How to Accelerate Training Data Generation with SDXL-Turbo or LCM?

[HaozheZhao]

Employing the SD-1.5 for data generation necessitates a 100-step ddim for each of 100 samples. This can be quite time-consuming.

Would it be feasible to use the SDXL-Turbo or LCM instead, to hasten the process? Since it only requires about 1-5 steps per sample, which could simplify and expedite data generation and scaling.

However, it appears that the original stable-diffusion codebase is not compatible with the SDXL-Turbo or LCM codebase.