A learned distribution prior, predicted by a VAE-style decoder, lets flow-matching image models generate in fewer ODE steps than with a Gaussian prior.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
LeDiFlow: Learned Distribution-guided Flow Matching to Accelerate Image Generation
A learned distribution prior, predicted by a VAE-style decoder, lets flow-matching image models generate in fewer ODE steps than with a Gaussian prior.