Developed by Microsoft, Trellis 2 is a high-performance 3D generation model designed to transform 2D images into fully textured, production-ready 3D assets. Utilizing a massive 4B parameter model with a native 3D VAE, it represents a significant leap in structural versatility and texture fidelity.
Core Capabilities: Precision and Versatility
Trellis 2 is engineered to handle geometry that traditional AI models often struggle with. By employing a two-stage Diffusion Transformer (DiT) pipeline with 16x spatial compression, it achieves exceptional results in several key areas:
Complex Topologies: Trellis 2 seamlessly reconstructs open surfaces and transparent materials that often baffle standard reconstruction models.
High-Resolution Output: The model generates assets with up to 1536³ voxel resolution and high-fidelity PBR textures.
Structural Versatility: It is highly effective for creating stylized characters or complex environmental props with arbitrary geometry from diverse reference inputs.
Full PBR Support: The model outputs complete Physically Based Rendering material maps, ensuring realistic light interactions in professional renderers.
Key Generation Settings
The Trellis 2 interface allows for precise control over the final asset's quality and performance.
Setting | Options / Range | Description |
Resolution | 512, 1024, 1536 | Controls the voxel grid resolution used during the generation process. |
Decimation Target | 100k to 2M | Specifies the target triangle count for the final mesh, balancing geometric detail with file size. |
Texture Size | 1024, 2048, 4096 | Determines the resolution of the baked PBR texture maps for the final model. |
Remesh | Toggle (On/Off) | An optimization tool that refines the mesh topology for cleaner surface continuity. |
Remesh Band | Slider (1 to 4) | Adjusts the intensity or threshold of the remeshing process. |
Advanced Pipeline Configuration
For professional workflows, Trellis 2 exposes a unique three-stage refinement pipeline. This allows creators to granularly control the generation process by adjusting Guidance, Steps, and Rescale T for each specific phase.
Stage 1: Sparse Structure Generation
This initial phase sets the primary 3D volume and defines the basic structural silhouette of the asset.
Guidance (Default 8.5): Dictates how strictly the model follows the reference image's silhouette.
Steps (Default 20): Determines the number of iterations used to establish the core 3D structure.
Rescale T (Default 5): Adjusts the noise schedule to ensure a solid foundational volume.
Stage 2: Shape Structured Latent
The second stage focuses on refining surface geometry, ensuring continuity and structural integrity across the mesh.
Guidance (Default 7.5): Balances geometric complexity with the structural stability established in Stage 1.
Steps (Default 20): Refines the latent representation to smooth out surfaces and sharpen edges.
Guidance Rescale (Default 0.5): Fine-tunes the influence of the guidance to prevent geometric artifacts.
Stage 3: Texture Structured Latent
The final phase synthesizes surface colors and material properties like roughness and metallic maps.
Guidance (Default 2): Setting this to a lower value, such as 2, is critical for color accuracy. This ensures the resulting texture remains highly faithful to the original reference image rather than becoming overly stylized.
Steps (Default 20): The number of passes dedicated to baking high-resolution details into the PBR maps.
Rescale T (Default 3): Specifically targets the noise levels for material synthesis, ensuring realistic surface light interactions.
Was this helpful?