Researchers introduce Weighted Adjoint Matching, a finetuning technique for pretrained diffusion and flow matching models. The method assumes access to a differentiable reward function and adjusts the model to optimise that reward.
The approach builds on Adjoint Matching, which frames finetuning as a stochastic optimal control problem. Weighted Adjoint Matching adds a time-dependent control cost, allowing the finetuning to weigh adjustments differently across the generation trajectory.
Because the source is a single arXiv abstract, there are no conflicting findings to compare. The abstract does not yet report experimental results or comparisons to other finetuning methods.