DanceOPD: On-Policy Generative Field Distillation
The episode discusses 'DanceOPD: On-Policy Generative Field Distillation,' a paper by Zhou et al. that teaches a single AI model multiple image generation skills without conflict. The hosts explain how the method uses 'on-policy' learning and 'hard-routed sample-wise field matching' to select the correct expert teacher for each step, leading to improved performance in editing and text-to-image tasks.