The SDXL-Lightning model is doing something very similar. It is a "distilled" version of SDXL for very low step counts (2, 4, 8). In addition to a full model,

The SDXL-Lightning model is doing something very similar. It is a "distilled" version of SDXL for very low step counts (2, 4, 8). In addition to a full model, it can be integrated into other trained SDXL models as a LORA. This seems to be an attempt to get the benefits of the Lightning model without the problems that come with it (worse prompt adherence)