Return Tensor array from inference instead of numpy array or pil image

[joansc]

Hello,

I was wondering if there's a way that when doing inference image = pipe(prompt).images[0] it returns the image as Tensor array instead of the options of a numpy array or a pil image.

Thanks,

Joan

[asomoza]

You can use output_type="pt" and if you want the latents output_type="latent"

pipeline = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16,
    variant="fp16",
).to("cuda")

generator = torch.Generator(device="cpu").manual_seed(0)
tensor = pipeline(
    prompt="a cat", negative_prompt="", num_inference_steps=20, generator=generator, output_type="pt"
).images[0]

[joansc]

It works, thanks!