Struct DdimScheduler 

pub struct DdimScheduler { /* private fields */ }

DDIM scheduler state.

Implementations

impl DdimScheduler

pub fn new(num_train_timesteps: usize, prediction_type: PredictionType) -> Self

Create a new DDIM scheduler.

Uses a “scaled linear” beta schedule matching SD 2.1 defaults: beta_start=0.00085, beta_end=0.012, 1000 training steps.

pub fn set_timesteps(&mut self, num_inference_steps: usize)

Configure evenly-spaced timesteps for a given number of inference steps.

pub fn timesteps(&self) -> &[usize]

Return the current list of timesteps (descending).

pub fn step( &self, model_output: &Tensor, t: usize, sample: &Tensor, ) -> Result<Tensor>

Perform one DDIM step (deterministic, η=0).

• model_output: the raw network prediction at timestep t.
• t: current timestep index.
• sample: the current noisy latent x_t.

Returns the denoised latent x_{t-1}.

pub fn add_noise( &self, original: &Tensor, noise: &Tensor, timestep: usize, ) -> Result<Tensor>

Add noise to latents for a given timestep (forward diffusion process).

x_t = sqrt(α_t) · x_0 + sqrt(1-α_t) · noise

pub fn timestep_tensor( &self, t: usize, batch_size: usize, device: &Device, ) -> Result<Tensor>

Create a tensor of timestep values on the given device.

Trait Implementations

impl Debug for DdimScheduler

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.