import os
import torch
from PIL import Image
from diffusers.utils import load_image
from diffusers import FluxControlNetModel
from diffusers.pipelines import FluxControlNetPipeline
from io import BytesIO

class EndpointHandler:
    def __init__(self, model_dir="huyai123/Flux.1-dev-Image-Upscaler"):
        # Access the environment variable
        HUGGINGFACE_API_TOKEN = os.getenv('HUGGINGFACE_API_TOKEN')
        if not HUGGINGFACE_API_TOKEN:
            raise ValueError("HUGGINGFACE_API_TOKEN")

        # Load model and pipeline
        self.controlnet = FluxControlNetModel.from_pretrained(
            model_dir, torch_dtype=torch.bfloat16, use_auth_token=HUGGINGFACE_API_TOKEN
        )
        self.pipe = FluxControlNetPipeline.from_pretrained(
            "black-forest-labs/FLUX.1-dev",
            controlnet=self.controlnet,
            torch_dtype=torch.bfloat16,
            use_auth_token=HUGGINGFACE_API_TOKEN
        )
        self.pipe.to("cuda")

    def preprocess(self, data):
        # Load image from file
        image_file = data.get("control_image", None)
        if not image_file:
            raise ValueError("Missing control_image in input.")
        image = Image.open(image_file)
        w, h = image.size
        # Upscale x4
        return image.resize((w * 4, h * 4))

    def postprocess(self, output):
        # Save output image to a file-like object
        buffer = BytesIO()
        output.save(buffer, format="PNG")
        buffer.seek(0)  # Reset buffer pointer
        return buffer

    def inference(self, data):
        # Preprocess input
        control_image = self.preprocess(data)
        # Generate output
        output_image = self.pipe(
            prompt=data.get("prompt", ""),
            control_image=control_image,
            controlnet_conditioning_scale=0.6,
            num_inference_steps=28,
            height=control_image.size[1],
            width=control_image.size[0],
        ).images[0]
        # Postprocess output
        return self.postprocess(output_image)