Reinforcing Few-step Generators via Reward-Tilted Distribution Matching

Researchers introduced RTDMD, a two-stage method for few-step image generation alignment.

It first distills generators with consistency-aware distribution matching, then fine-tunes them with reward optimization.

The method combines hybrid policy gradients and reduced-variance SubGRPO to better match human preferences.

RTDMD achieves state-of-the-art 4-step text-to-image results while keeping inference fast across SD3, SD3.5, and FLUX.2.