LIDA

TRELLIS.2 is an open-source image-to-3D model generator, designed to produce high fidelity textured assets using native 3D VAEs. Its core feature makes use of native and compact structured latents for providing both high fidelity and compression capabilities. This tool can handle complex structures, including open surfaces, non-manifold geometry, and enclosed interior structures, thus overcoming the limitations of iso-surface fields. It's a research project with Responsible AI considerations factored into all stages of its development.

TRELLIS.2 generates 3D models from images through a process that begins with an Instant Bidirectional Conversion that transforms meshes into a new representation termed O-Voxel. The Sparse Compression VAE then encodes these voxels into a compact Structured Latent space. The result is a highly compact representation of a fully textured 3D asset with negligible perceptual degradation.

Key features of TRELLIS.2 include the ability to handle complex structures, model arbitrary surface attributes such as base colour, roughness, metallic, and opacity, and optimize pre and post-processing of data for training and inference. Additional features include the utilization of 'O-Voxel', a novel 'field-free' sparse voxel structure to encode detailed geometry and complex appearance simultaneously, and a Sparse Compression VAE component for efficient voxel data compression.

TRELLIS.2 can handle complex structures including open surfaces, non-manifold geometry, and enclosed interior structures.

TRELLIS.2 handles non-manifold geometry by using a flexible dual grids representation within the O-Voxel structure. This methodology allows it to handle arbitrary topologies while preserving sharp edges.