FVD: Inference-Time Alignment of Diffusion Models via Fleming-Viot Resampling
A novel resampling method that preserves trajectory diversity while optimizing diffusion model outputs.
We introduce Fleming-Viot Diffusion (FVD), an inference-time alignment method that resolves the diversity collapse commonly observed in Sequential Monte Carlo (SMC) based diffusion samplers. Existing SMC-based diffusion samplers often rely on multinomial resampling or closely rel...
ArXiv CS.AI
Source
Generated insight by OnAirToday AI
Fleming-Viot Diffusion addresses the fundamental trade-off between sample quality and diversity in generative models by replacing traditional resampling with a more stable stochastic mechanism. This advancement enables high-performance reward-based alignment without the computational overhead of value function training, marking a significant step toward efficient steerable generation.

Comments
Keep comments relevant and respectful. No links, advertising, self-promotion, or spam.