VarDiU optimizes a variational upper bound on the diffusive KL divergence with an unbiased gradient estimator, improving one-step generation on a 2D 40-Gaussian toy benchmark compared with Diff-Instruct.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
VarDiU: A Variational Diffusive Upper Bound for One-Step Diffusion Distillation
VarDiU optimizes a variational upper bound on the diffusive KL divergence with an unbiased gradient estimator, improving one-step generation on a 2D 40-Gaussian toy benchmark compared with Diff-Instruct.