Why autopipeline and text2img pipeline cost time much different?

[zengjie617789]

I used the autopipeline and text2imgpipeline to load model respectively, but the time is much different. Here are code I used and the time screenshot：

pipe = StableDiffusionXLPipeline.from_pretrained(data_dir,  use_safetensors=True, variant="fp16").to("cuda")
pipe= AutoPipelineForText2Image.from_pretrained(data_dir, torch_dtype=torch.float16, variant="fp16", use_safetensors=True).to("cuda")

the same prompt and steps.
Any help will be appreciated.

[tolgacangoz]

To equalize them totally, one needs to add torch_dtype=torch.float16 to the first pipe; because the default is torch_dtype=torch.float32.