test-time-reward-alignment

TILT is a training-free framework that improves compositional generation in diffusion models by using a model-intrinsic reward to align sampling trajectories at test time, enhancing fidelity to complex prompts without external supervision.