The core of HoloPart is a novel diffusion-based generative model built upon the TripoSG foundation, which has been extensively pretrained on large-scale 3D datasets and fine-tuned for part-whole understanding. The segmentation process begins by using advanced methods like SAMPart3D to extract initial, often incomplete, surface fragments corresponding to different parts. HoloPart then applies a dual attention mechanism: local attention captures fine geometric details of each part, while context-aware attention ensures that the completed part remains consistent with the overall object’s structure. This approach allows HoloPart to intelligently reconstruct missing geometry, even in complex or highly occluded scenarios, producing high-fidelity, complete parts that align with the original model.
HoloPart’s capabilities open up a range of powerful applications for 3D content creators. By generating complete, editable parts, users can intuitively select, scale, move, or replace individual components, assign materials or textures with precision, and prepare assets for animation or rigging. The technology also facilitates robust geometric processing, such as remeshing and geometric super-resolution, and lays the groundwork for future generative models that operate at the part level. HoloPart has been evaluated on new benchmarks and consistently outperforms existing shape completion methods, making it a vital tool for anyone working with 3D models in design, animation, or virtual production.
Key features include:
- Performs 3D part amodal segmentation, inferring complete geometry even for occluded parts
- Built on a diffusion-based generative model with dual attention mechanisms
- Supports intuitive editing and animation-ready asset creation
- Enables precise material assignment and robust geometric processing
- Open-source with pretrained models and interactive demo available
- Outperforms state-of-the-art methods on new 3D part completion benchmarks