this experience honestly makes me wonder what's the right approach to design parameter tuning. the old iterative ways like adjoint optimization were guided only by objectives you as the designer could think of and take several iterations to get to the right design spacedefinitely feel like there is some hybrid approach to find the global valley faster and more intelligently with AI then use traditional optimizers to fine tune parameters

1/7 When rewards conflict, what should RL post-training of diffusion models optimize?In visual generation, objectives are often in tension:Prompt adherence can conflict with source preservation.Photorealism can conflict with stylization.In our new paper, ParetoSlider, we introduce a multi-objective RL framework that trains a single diffusion model for continuous control over competing reward objectives 🧵