Source-Prior-Driven Selective Adaptation Reduces Diffusion Model Fine-Tuning Degradation

This method uses a learned static mask to identify parameters for adaptation that minimize the loss of general generative capabilities. It addresses the trade-off between target-specific style improvement and the preservation of pretrained model breadth in diffusion models.