trellis.cpp
The native C++ and GGML runtime that makes local TRELLIS.2 generation practical across supported GPU backends.
ORIGINAL REPOSITORY
IMAGE TO 3D ASSET, ON YOUR MACHINE
CHOOSE YOUR EXPERIENCEExplore at your own pace. You can reopen this guide whenever you need it.
Each room follows one stage of the Triastasis workflow, from local generation to a versioned asset ready for another tool.
Triastasis connects local TRELLIS.2 generation to the work around each result. Start with an image, then generate, inspect, compare, organize, and export without sending your work to a hosted service.
Generation stays on your machine. Triastasis builds a product workflow around Microsoft TRELLIS.2 and Piotr Wilkin’s native trellis.cpp runtime while keeping their upstream contributions explicit.
The native C++ and GGML runtime that makes local TRELLIS.2 generation practical across supported GPU backends.
ORIGINAL REPOSITORYThe image-to-3D research, reference implementation, and model family beneath the reconstruction process.
REFERENCE PROJECTBring in one image, choose a practical quality level, and generate locally with clear progress and reproducible settings.
Run TRELLIS.2 locally on supported CUDA or Vulkan hardware. ROCm support is experimental and requires a compatible external runtime.
We built this to make testing different seeds easy. Generate several candidates from the same source and compare consecutive seeds as one group.
Use approachable presets or open detailed controls for masking, resolution, geometry, UVs, and textures.
Recover interrupted generations by requeuing the original seed and settings. This starts a new run while preserving the asset lineage.
A convincing texture can hide difficult geometry. Pair every source with its generated model, then examine topology, materials, UVs, dimensions, mesh parts, and generation metadata before choosing what to keep.

Move between textured, clay, wireframe, normal, UV, and material-channel views while keeping the exact source reference and technical record close.
The current viewer helps you find a problematic mesh part before editing begins. Selection and visibility controls are available now. The derived-edit workflow is still in development and remains deliberately separate from the original asset.
Choose a component from the model or the mesh-parts list and frame it in the viewer.
Isolate, hide, or reveal parts so geometry and material issues are easier to inspect.
Cleanup, transforms, material changes, normal repair, and derived exports are not exposed in the current release UI.
The planned tools target static assets. They do not create rigs or promise animation-ready topology.
A useful asset is more than one GLB in an output folder. Triastasis keeps the source, settings, measurements, candidates, versions, and exports connected.
The library makes related results findable while a portable manifest records how each generation was produced.
Search, filter, favourite, compare, rename, and reopen saved work.
Preserve originals and understand which parent produced each related version.
Keep the source, seed, settings, measurements, hashes, and runtime information together.
Local generation becomes part of a larger creative workflow instead of an isolated button.
The loopback-only automation API accepts queued jobs and runs one generation at a time. Local tools and the optional Codex workflow can monitor progress and export GLBs or portable packages for Blender and game engines. A built-in Open in Blender button and multi-asset scene export remain planned.
The first public alpha is here for Windows x64. Install the desktop app, then let onboarding guide you through the runtime and AI models needed to generate 3D assets. Linux support is still in development.
Version 0.0.1-alpha.1. Open the release page to choose the Windows installer or portable ZIP. The runtime and AI model bundle are separate downloads inside the app. This alpha exports static GLBs without rigging or animation. The installer is not yet code-signed, and updates are installed manually.
VIEW ALPHA RELEASETriastasis detects your GPU and downloads the matching runtime, the engine that runs generation locally. Next, review the model terms, choose a storage folder, and download the recommended AI model bundle. Onboarding shows its download size, verifies the files, and can resume interrupted model downloads. Once setup is complete, add an image to generate your first 3D asset.
OPEN SETUP GUIDERead the current README, build instructions, notices, and application source.
OPEN PROJECTCheck active work and report problems against the current source tree.
VIEW ISSUES