Style2Fab: Functionality-Aware Segmentation for Fabricating Personalized 3D Models with Generative AI

• Global geometric edits can break a printed model’s intended functionality, while identifying functional regions in unfamiliar meshes is difficult for users.Source: §1 · p. 2
• Style2Fab derives a taxonomy from 1000 Thingiverse designs and classifies segments as aesthetic, internally functional, or externally functional.Source: §1 · p. 2
• Its topology-based segmentation supports selective stylization of aesthetic segments while preserving functional segments.Source: §1 · p. 2
• The system targets practical workflows by letting makers verify functional regions before applying styles and fabricating the result.Source: §1 · p. 2
• The paper evaluates the classification against formative-study labels and examines whether Style2Fab helps users modify designs without breaking functionality.Source: §1 · p. 2

2.1 Personalizing Open-Source 3D Designs

• Novice makers struggle to modify shared 3D designs because mesh-based workflows require expertise and existing models often lack sufficiently exposed customizable features.Source: §2.1 · p. 2

2.2 Supporting users in Fabrication Workflows

• Fabrication systems address connectors, physical interfaces, mechanisms, balance, and other post-fabrication properties through specialized tools and optimization.Source: §2.2 · p. 2

2.3 Functionality-Aware Design in 3D Models

• Parametric and intent-aware systems support manufacturable customization, but most online meshes lack the metadata required for such manipulations.Source: §2.3 · p. 3
• Style2Fab extends segment correspondence and reuse methods by using segment-level similarity to classify functionality and separate aesthetic from functional elements.Source: §2.3 · p. 3

2.4 Data-Centric Methods for 3D Manipulation

• Deep-learning approaches enable mesh segmentation, semantic relationships, text-guided manipulation, localization, texture application, and 3D generation.Source: §2.4 · p. 3
• The paper focuses these data-centric capabilities on stylizing aesthetic mesh regions without compromising functionality.Source: §2.4 · p. 3

3 FORMATIVE STUDY

• The study qualitatively codes 1000 Thingiverse designs to identify descriptors of functionality and develop a taxonomy.Source: §3 · p. 3, §3.1 · p. 3

3.1 Data Collection

• Thingiverse sampling targets fabrication-oriented variety, while preprocessing retains standardized meshes and yields 993 things containing 10,945 unique models.Source: §3.1 · pp. 3–4

3.2 Inductive Taxonomy Development

• The taxonomy distinguishes Artifact and Task-Related models, with each potentially containing single or multiple assembled components.Source: §3.2 · p. 4, Figure 2 · p. 4
• External context captures how a segment interfaces with the environment, whereas internal context captures how it interfaces with other segments to affect functionality.Source: §3.2 · p. 4, Figure 2 · p. 4
• Segments lacking both internal and external context are considered aesthetic because they do not affect the model’s functionality.Source: §3.2 · p. 4

3.3 Deductive Functionality Classification

• Two annotators independently classified models using Thingiverse metadata and usage images, achieving Cohen’s Kappa of 0.94 before negotiating disagreements.Source: §3.3 · p. 4
• Removing models without aesthetic segments, including printer-calibration models, produced a final labeled dataset of 938 models.Source: §3.3 · p. 4
• The classification uses external-context rows and internal-context columns to count Thingiverse designs across functionality dimensions.Source: Table 1 · p. 4

• The method defines a segmentation problem, classifies internal and external functionality, tunes influential hyperparameters, and evaluates classification against formative-study labels.Source: §4 · pp. 4–5
• Unsupervised spectral segmentation leverages mesh geometry to predict the number of segments separately for each model.Source: §4 · p. 5
• The geometry-based approach is intended to generalize across 3D-printable models with diverse functionality.Source: §4 · p. 5
• Evaluation measures accuracy in predicting functional-segment counts and robustness across a wide range of mesh resolutions.Source: §4 · p. 5

4.1 Segmentation Approach

• The method represents a mesh as a face-adjacency graph and partitions it into k nonoverlapping subgraphs with consistent features.Too few segments can merge functionally distinct components, while too many can split one functional component into incompatible parts.Source: §4.1 · p. 5

