handler.py

import os
import torch
import base64
import io
from PIL import Image
from trellis.pipelines import TrellisImageTo3DPipeline
from trellis.utils import postprocessing_utils
from typing import Dict, Any

class EndpointHandler:
    def __init__(self, model_dir: str):
        """
        Initialize the TRELLIS pipeline.
        """
        # Set algorithm to 'native' for faster startup on Inference Endpoints
        os.environ['SPCONV_ALGO'] = 'native'
        
        # Load the pipeline from the local directory or HF hub
        # 'microsoft/TRELLIS-image-large' is the standard model
        self.pipeline = TrellisImageTo3DPipeline.from_pretrained("microsoft/TRELLIS-image-large")
        self.pipeline.cuda()
        self.pipeline.preprocess_image = self.pipeline.preprocess_image # Ensure visibility

    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
        """
        Args:
            data (:obj:`Dict[str, Any]`):
                - "inputs": The image as a base64 string or URL.
                - "params": Dictionary of optional parameters (seed, steps, etc.)
        """
        inputs = data.pop("inputs", data)
        params = data.pop("params", {})

        # 1. Decode Image
        if isinstance(inputs, str):
            image = Image.open(io.BytesIO(base64.b64decode(inputs))).convert("RGB")
        else:
            image = inputs

        # 2. Run Pipeline
        # You can adjust 'sparse_structure_sampler_params' and 'slat_sampler_params' here
        outputs = self.pipeline.run(
            image,
            seed=params.get("seed", 42),
            sparse_structure_sampler_params=params.get("sparse_params", {"steps": 12, "cfg_strength": 7.5}),
            slat_sampler_params=params.get("slat_params", {"steps": 12, "cfg_strength": 3.0})
        )

        # 3. Post-process to GLB
        # We extract the mesh and simplify it for export
        glb = postprocessing_utils.to_glb(
            outputs['gaussian'][0], 
            outputs['mesh'][0],
            simplify=params.get("simplify", 0.95),
            texture_size=params.get("texture_size", 1024)
        )

        # 4. Encode to Base64 for response
        buffered = io.BytesIO()
        glb.export(buffered)
        glb_str = base64.b64encode(buffered.getvalue()).decode("utf-8")

        return {
            "mesh_base64": glb_str,
            "format": "glb"
        }