[Task] image-to-3d model support

Overview

Image-to-3D models generate a 3D representation (mesh, NeRF, or structured latent) from a single input image. TRELLIS uses a structured latent representation (SLAT) and a powerful diffusion backbone, supporting both image and text conditioning and outputting high-quality meshes, radiance fields, and Gaussian splats.

Agent Scenarios

• E-commerce product agent: convert a product photo into a 3D model for AR try-on or 360° viewing directly on Windows
• Game asset creation agent: generate 3D props or characters from reference images for use in a game engine
• Interior design agent: lift a photo of a furniture piece into a 3D object to place and visualize in a room layout
• Digital twin agent: reconstruct physical objects from photos for simulation or inspection workflows

ModelKit Integration

Models must pass the full wmk pipeline on all EPs:

wmk config → wmk build (ONNX export) → wmk perf → wmk eval

Acceptance Criteria

• microsoft/TRELLIS-image-large

Start by reviewing existing model integrations and the ModelKit entry points used by wmk config, wmk build, wmk perf, and wmk eval. Determine how microsoft/TRELLIS-image-large should pass the full pipeline on all EPs. Done means the model is integrated and all four commands complete successfully across the supported EPs.