SegviGen turns 3D models into modular parts with a single click

SegviGen is an open-source framework that slices monolithic 3D meshes into clean, modular parts with pinpoint accuracy. Whether you are prepping digital toys for 3D printing or reorganizing game assets, it turns messy meshes into workable, isolated components.

Built on Microsoft’s TRELLIS.2 architecture, SegviGen sidesteps the usual headache of training heavy segmentation models from scratch. Instead, it reuses the structural and textural knowledge already baked into large-scale 3D generative backbones, learning to predict part-indicative colors while reconstructing geometry.

What it can do

SegviGen unifies three distinct segmentation workflows under a single setup.

The first mode is interactive. You drop a single click on a mesh feature, and the model instantly segments out that specific part. On benchmarks like PartObjaverse-Tiny and PartNeXT, its single-click IoU@1 outperforms previous state-of-the-art model P3-SAM by 40%.

The second mode is unguided full segmentation. Without any prompt, it carves up the entire object into categorized, color-coded parts, boosting average full segmentation IoU by 15% across datasets.

The third mode takes a 2D guide image. Feed it a 2D segmentation map, and SegviGen breaks down the 3D asset matching whatever level of granularity you specified in the picture.

The killer feature here is data efficiency. Because it borrows existing generative priors, SegviGen needed minimal task-specific supervision, pulling off these results with just 0.32% of traditional training data.

Getting started

The project is fully open source on GitHub under an MIT license, while its checkpoint weights are hosted on Hugging Face. The paper has been accepted by the SIGGRAPH 2026 Journal, and an interactive online demo is available for quick tests.

Running it locally requires Linux, Python 3.10, and an Nvidia GPU with at least 24GB of VRAM. You will need to clone and install TRELLIS.2 with its custom dependencies before grabbing SegviGen's scripts, Blender bindings, and checkpoint files.

Inference uses straightforward CLI commands. Separate scripts handle single-click extraction, automated full segmentation, and 2D-guided slicing, directly spitting out partitioned GLB files ready for downstream editing.

Why it matters

Segmenting 3D assets has long been a tedious bottleneck requiring either hours of manual poly cutting or brittle models trained on scarce labeled datasets. SegviGen cleverly shows that 3D foundation models already understand object anatomy, and repurposing those generative weights cuts training costs down to virtually nothing while setting new performance benchmarks.

If you work in digital fabrication, game modding, or robotics perception, this is an immediate download. While the 24GB VRAM ceiling rules out basic laptops, anyone with an RTX 3090, 4090, or a cheap cloud instance can now automate hours of mesh cleanup in seconds.