handler.py

from typing import Dict, List, Any
import torch
import base64
import io
from PIL import Image
from tryon_core import TryOnEngine
from api_utils import prepare_image_for_processing, image_to_base64

class EndpointHandler:
    def __init__(self, path=""):
        # Initialize the engine
        # path is the path to the model files on the HF container
        print("Initializing IDM-VTON Handler...")
        self.engine = TryOnEngine(load_mode="4bit", enable_cpu_offload=False, fixed_vae=True)
        
        # Override model_id to load from local path if needed, 
        # or let it download from Hub if path is just a directory
        # self.engine.model_id = path 
        
        self.engine.load_models()
        self.engine.load_processing_models()
        print("Handler Initialized!")

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        """
        Args:
            data (:obj:):
                includes the input data and the parameters for the inference.
        """
        # 1. Extract inputs
        inputs = data.pop("inputs", data)
        human_img_b64 = inputs.get("human_image")
        garment_img_b64 = inputs.get("garment_image")
        description = inputs.get("garment_description", "a photo of a garment")
        category = inputs.get("category", "upper_body")
        
        # 2. Decode images
        human_img = Image.open(io.BytesIO(base64.b64decode(human_img_b64)))
        garment_img = Image.open(io.BytesIO(base64.b64decode(garment_img_b64)))
        
        # 3. Process
        human_img = prepare_image_for_processing(human_img)
        garment_img = prepare_image_for_processing(garment_img)
        
        # 4. Generate
        generated_images, masked_image = self.engine.generate(
            human_img=human_img,
            garment_img=garment_img,
            garment_description=description,
            category=category,
            use_auto_mask=True,
            use_auto_crop=True,
            denoise_steps=30,
            seed=42,
            num_images=1
        )
        
        # 5. Return result
        return [{
            "generated_image": image_to_base64(generated_images[0]),
            "masked_image": image_to_base64(masked_image)
        }]