4.1.1 Predicting the Number of Segments.

• Spectral segmentation constructs weighted face similarities from geodesic and angular distances, then uses the graph Laplacian’s eigenvectors and k-means clustering to assign faces to segments.Large eigenvalue gaps indicate weak graph connectedness.Source: §4.1.1 · p. 5
• The system estimates k from the mesh eigenvalue distribution, counting eigenvalues whose dispersion exceeds the distribution’s mean plus standard deviation.Higher dispersion indicates more varied connectivity and potentially more distinct segments, whereas tightly clustered eigenvalues support fewer clusters.Source: §4.1.1 · p. 5
• Because the approach depends on uniform mesh resolution, all models are remeshed before segmentation to avoid incoherent partitions caused by uneven face density.Runtime increases with resolution, motivating the search for a low resolution that preserves stable segmentation.Source: §4.1.2 · p. 5

4.1.2 Uniform Mesh Resolution for Segmentation.

• 88% of models stabilized their predicted segment count at 25K vertices, with 72 seconds average segmentation time versus 102 seconds at 30K.The authors selected 25K as the lowest tested resolution whose segment count remained unchanged at higher resolutions.Source: §4.1.2 · p. 5

4.2.1 Using Similarity as a Heuristic:

• Contextual similarity multiplies segment-level and parent-mesh-level topology similarity, producing a value from 0 for complete mismatch to 1 for identical topology.Comparing parent meshes reduces noise from geometrically similar segments used differently in different models.Source: §4.2.1 · p. 6, §4.2 · p. 6
• The external classifier labels a segment by uniformly voting over its five most similar labeled segments, after pruning the search to a small informative mesh set.Classification accuracy was found to converge after five comparisons, while larger mesh sets increase computation and noise.Source: §4.2.1 · p. 6

4.2.2 Classifying External Functionality.

• External-functionality classification achieves 74% precision and 88% recall using five similar meshes, favoring recall because modifying functional components is more costly than missing aesthetic ones.The labeled dataset comprised 1,151 segments from 137 single-component models, with 0.97 inter-rater reliability.Source: §4.2.2 · pp. 6–7

4.2.3 Classifying Internal-Functionality.

• Internal functionality is assigned when segments in different component meshes exceed similarity threshold α, with the highest-similarity partner forming the linkage.Both linked segments receive the internal-functionality label.Source: §4.2.3 · p. 7
• Internal-linkage detection uses ground truth from 50 multi-component models and selects α=0.86, yielding 64% precision and 86% recall.The threshold maximized functional-segment identification before minimizing aesthetic-segment misidentification.Source: §4.2.3 · p. 7

4.3 Stylization of Segments

• Text2Mesh optimizes vertex colors and normal-direction displacements so rendered views increasingly match a natural-language style prompt.The optimization uses a CLIP-based neural representation and gradient descent across different viewing angles.Source: §4.3 · p. 7
• Global Text2Mesh stylization can alter functional-segment topology and render an articulated cat inoperable, whereas Style2Fab masks gradients on functional vertices.The system still permits color and displacement changes on unmasked aesthetic vertices and runs the optimization for 1500 iterations.Source: §4.3 · p. 7

4.4 Style2Fab User Interface and Workflow

• Style2Fab organizes preprocessing, segmentation and classification, selective styling, and review into four Blender-plugin workflow menus.The workflow is designed to let users inspect and revise automated functionality decisions before stylization.Source: §4.4 · p. 8

4.4.1 Pre-Processing and Segmentation.

• The preprocessing interface standardizes resolution, automatically estimates k, colors segments for inspection, and supports manual re-segmentation or parallel processing of multiple meshes.The default resolution is 25K faces, based on the evaluation.Source: §4.4.1 · p. 8

4.4.2 Functionality-Verification of Segments.

• Users can review predicted functional segments, toggle external-functionality labels, and inspect or correct internal linkages between component meshes.For internal assignments, the Assembly panel supports separating incorrect pairs and adjusting α between 0 and 1.Source: §4.4.2 · p. 8

