trellis.cpp is a standalone C++/GGML reimplementation of Microsoft's TRELLIS.2-4B image-to-3D pipeline. It carries the whole chain — background removal, image conditioning, three flow transformers, the VAE decoders, mesh extraction and UV-textured GLB export — with no Python at runtime. The repository is MIT-licensed and has drawn 261 stars and 24 forks since its 28 June 2026 creation.

The pipeline mirrors the reference implementation stage for stage. A BiRefNet or RMBG matte cuts the subject out, DINOv3 ViT-L/16 turns the RGBA cutout into patch tokens, and three 1.3B flow DiTs run in sequence: a dense sparse-structure transformer at 16³ that decodes to a 64³ occupancy grid, a sparse Shape-SLAT transformer whose FlexiDualGrid decoder feeds FlexiCubes mesh extraction, and a sparse Texture-SLAT transformer that emits six-channel PBR values per voxel. Every one of those blocks runs on GGML, so the same source tree builds against CUDA, ROCm, Vulkan and CPU.

The postprocess is where a port usually diverges, and this one does not. The raw dual-grid mesh is welded and hole-filled, remeshed with narrow-band UDF dual contouring into a single manifold, quadric-simplified to a face budget (300K at res-1024), cluster-merged with the reference's normal-cone pass, unwrapped with stock xatlas, baked per texel by trilinear sampling of the voxel PBR volume with a BVH closest-point snap, gutter-filled with a Telea inpaint port, and written out as a GLB with lossy-WebP textures via EXT_texture_webp. The project documents this op-by-op against the reference in its spec and divergence-matrix docs and states parity with the reference CUDA postprocess on identical inputs.

Two front ends sit on top of the same engine. trellis-server stays resident so Vulkan is not re-initialised per request, exposing GET /health and a multipart POST /generate that returns model/gltf-binary; launch flags set defaults and each request can override seed, resolution and background-removal mode. Trellis Studio is a Tauri desktop app installed by one curl | bash (or irm | iex on Windows) that auto-detects the GPU runtime, pulls the matching server build plus roughly 16.5 GB of weights, supervises the server, and keeps every result in a local IndexedDB gallery. The project also serves as the trellis backend of Lemonade, and can be driven end-to-end from a text prompt by pairing it with stable-diffusion.cpp for the input image.

Release v0.6.0 landed on 19 August 2026 and is mostly a reliability release. It fixes the severe holes and corrupted geometry the 0.5.4 CUDA build produced at high resolution — FlashAttention values are power-of-two scaled around the tensor-core path so large Shape-SLAT activations stay finite, and the pinned GGML revision uses 64-bit mask strides for attention masks above 2 GiB. BiRefNet now evaluates its final high-resolution convolutions in bounded-height stripes and resizes the matte back to the source dimensions, so non-square inputs stop being stretched. Separate cuda12 packages built with CUDA 12.9 restore Pascal and Volta support (compute capability 6.0, 6.1 and 7.0, cards such as the Tesla P100), with the installers picking the legacy runtime automatically. CPU execution now uses all available cores by default: a contributed 20-core benchmark moved the sparse-structure flow stage from 305 s/step to 118 s/step, a 2.6x improvement on that machine.

Worth knowing before you commit: the CPU path is described by the project itself as very slow and RAM-hungry — --require-gpu exists precisely to fail rather than silently fall back — and the ROCm backend is still labelled experimental in the release assets.