Grafting is a simple method for editing pre-trained diffusion transformers (DiT) proposed by a team from Stanford University in 2025, which only requires a small computational budget.Exploring Diffusion Transformer Designs via Grafting"

The specific process of Grafting is as follows:

Activation Distillation: Transfer the functionality of the original operator to a new operator and distill its activation values through the regression objective.
Lightweight fine-tuning: Mitigating the error propagation caused by integrating multiple new operators by performing end-to-end fine-tuning with limited data.