Kang et al.
Single-stage RL method unifies preference alignment and few-step image generation
AdvFidelity uses an online adversarial discriminator to penalize low-level artifacts that task rewards miss, matching full-step RL performance with only 4–8 sampling steps.
Chinese Tech
Sicong Kang · Fangao Zeng · Yuhan Li · Hao Zhou · Mengfei Xu · Wenxiang Shang · +3 more
Alibaba · Shanghai Jiao Tong University
Research Digest··3 min read
The authors propose AdvFidelity, which integrates artifact-sensitive fidelity supervision into reinforcement learning (RL) through an online adversarial reward.
§