Bambu Lab discloses SegviGen 3D model using 0.32% labels to beat P3-SAM by 15% IoU

Bambu Lab has appeared alongside Beihang University, Renmin University of China and Tsinghua University on SegviGen, a paper accepted to the SIGGRAPH 2026 Journal Track and its first public disclosure of research into a 3D generative model. The work repurposes a pretrained native-3D generative model for part segmentation, converting segmentation into part-level colouring rather than training a separate model from large, densely annotated 3D datasets. During conditional denoising, SegviGen reconstructs geometry while assigning indicator colours to structurally meaningful components, supporting interactive selection, full segmentation and 2D-mask-guided segmentation at arbitrary granularity.

The reported results are technically notable but remain research results: SegviGen used 0.32% of labelled training data, improved single-click IoU@1 by 40% versus P3-SAM, and led the best baseline by 15% in average overall IoU across datasets. That targets a real weakness in consumer FDM/FFF workflows, where generated meshes may be visually plausible yet difficult to edit, divide, colour, orient or manufacture as separate parts. A reliable part-aware layer could connect Bambu Lab's hardware, Bambu Studio and MakerWorld with more manufacturable AI-generated content, rather than limiting AI to prompt assistance, model search or slicer-parameter recommendations. Bambu Lab did not confirm whether SegviGen will enter its software stack, and the paper itself cautions that production deployment would still require work on complex user files, inference speed and slicer integration.

What readers newly know is that Bambu Lab's AI work extends below consumer-facing generation tools into native 3D structure understanding, a more consequential capability for automated part splitting and multi-colour or multi-material preparation. Its MakerLab platform has already moved to paid access as AI-generation usage rose, but this disclosure does not move an integration date: printer buyers and software users should treat autonomous segmentation as an unannounced roadmap option, not a current feature. AM Pulse logged 20 Bambu Lab name-matched company-news items in the past 90 days versus 31 in the prior period; that count is name-matched only and does not measure sector activity.