Objective option

Diffusion-loss autoregression

Autoregress over a sequence of continuous tokens, using a small per-token diffusion model as the loss instead of a categorical cross-entropy over a codebook.

Introduced by

MAR 2024-06

Used by (1)

MAR

Alternatives on this axis

Diffusion (epsilon/v/x0-prediction), Drifting field, Flow matching / rectified flow, Stochastic interpolant framework, Average (mean) velocity, Next-scale autoregression, Normalizing flow, Transition matching

← All Objective options