ZeroTerminalSNRSchedule<T>
Models & Types · AiDotNet.Diffusion.Schedulers.NoiseSchedules
Zero Terminal SNR noise schedule ensuring signal-to-noise ratio reaches exactly zero at the final timestep.
For Beginners
This fixes a common issue in diffusion training where the model never sees completely noisy images. By ensuring SNR reaches zero, the model learns to generate from pure noise, improving overall sample quality.
How It Works
Standard beta schedules often have non-zero SNR at the final timestep, meaning the model never sees pure noise during training. Zero Terminal SNR rescales the schedule so that alpha_cumprod[T] = 0, ensuring the final timestep is pure noise.
Reference: Lin et al., “Common Diffusion Noise Schedules and Sample Steps are Flawed”, WACV 2024
Constructors
| Constructor | Summary |
|---|---|
ZeroTerminalSNRSchedule() |
Methods
| Method | Summary |
|---|---|
Vector<T> Apply(Vector<T> alphasCumprod) |
Rescales alpha cumulative products to enforce zero terminal SNR. |