microsoft/TRELLIS.2 ? reverse-engineered prompt

Build me an image to 3D tool that takes a single photo of any object and turns it into a high quality fully textured 3D model in just a few seconds. It should handle really complex shapes like clothing with folds, plants with leaves, or objects that have holes and internal parts without breaking. The output needs rich PBR materials including color, roughness, metallic properties and transparency so it looks photorealistic when rendered.

Give me a clean Python pipeline that loads the 4B model from Hugging Face, runs inference on a GPU, produces a mesh, renders a spinning video preview with good lighting, and exports everything to a clean GLB file ready for Unity or Blender. Add a simple web demo where I can drag in a picture and get the result back plus an example script that shows it all end to end.

Make it efficient enough to run on an A100 or H100 and keep the code modular so I can later add training if I want. Feel free to look up the latest docs for any rendering libraries you need.