LogSNRImportanceSampling<T>

Models & Types · AiDotNet.Diffusion.Schedulers.NoiseSchedules

Log-SNR importance sampling for efficient timestep selection during diffusion training.

For Beginners

During training, some timesteps are more important than others for the model to learn from. This utility picks timesteps that give the model the most useful learning signal, making training more efficient.

How It Works

Instead of uniformly sampling timesteps during training, samples proportionally to the gradient magnitude at each timestep. Approximates this by sampling from a distribution that is uniform in log-SNR space, focusing training on the most informative timesteps.

Reference: Hang et al., “Efficient Diffusion Training via Min-SNR Weighting Strategy”, ICCV 2023

Constructors

Constructor Summary
LogSNRImportanceSampling()

Methods

Method Summary
int[] SampleTimesteps(int batchSize, Vector<T> alphasCumprod, Random random) Samples timesteps with importance weighting based on log-SNR distribution.