handler.py

from typing import Dict, List, Any
from diffusers import StableDiffusionPipeline
import torch
from PIL import Image
import numpy as np

class EndpointHandler():
    def __init__(self, path=""):
        # Preload all the elements you are going to need at inference.
        # pseudo:
        # self.model= load_model(path)
        model_id = "selectmixer/SDXL_NIJI_SIX"
        self.pipe = StableDiffusionPipeline.from_pretrained(model_id)
        self.pipe.to("cuda")

        self.pipe.load_lora_weights("selectmixer/LORA_NIJI_DLCV6_SDXL", weight_name="SDXL_Niji_V6_DLC_LoRa_V2.safetensors", adapter_name="test")

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        """
       data args:
            inputs (:obj: `str` | `PIL.Image` | `np.array`)
            kwargs
      Return:
            A :obj:`list` | `dict`: will be serialized and returned
        """

        # pseudo
        # self.model(input)

        
        # generator = torch.Generator("cuda").manual_seed(31)
        # image = self.pipe("Astronaut in a jungle, cold color palette, muted colors, detailed, 8k", generator=generator).images[0]
        # return image
    
        inputs = data.get("inputs")
        kwargs = data.get("kwargs", {})

        # Convert inputs to the required format
        if isinstance(inputs, str):
            # Assuming the string is a prompt for text-to-image generation
            prompt = inputs
        elif isinstance(inputs, Image.Image):
            # Convert PIL.Image to a format suitable for the model
            prompt = self.pipe.feature_extractor(images=inputs, return_tensors="pt").to("cuda")
        elif isinstance(inputs, np.ndarray):
            # Convert np.array to PIL.Image and then to the required format
            image = Image.fromarray(inputs)
            prompt = self.pipe.feature_extractor(images=image, return_tensors="pt").to("cuda")
        else:
            raise ValueError("Unsupported input type")

        lora_scale = kwargs.get("lora_scale", 0.5)

        # Generate images using the model
        results = self.pipe(prompt, cross_attention_kwargs={"scale":lora_scale}, **kwargs)

        # Convert results to a list of dictionaries
        output = [{"generated_image": result} for result in results.images]

        return output