cs.CVSep 22, 2026
SavePartLLM: A Unified Multimodal Foundation for 3D Part Segmentation
Organizations: Taiyuan University of Technology, Tencent Visvise, China · Hong Kong University of Science and Technology, Tencent Visvise, China · Taiyuan University of Technology, China · Tencent Visvise, China ·
Abstract
Part segmentation is a fundamental problem in computer graphics and 3D vision. Recent works have expanded 3D part segmentation beyond fixed taxonomies, but existing approaches typically only address a specific setting, such as text-guided part segmentation or point-based interaction. In this work, we argue that these settings can be unified as an intent-conditioned generative problem, where different prompts specify the desired part decomposition. To this end, we introduce PartLLM, a unified multimodal model that formulates 3D part segmentation as autoregressive semantic decomposition. Conditioned on an input shape and a user prompt, PartLLM autoregressively generates semantic part hypotheses as queries for mask prediction and feeds them to a decomposition-aware decoder that jointly predicts coherent part masks. This unified design supports text-guided part segmentation, interactive segmentation, and full-shape semantic decomposition with controllable granularity within a single model. Extensive experiments across these task settings show that PartLLM consistently outperforms task-specific baselines, demonstrating the effectiveness of unifying 3D part segmentation under an intent-conditioned generative formulation.
Figures & tables
Figure 2. PartLLM achieves unified part segmentation within a single model. Given the same shape, it produces different segmentations conditioned on user intent, including text-guided part segmentation, interactive segmentation, and semantic full-shape decomposition at controllable granularities. Teaser figure showing one 3D shape segmented by PartLLM under different user intents, including text-guided, point-prompted, and full-shape semantic decomposition settings.
Figure 1. Gallery of PartLLM on diverse 3D shapes. Across complex artist-created and AI-generated shapes, PartLLM produces semantically meaningful part segmentation with coherent functional structure, clean boundaries, and fine geometric details. Gallery of PartLLM results on diverse 3D objects, showing semantic part decompositions across artist-created and AI-generated assets.
| Method | Promptable segmentation | Full-shape segmentation | ||
|---|---|---|---|---|
| Interactive Point | Part Text | Category-agnostic | Semantic | |
| PartField ( Liu et al., 2025 ) | – | – | ✓ | – |
| PartSAM ( Zhu et al., 2026 ) | ✓ | – | ✓ | – |
| P3-SAM ( Ma et al., 2025 ) | ✓ | – | ✓ | – |
| S2AM3D ( Su et al., 2025 ) | ✓ | – | ✓ | – |
| SegviGen ( Li et al., 2026 ) | ✓ | – | ✓ | – |
Table 1. Task coverage of representative part segmentation methods. Existing methods cover only subsets of promptable and full-shape segmentation capabilities, whereas PartLLM supports point-prompted interaction, text-guided part segmentation, category-agnostic segmentation, and semantic full-shape segmentation within a single unified model.
Figure 3. Overview of PartLLM. Different part segmentation tasks are unified as autoregressive semantic decomposition: the user intent (right) conditions a 3D-aware multimodal language model to generate task-specific part hypotheses, which are then jointly decoded into mutually consistent masks. Overview of the PartLLM pipeline. A point cloud encoder produces 3D tokens for a multimodal language model conditioned on user intent. The model generates named and localized part hypotheses, which a decomposition-aware decoder converts into a complete colored segmentation.
Figure 4. Comparison of two mask decoding paradigms. SAM-style decoders predict each mask independently, using a sigmoid decision for every query. PartLLM instead performs decomposition-aware joint decoding. Autoregressively generated part hypotheses form a dynamic query set, and a softmax competition assigns each point to one generated part or background. This converts mask prediction from independent binary decisions into a joint decomposition of the shape. Diagram of the PartLLM decoder. Generated part and background queries interact with point features, produce an N by S plus one score matrix, and use row-wise softmax for joint point assignment. A small SAM-style reference indicates independent binary mask decoding with sigmoid.
| Task | Prompt input | Generated response |
|---|---|---|
| Full-shape | general: Please segment and name all parts in <point_cloud>. coarse: Please segment and name the main parts in <point_cloud>. fine: Please segment and name all detailed parts in <point_cloud>. number: Please segment and name about {n} parts in <point_cloud>. | label=Backrest, bbox=[...] <|PART|> label=Seat, bbox=[...] <|PART|> label=Armrest, bbox=[...] <|PART|> label=Base, bbox=[...] <|PART|> <|BG|> |
| Text-guided | Please segment the Backrest and Armrest in <point_cloud>. | label=Backrest, bbox=[...] <|PART|> label=Armrest, bbox=[...] <|PART|> <|BG|> |
| Interactive | First click: segment the part at (x, y, z) in <point_cloud>. | label=Armrest, bbox=[...] <|PART|> <|BG|> |
| Refinement: previous mask + include/exclude point (x, y, z). | label=Armrest, bbox=[...] <|PART|> <|BG|> |
Table 2. Prompt-response template for the unified task formulation.
