Distillation of Discrete Diffusion by Exact Conditional Distribution Matching
By: Yansong Gao, Yu Sun
Discrete diffusion models (DDMs) are a powerful class of generative models for categorical data, but they typically require many function evaluations for a single sample, making inference expensive. Existing acceleration methods either rely on approximate simulators, such as $τ$-leaping, or on distillation schemes that train new student models and auxiliary networks with proxy objectives. We propose a simple and principled distillation alternative based on \emph{conditional distribution matching}. Our key observation is that the reverse conditional distribution of clean data given a noisy state, $p_{0\mid t}(x_0 \mid x_t)$, admits a Markov decomposition through intermediate times and can be recovered from marginal density ratios and the known forward CTMC kernel. We exploit this structure to define distillation objectives that directly match conditional distributions between a pre-trained teacher and a low-NFE student, both for one-step and few-step samplers.
Similar Papers
Towards Overcoming Data Scarcity in Nuclear Energy: A Study on Critical Heat Flux with Physics-consistent Conditional Diffusion Model
Machine Learning (CS)
Creates fake nuclear power data for better predictions.
MissDDIM: Deterministic and Efficient Conditional Diffusion for Tabular Data Imputation
Artificial Intelligence
Fills in missing table data quickly and reliably.
One-step Diffusion Models with Bregman Density Ratio Matching
CV and Pattern Recognition
Makes AI create pictures much faster.