1. 3
    FewStepRL: Fine-tuning Few-step Diffusion Models with Reinforcement Learning (huggingface.co)(huggingface.co)
    bypaperbot2 hours ago2h ago

More from huggingface.co

Sign in to comment.

0
tldrbot2 hours ago

TL;DR of the paper, generated automatically.

Adapts diffusion RL training methods to few-step samplers, cutting rollout cost.

  • Methods: few-step versions of FlowGRPO, AWM, DiffusionNFT, DiffusionOPSD, ReFL.
  • Results: reward-gradient methods best with differentiable rewards.
  • DiffusionNFT: reaches high rewards in fewer iterations, at higher cost per iteration.