Figure 5. Qualitative comparison with baseline methods ( Liu et al., 2025 ; Ma et al., 2025 ; Zhu et al., 2026 ; Li et al., 2026 ) on full-shape segmentation. Comparison of full-shape part decompositions produced by PartField, P3-SAM, PartSAM, SegViGen, and PartLLM on five 3D objects. Different colors denote the predicted parts.
Figure 6. Qualitative full-shape segmentation results on AI-generated meshes. The first row contains dense meshes produced by geometry-based generative models ( Lai et al., 2025 ; Xiang et al., 2025 ) , while the second row contains low-poly meshes produced by autoregressive topology generative models ( Zhao et al., 2025 ; Liu et al., 2026 ) . Colored full-shape part segmentations of dense and low-poly AI-generated 3D meshes, including characters, creatures, and objects.
| Method | PartNeXt | 3DCoMPaT200 | HY3D-Bench | PartObjaverse -Tiny | |
|---|---|---|---|---|---|
| Coarse | Fine | ||||
| PartField | 42.7 | 45.8 | 35.0 | 33.6 | 51.5 |
| SAMPart3D | 40.8 | 51.6 | 41.1 | 36.9 | 53.5 |
| PartSAM | 37.2 | 42.6 | 42.2 | 40.5 | 69.5 |
| P3-SAM | 40.3 | 41.4 | 40.3 | 38.7 | 59.9 |
| SegviGen | 34.2 | 43.8 | 37.2 | 35.4 | 50.6 |
Table 3. Quantitative results on category-agnostic full-shape segmentation. We report mask-only mIoU by matching each ground-truth part to its best-IoU prediction across granularities.
Figure 7. Number-based granularity control. By changing the requested part counts, PartLLM produces progressively finer decompositions while preserving semantic coherence. A number-control figure showing the same object decomposed by PartLLM under prompts requesting about 2 parts, about 5 parts, and about 10 parts.
| Method | PartNeXt | 3DCoMPaT200 Coarse | 3DCoMPaT200 Fine | |||
|---|---|---|---|---|---|---|
| Mask mIoU | SA-mIoU | Mask mIoU | SA-mIoU | Mask mIoU | SA-mIoU | |
| PartSAM+MLLM | 37.2 | 13.8 | 42.6 | 25.1 | 42.2 | 20.8 |
| P3-SAM+MLLM | 40.3 | 18.3 | 41.4 | 23.5 | 40.3 | 16.7 |
| Ours | 56.7 | 43.2 | 86.1 | 71.9 | 74.2 | 63.1 |
Table 4. Quantitative results on semantic full-shape segmentation. We report semantic-aware mIoU (SA-mIoU) together with mask-only mIoU.
Figure 8. Qualitative comparison with baseline methods ( Ma et al., 2024 ; Hadgi et al., 2026 ; Jin et al., 2026 ) on text-guided part segmentation. Text-guided part segmentation comparison among Find3D, PatchAlign3D, CoSMo3D, and PartLLM on a strap, drill, fishing rod, bicycle, and laptop. Input part names are listed at left, and colors show the predicted regions.
| Method | PartNet-E | PartNeXt | 3DCoMPaT200 | |
|---|---|---|---|---|
| Coarse | Fine | |||
| FIND3D | 16.7 | 28.8 | 31.4 | 10.4 |
| PatchAlign3D | 41.4 | 29.3 | 32.8 | 10.3 |
| CoSMo3D | 17.6 | 37.9 | 48.8 | 26.1 |
| Ours | 50.1 | 78.7 | 86.1 | 80.0 |
Table 5. Quantitative results on text-guided part segmentation.
| PartNeXt | 3DCoMPaT200 | HY3D-Bench | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Coarse | Fine | |||||||||||||||
| Method | @1 | @3 | @5 | @7 | @1 | @3 | @5 | @7 | @1 | @3 | @5 | @7 | @1 | @3 | @5 | @7 |
| Point-SAM | 36.6 | 49.0 | 55.0 | 60.3 | 45.6 | 65.9 | 71.1 | 74.3 | 37.3 | 46.9 | 49.4 | 51.6 | 20.7 | 26.2 | 29.7 | 31.1 |
| P3-SAM | 43.8 | - | - | - | 47.7 | - | - | - | 49.9 | - | - | - | 31.8 | - | - | - |
| S2AM3D | 40.8 | - | - | - | 62.6 | - | - | - | 43.3 | - | - | - | 39.8 | - | - | - |
| PartSAM | 45.9 | 61.3 | 65.0 | 69.6 | 51.3 | 76.8 | 79.6 | 81.8 | 50.6 | 66.5 | 71.8 | 73.5 | 34.6 | 45.1 | 49.6 | 50.4 |
Table 6. Quantitative results on interactive part segmentation. IoU@ denotes mean IoU after interaction rounds.
