At image-diffusion scales the personal history is destroyed. The
scale knob keeps it recoverable.
Usage
noise_schedule(T, scale = 0.001, beta_min = 1e-04, beta_max = 0.02)
Arguments
- T
Number of diffusion steps.
- scale
Noise scale multiplier.
- beta_min
Lower endpoint of the unscaled linear schedule.
- beta_max
Upper endpoint of the unscaled linear schedule.
Value
A list with beta, alpha_bar, T,
scale, signal_retained, note.
Examples
s <- noise_schedule(T = 50)
c(min(s$alpha_bar), s$signal_retained)
#> [1] 0.9994976 0.9994976
