The 'nice property' lets you sample any diffusion timestep directly from the original image without iterating the forward process

A key mathematical convenience in DDPM is that, because sums of Gaussians are Gaussian, you can jump to any noise level t in one step: q(x_t | x_0) = N(x_t; sqrt(alpha_bar_t) * x_0, (1 - alpha_bar_t) * I), where alpha_bar_t is the cumulative product of (1 - beta_s) from s=1 to t. This means you do not need to repeatedly apply the forward process to generate a training example at timestep t. During training, you randomly sample a timestep t, use this formula to corrupt x_0 to x_t, and then ask the network to predict the noise. This makes the training loop simple and allows random-timestep sampling for SGD.

Examples

In code: q_sample(x_start, t, noise) computes sqrt_alphas_cumprod_t * x_start + sqrt_one_minus_alphas_cumprod_t * noise in one shot. No loop needed.

Assessment

Given alpha_bar_5 = 0.8, write the formula for sampling x_5 from x_0 and a noise sample epsilon. Explain why this avoids iterating the forward process.