Nelipot-Lee / SegviGen

What is SegviGen?

SegviGen repurposes pretrained 3D generative models in Python to perform part segmentation on 3D assets like GLB files. Feed it a rendered image or 2D map alongside the model, and it outputs a colored GLB where parts are distinctly segmented—via single-click interactive mode, unguided full segmentation, or 2D-guided refinement. Built on TRELLIS.2, it leverages generative priors for accurate 3D structure and texture understanding with minimal task-specific data.

Why is it gaining traction?

It achieves state-of-the-art IoU gains (40% on single-click, 15% overall) over tools like P3-SAM while needing just 0.32% training data, making it data-efficient for segmentation tasks. A single architecture handles interactive, full, and guided modes with arbitrary part counts, plus pretrained weights on Hugging Face for instant inference on 24GB GPUs. Developers dig the flexible CLI scripts for quick GLB processing without retraining.

Who should use this?

3D computer vision engineers segmenting assets for AR/VR apps or robotics. Researchers fine-tuning generative models for custom part detection on Objaverse-like datasets. Game devs needing interactive part labeling from single image clicks.

Verdict

Promising research prototype for repurposing generative models in segmentation—grab it if you're experimenting with 3D ML pipelines. Low 1.0% credibility score and 46 stars reflect early-stage code; expect setup hurdles and sparse docs, but solid for proofs-of-concept.

(198 words)