Noise Schedule
The sequence of noise levels added to (or removed from) an image during diffusion - how much to corrupt at each step.
Think of it like
A recipe: "First scatter a bit of flour (10% noise), then more (20% noise), then keep stirring. Now reverse: pour out flour, then a bit less."
Example
A linear schedule: noise increases from 0 to 1 over 1000 steps. A cosine schedule: slow at first, accelerating in the middle. DDPM uses a specific b sequence.
How it actually works
The schedule (b0, b1, ..., bt) controls how much Gaussian noise is added at each step. Too fast: model never learns to denoise gradual changes. Too slow: needs many steps. Cosine and linear schedules are common. Variance schedule also affects training - whether the model predicts noise, the mean, or the variance. Choice impacts convergence speed and final quality. Recent work (EDM, v-prediction) shows better schedules and parameterizations.
For product teams
Tuning the schedule speeds up generation (fewer steps) or improves quality (better learned schedule).
For engineers
Typically defined by b values. Common: linear b in [0.0001, 0.02], cosine b in [0.0001, 0.02].
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome