Decouple-Then-Merge: Finetune Diffusion Models as Multi-Task Learning

In Sec. 1.1 and Sec. 1.2, we demonstrate that DeMe can be formally transformed into a loss reweighting framework, just like previous works [3, 6, 15] .