Ancestral Sampling
A stochastic sampling method commonly used in discrete diffusion models, which samples step by step according to the transition probabilities of the reverse diffusion process. It guarantees correct distributional results but produces random trajectories. Ancestral sampling is a standard approach for generating samples from discrete diffusion models by sequentially applying the reverse process.
Timeline (last 90 days)
Ancestral sampling is the most commonly used sampling method for discrete diffusion models; it strictly guarantees that generated samples follow the model distribution, but requires a number of iterations comparable to the number of diffusion steps used in training (hundreds to thousands of steps), and its trajectory is stochastic.
Because the ancestral sampling trajectory is stochastic, it cannot, as in continuous diffusion, be compressed via consistency distillation from multi-step sampling into a few-step deterministic mapping.
All Facts (2)
Ancestral sampling is the most commonly used sampling method for discrete diffusion models; it strictly guarantees that generated samples follow the model distribution, but requires a number of iterations comparable to the number of diffusion steps used in training (hundreds to thousands of steps), and its trajectory is stochastic.
50%UnverifiedBecause the ancestral sampling trajectory is stochastic, it cannot, as in continuous diffusion, be compressed via consistency distillation from multi-step sampling into a few-step deterministic mapping.
50%