Figure 9. Qualitative comparison with baseline methods ( Zhou et al., 2025 ; Ma et al., 2025 ; Zhu et al., 2026 ; Su et al., 2025 ; Li et al., 2026 ) on one-click interactive part segmentation. Yellow dots denote prompt points, and red regions denote predicted masks. One-click interactive segmentation comparison among Point-SAM, P3-SAM, PartSAM, S2AM3D, SegViGen, and PartLLM on three objects. Yellow dots mark the input clicks and red regions show the predicted masks.
| Dataset | PartSAM | P3-SAM | SegviGen | Ours |
|---|---|---|---|---|
| FAUST | 55.3 | 35.0 | 37.7 | 72.2 |
Table 7. Category-agnostic full-shape segmentation on FAUST.
| Method | FAUST | AKBSeg |
|---|---|---|
| PartSLIP | 48.6 | 22.6 |
| ZeroPS | 49.5 | 33.5 |
| SATR | 79.8 | 29.0 |
| PatchAlign3D | 66.5 | 35.6 |
| Ours | 87.6 | 54.8 |
Table 8. Text-guided segmentation on unseen real-world scans.
Figure 10. Results on real-world scans with noise and partial structure. Despite missing surfaces and capture artifacts, PartLLM still produces usable semantic part decompositions. PartLLM results on noisy and partial real-world scans.
| Variant | 3DCoMPaT200-Fine | PartObjaverse-Tiny | PartNet-E | |||
|---|---|---|---|---|---|---|
| Text | Interactive | Full | Semantic | |||
| w/ SAM-style decoder | 76.2 | 57.5 | 62.9 | 54.6 | 67.0 | 47.7 |
| w/o coarse 3D location | 55.4 | 46.3 | 57.6 | 38.1 | 61.2 | 34.8 |
| w/o semantic label | 62.0 | 53.8 | 61.3 | – | 65.0 | 38.8 |
| Full PartLLM | 80.0 | 60.7 | 74.2 | 63.1 | 78.8 | 50.1 |
Table 9. Ablation study on method design choices. All entries use mIoU except Semantic, which uses SA-mIoU.
Figure 11. Qualitative ablation of the decoding strategy. Compared with the SAM-style decoder that predicts masks independently, our decoder produces cleaner mutually exclusive regions with sharper boundaries. Side-by-side qualitative comparison of a SAM-style decoder and the PartLLM decoder on a drone and scissors. Enlarged regions show that the PartLLM decoder produces cleaner boundaries between adjacent parts.
Figure 12. Scaling curve on the fine-grained split of 3DCoMPaT200. Line plot of Full, Text, and Interactive scores as training data scale increases.
| Task | Setting | PartSAM | Ours |
|---|---|---|---|
| Full | Canonical | 69.5 | 78.8 |
| Rotated | 67.9 | 78.3 | |
| Interactive | Canonical | 34.6 | 51.2 |
| Rotated | 35.3 | 50.7 |
Table 10. Robustness to random 3D rotations. We report full-shape mIoU on PartObjaverse-Tiny and one-click mIoU on HY3D-Bench.
| Dataset | All-part | Single-part |
|---|---|---|
| PartNet-E | 50.1 | 53.3 |
| PartNeXt | 78.7 | 77.0 |
| 3DCoMPaT200-Coarse | 86.1 | 87.4 |
| 3DCoMPaT200-Fine | 80.0 | 78.2 |
Table 11. Text-guided mIoU with all-part and single-part queries.
| Training data | Evaluation | ||||||
| Full | Text | Interactive | Geometry | Full | Semantic | Text | Interactive |
| ✓ | 56.2 | 48.3 | - | - | |||
| ✓ | ✓ | 62.7 | 58.5 | 72.0 | - | ||
| ✓ | ✓ | ✓ | 67.2 | 61.8 | 77.5 | 56.5 | |
| ✓ | ✓ | ✓ | ✓ | 74.2 | 63.1 | 80.0 | 60.7 |
Table 12. Ablation study on training data on 3DCoMPaT200-Fine.
| Method | Runtime (s) | mIoU | |||
|---|---|---|---|---|---|
| PartField ( Liu et al., 2025 ) | 10 | 35.0 | |||
| PartSAM ( Zhu et al., 2026 ) | 12 | 42.2 | |||
| P3-SAM ( Ma et al., 2025 ) | 10 | 40.3 | |||
| Ours | 4.8 | 7.1 | 11.7 | 16.3 | 74.2 |
Table 13. Runtime comparison on the full-shape segmentation task on 3DCoMPaT200-Fine. denotes the number of parts.
Figure 13. Editing applications enabled by PartLLM. A target part can be selected either interactively or through a semantic text query such as Wheel , and the resulting mask can serve as a direct handle for downstream material and geometry editing. Part-aware editing application figure showing part selection by interactive prompting or text query, followed by material editing and geometry editing on the selected part.
Figure 14. Representative failure case of category confusion. When the object-level prior is incorrect, PartLLM may produce a semantically coherent but wrong decomposition. Adding the object category to the prompt can correct the prediction. Failure case figure showing a router that is incorrectly decomposed with bed-related part names, followed by a corrected prediction when the object category is provided in the prompt.
More in this topic
3D Part Segmentation
3D part segmentation divides objects and scenes into meaningful components using geometry, appearance, language, and motion cues.
See all 35 papers