src.diffusion.ddim module

src.diffusion.ddim module#

class src.diffusion.ddim.DDIM(args: Namespace, eta=0.0)[source]#

Bases: DDPM

Denoising Diffusion Implicit Model (DDIM).

Inherits from DDPM. DDIM enables faster sampling with skipped steps while preserving generation quality through a non-Markovian reverse process.

__init__(args: Namespace, eta=0.0)[source]#

Initialize the DDIM sampler.

Parameters:

  • args (Namespace) – Configuration argument object.

  • eta (float, optional) – Hyperparameter controlling the stochasticity of the sampling process. - eta=0.0: deterministic sampling (Standard DDIM). - eta=1.0: DDPM-equivalent variance (Standard DDPM). Default is 0.0.

denoise(xt, denoise_t, x0=None, noise=None, stride=1)[source]#

DDIM reverse process: deterministically or semi-deterministically derive x_{t-stride} from x_t.

This method overrides the parent DDPM.denoise implementation and uses the DDIM update rule.

x_{t-1} = sqrt(alpha_bar_{t-1}) * “predicted x0” +

sqrt(1 - alpha_bar_{t-1} - sigma_t^2) * “predicted noise” + sigma_t * epsilon_t

Parameters:

  • xt (torch.FloatTensor) – Noisy data at the current timestep t.

  • denoise_t (torch.LongTensor) – Index of the current timestep t.

  • x0 (torch.FloatTensor, optional) – Model-predicted original data x0.

  • noise (torch.FloatTensor, optional) – Model-predicted noise epsilon. Note: exactly one of x0 and noise must be provided.

  • stride (int, optional) – Sampling step size used for acceleration. Default is 1.

Returns:

Denoised data at the previous step x_{t-stride}.

Return type:

torch.FloatTensor

Contents