Converting the checkpoint model file to the Diffusers format

Loading checkpoint model data from a .ckpt or safetensors file is slow compared with the Diffusers format because every time we load a .ckpt or safetensors file, Diffusers will unpack and convert the file to the Diffusers format. To save the conversion every time we load a model file, we may consider converting checkpoint files to the Diffusers format.

We can use the following code to convert a .ckpt file to the Diffusers format:

ckpt_checkpoint_path = r"D:\temp\anythingV3_fp16.ckpt"
target_part = r"D:\temp\anythingV3_fp16"
pipe = download_from_original_stable_diffusion_ckpt(
    ckpt_checkpoint_path,
    from_safetensors = False,
    device = "cuda:0"
)
pipe.save_pretrained(target_part)

To convert a safetensors file to the Diffusers format, simply change the from_safetensors parameter to True as shown in the following sample...