preloader
  • Home
  • Multi-Objective Alignment of Diffusion Model with Reverse-Time Drift Blending

Multi-Objective Alignment of Diffusion Model with Reverse-Time Drift Blending

Multi-Objective Alignment of Diffusion Model with Reverse-Time Drift Blending

We propose DiffusionBlend which enables test-time alignment of Diffusion Models to user-defined weights for objectives and regularization without additional fine-tuning. DiffusionBlend leverages the closed-form solution of regularized fine-tuning to combine pre-finetuned drift terms for individual reward functions at test time, guided by user-specified objectives and regularization weights.