarial Networks and Image Synthesis
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

AI Diffusion Model Based on Stochastic Differential Equations: Noise Scheduling Variable Discrete Approximation

Dongqing Du, Yan Liu
2 links
ICST Transactions on Scalable Information Systems
Generative Adversarial Networks and Image Synthesis
article

AI Diffusion Model Based on Stochastic Differential Equations: Noise Scheduling Variable Discrete Approximation

Dongqing Du, Yan Liu
article en

Abstract

Recently, diffusion probabilistic models have shown great success in image synthesis, but their performance relies heavily on the pre-defined noise schedules, which have a direct impact on the convergence speed, stability, and quality of the generated images. Moreover, the traditional score-based SDE model is very sensitive to the choice of the step size in the numerical integration process. To overcome the above limitations, a variable discrete noise scheduling method combined with a Fully Connected Neural Network (FC-NN) denoiser is proposed in the SDE diffusion model. Experimental analysis on the CIFAR-10 dataset shows that the proposed Variable SDE model results in a final FID value of 8.7, which is better than DDPM (11.5), DDIM (10.9), and Score SDE (10.2), and an Inception Score of 9.4, which is better than 8.6 for the fixed scheduling approach. The proposed approach results in a reduction of convergence time by 20-25% and peak discretization error by 30-35%. The t-SNE analysis shows that the overlap between the real and generated distributions is good, and there is no mode collapse. The proposed approach provides faster convergence, better numerical stability, and improved generative quality compared to existing standard diffusion models.

ICST Transactions on Scalable Information SystemsVol. 13(8)
Wuhu University (CN), Jiangsu Union Technical Institute (CN)
American Indian Graduate Center, Government of Jiangsu Province
Generative Adversarial Networks and Image Synthesis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon