Denoising Diffusion Probabilistic Models (DDPM)

Open In Colab

This is a PyTorch implementation/tutorial of the paper Denoising Diffusion Probabilistic Models.

In simple terms, we get an image from data and add noise step by step. Then We train a model to predict that noise at each step and use the model to generate images.

The following definitions and derivations show how this works. For details please refer to the paper.

Forward Process

The forward process adds noise to the data , for timesteps.

where is the variance schedule.

We can sample at any timestep with,

where and

Reverse Process

The reverse process removes noise starting at for time steps.

are the parameters we train.


We optimize the ELBO (from Jenson's inequality) on the negative log likelihood.

The loss can be rewritten as follows.

is constant since we keep constant.


The forward process posterior conditioned by is,

The paper sets where is set to constants or .


For given noise using

This gives,

Re-parameterizing with a model to predict noise

where is a learned function that predicts given .

This gives,

That is, we are training to predict the noise.

Simplified loss

This minimizes when and for discarding the weighting in . Discarding the weights increase the weight given to higher (which have higher noise levels), therefore increasing the sample quality.

This file implements the loss calculation and a basic sampling method that we use to generate images during training.

Here is the UNet model that gives and training code. This file can generate samples and interpolations from a trained model.

162from typing import Tuple, Optional
164import torch
165import torch.nn.functional as F
167from torch import nn
169from labml_nn.diffusion.ddpm.utils import gather

Denoise Diffusion

172class DenoiseDiffusion:
  • eps_model is model
  • n_steps is
  • device is the device to place constants on
177    def __init__(self, eps_model: nn.Module, n_steps: int, device: torch.device):
183        super().__init__()
184        self.eps_model = eps_model

Create linearly increasing variance schedule

187        self.beta = torch.linspace(0.0001, 0.02, n_steps).to(device)

190        self.alpha = 1. - self.beta

192        self.alpha_bar = torch.cumprod(self.alpha, dim=0)

194        self.n_steps = n_steps

196        self.sigma2 = self.beta

Get distribution

198    def q_xt_x0(self, x0: torch.Tensor, t: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:

gather and compute

208        mean = gather(self.alpha_bar, t) ** 0.5 * x0

210        var = 1 - gather(self.alpha_bar, t)

212        return mean, var

Sample from

214    def q_sample(self, x0: torch.Tensor, t: torch.Tensor, eps: Optional[torch.Tensor] = None):

224        if eps is None:
225            eps = torch.randn_like(x0)


228        mean, var = self.q_xt_x0(x0, t)

Sample from

230        return mean + (var ** 0.5) * eps

Sample from

232    def p_sample(self, xt: torch.Tensor, t: torch.Tensor):

246        eps_theta = self.eps_model(xt, t)


248        alpha_bar = gather(self.alpha_bar, t)

250        alpha = gather(self.alpha, t)

252        eps_coef = (1 - alpha) / (1 - alpha_bar) ** .5

255        mean = 1 / (alpha ** 0.5) * (xt - eps_coef * eps_theta)

257        var = gather(self.sigma2, t)

260        eps = torch.randn(xt.shape, device=xt.device)


262        return mean + (var ** .5) * eps

Simplified Loss

264    def loss(self, x0: torch.Tensor, noise: Optional[torch.Tensor] = None):

Get batch size

273        batch_size = x0.shape[0]

Get random for each sample in the batch

275        t = torch.randint(0, self.n_steps, (batch_size,), device=x0.device, dtype=torch.long)

278        if noise is None:
279            noise = torch.randn_like(x0)

Sample for

282        xt = self.q_sample(x0, t, eps=noise)


284        eps_theta = self.eps_model(xt, t)

MSE loss

287        return F.mse_loss(noise, eps_theta)