TPSO Improves Diffusion Model Image Diversity Without Retraining

TPSO is a training-free, model-agnostic module that optimizes token-prompt embedding space to prevent diffusion models from collapsing into high-probability modes. It uses learnable parameters to explore underrepresented embedding regions while employing semantic constraints to maintain image quality and fidelity.