4.4.3 Selective Stylization of Aesthetic Elements.

• After verification, users enter a natural-language style description to stylize aesthetic segments and can review, iterate, or re-segment the resulting model.The completed model is rendered alongside the original for comparison.Source: §4.4.3 · p. 8

5 USER STUDY

• Eight participants segmented and stylized Thingiverse models with or without automatic classification support, covering single- and multi-component Artifact and Task-Related models.Each participant completed experimental and control conditions with counterbalanced model combinations.Source: §5 · pp. 8–9

5.1 Findings

• Style2Fab significantly reduced task completion time across all model types, with the greatest effects on complex multi-component and externally contextual Task-Related models.The within-subject t-tests were significant at p<0.05.Source: §5.1 · pp. 9–10, Table 3 · p. 9
• Automatic classification improved users’ functional-segment accuracy as model complexity increased, whereas manual users had more difficulty locating functional segments.The comparison used collaboratively annotated and fabrication-verified ground truth.Source: §5.1 · p. 9, Figure 7 · p. 9, Table 4 · p. 9
• Participants generally found Style2Fab helpful because its classifications and segmentation made stylization feel more precise and increased confidence in post-fabrication functionality.Participants also reported that results could be less effective or too broad for some complicated models.Source: §5.1 · pp. 9–10

6 DEMONSTRATIONS

• Across six scenarios, Style2Fab personalized Task-Related models in home decor, medical or assistive applications, and personal accessories.The demonstrations illustrate functionality-aware stylization across four model categories.Source: §6 · p. 10

6.1 Home Interior Design

• Self-watering planters and dispensers retained stability, self-watering, liquid-distribution, and other functional geometry while their aesthetic regions were stylized.For the planter, preserved regions included the base, watering cavity, and component-assembly features; for the dispenser, they included the base and interior cavities.Source: §6.1 · p. 10

6.2 Medical/Assistive Applications

• A thumb splint preserved its skin-contact surface and breathable holes while its exterior was styled, and a tactile heart graphic received distinct textures for touch-based region identification.These examples apply functionality-aware personalization to medical and assistive models.Source: §6.2 · p. 10

6.3 Personalizing Accessories

• An AirPods cover preserved internal geometry and its charging hole, while a whistle preserved its resonant chamber and mouthpiece and retained acoustic functionality after exterior styling.The globally styled whistle lost functionality because internal geometry was manipulated.Source: §6.3 · p. 10

7 DISCUSSION

• Semi-automatic segmentation separates aesthetic and functional segments, reducing the need for hyperparameter trial-and-error or vertex-level highlighting while retaining model functionality.The method is designed to be modular and adaptable to improved underlying classifiers.Source: §7 · p. 10

7.1 A Broader Definition of Functionality

• Functionality is currently limited to high-level external and internal contexts, although makers may repurpose models through novel affordances and usage contexts.The authors define functionality as topology-dependent and acknowledge that usage context is ignored.Source: §7.1 · p. 11
• A broader affordance taxonomy could improve specificity, but rare affordances are difficult to classify and require functionality labels beyond the current external/internal categories.The discussion contrasts standardized deep-learning datasets with the long-tail diversity of online fabrication designs.Source: §7.1 · p. 11

7.2 Limits of Topological Similarity

• The classifier relies on topology-only segment and model similarity, so alternative multi-factor metrics and more refined semantic labels could yield more robust segmentation.The chosen measure is computationally efficient and avoids expensive human-generated labels, but may be less effective than richer similarity measures.Source: §7.2 · p. 11

7.3 Opportunities for Richer Data Sets

• Improving classification requires larger functionality datasets or better labels and metadata, while the diversity of real-world models may demand datasets orders of magnitude larger for advanced deep learning.The paper notes that current 3D modeling and printing practices do not readily support creating such datasets.Source: §7.3 · pp. 11–12

8 CONCLUSION

• Style2Fab combines functionality-aware segmentation and classification with selective styling to help users personalize 3D-printable models without compromising functionality.The approach is informed by a formative study of Thingiverse models and defines functionality through external and internal contexts.Source: §8 · p. 12