판 이력 — TILT: Improving Compositional Generation in Diffusion Models with a Model-Intrinsic Reward | AIChainDay