These models are merged fp8 variants of the original models with this LoRA: https://huggingface.co/tianweiy/DMD2/resolve/main/dmd2_sdxl_4step_lora_fp16.safetensors
Use the ComfyUI workflow provided along with the checkpoints to further accelerate the process using Kohya Deep Shrink. Able to reduce SDXL gen time from 90 minutes to 10 in my computer.
ONLY RECOMMENDED IF YOU HAVE A REALLY SLOW / NO GPU.