from diffusers import DiffusionPipeline, LCMScheduler
import torch

loaded_pipe = None
loaded_pipe_id = None

def load_model(pipe_id):
    global loaded_pipe, loaded_pipe_id
    if loaded_pipe_id != pipe_id:
        loaded_pipe = DiffusionPipeline.from_pretrained(pipe_id, torch_dtype=torch.float16).to("cuda")
        loaded_pipe.scheduler = LCMScheduler.from_config(loaded_pipe.scheduler.config)
        loaded_pipe.load_lora_weights("latent-consistency/lcm-lora-sdxl")
        loaded_pipe_id = pipe_id
    return loaded_pipe

def generate_image(prompt, num_inference_steps, seed, guidance_scale, negative_prompt=None, pipe_id="Linaqruf/animagine-xl"):
    global loaded_pipe
    pipe = load_model(pipe_id)
    generator = torch.manual_seed(seed)
    image = pipe(prompt=prompt, negative_prompt=negative_prompt, num_inference_steps=num_inference_steps, generator=generator, guidance_scale=guidance_scale).images[0]
    return image