Denoising Diffusion Implicit Models (Song et al.)

May 15, 2026

💡

Takeaways:

1.DDPM training objective doesn't uniquely determine the reverse sampling process

2.Many reverse processes can share the same training loss:

-stochastic (DDPM),

-partially deterministic,

-fully deterministic.

3.Replace stochastic reverse diffusion

x_{t-1} \sim p(x_{t-1} | x_t)

with

x_{t-1} = f(x_t, \hat{\epsilon}_\theta)

1.smooth latent trajectory, ODE-like

2.sample with fewer steps, much faster than DDPM