inference_karras.py

import torch
from diffusers import StableDiffusionKDiffusionPipeline, AutoencoderKL
pipe = StableDiffusionKDiffusionPipeline.from_pretrained('/models/pseudo-test', 
    torch_dtype=torch.float16,
    safety_checker = None,
)
pipe.set_scheduler('sample_dpmpp_2m')
vae = AutoencoderKL.from_pretrained("stabilityai/sd-vae-ft-mse", use_safetensors=True, torch_dtype=torch.float16)
pipe.vae = vae
pipe.to('cuda')
image = pipe(prompt='best quality ((bear)) in the swiss alps cinematic 8k highly detailed sharp focus intricate fur',
    negative_prompt='malformed, disgusting',
    num_inference_steps=50, generator=torch.Generator(device='cuda').manual_seed(42), 
    width=1280, height=720, guidance_scale=3.0,
    use_karras_sigmas=True).images[0]
image.save('bear.png', format="PNG")