AI Designs 3D Models That Print With Less Waste

A new artificial intelligence framework is tackling one of the most persistent and wasteful problems in 3D printing: the need for temporary support structures. These supports, required to prevent overhanging features from collapsing during fabrication, lead to significant material waste, extended production times, and can leave blemished surfaces on finished objects. While current slicing technologies focus on post-processing optimizations, researchers have developed a system that addresses this issue at the design phase, generating 3D models that inherently require fewer supports from the start. This approach could make digital fabrication more sustainable and efficient, reducing the environmental and economic costs associated with bringing digital creations into the physical world.

The researchers introduced a framework called SEG (Support-Effective Generation), which uses a technique called Offset Direct Preference Optimization (ODPO) to refine 3D generative models. The system trains AI to produce geometries that minimize the volume of support material needed during printing, measured by a metric called normalized support volume (NSV). By incorporating support structure simulation directly into the training process, SEG encourages the generation of models with fewer "risky faces"—areas where overhangs exceed a 45-degree angle and typically require supports. builds on existing diffusion-based 3D generation models but aligns them with practical fabrication constraints that have been largely ignored in previous text-to-3D systems.

To achieve this, the team developed a pipeline that begins with data from the Thingi10k dataset, a collection of 3D printing models. They rendered multi-view images of these models and used Cap3D to generate detailed text captions, creating 9,840 prompt-model pairs for training. The core innovation lies in the ODPO training phase, where support structure simulation is integrated. This simulation uses ray tracing and geometric approximation to estimate support volume by identifying risky faces and calculating the tetrahedral volume extending from these faces to the print bed. The ODPO then applies an offset based on the difference in NSV values between preferred (low-support) and dispreferred (high-support) samples, reshaping the training signal to amplify meaningful differences in support efficiency.

Extensive experiments on two benchmark datasets—Thingi10k-Val and GPT-3DP-Val—demonstrate SEG's effectiveness. As shown in Table I and Table II of the paper, SEG significantly outperforms baseline models including TRELLIS, DPO, and DRO. On Thingi10k-Val, SEG achieved an NSV of 0.176 and an NSV* of 0.587, compared to TRELLIS's NSV of 0.343 and NSV* of 1.255. It also attained a support-efficient consistency (SEC) score of 0.870, indicating reliable reduction in support requirements. Visual in Figure 5 highlight how SEG-generated models (right) show fewer red-highlighted risky areas compared to baseline models (left), and Figure 3 shows physical prints where SEG models (bottom) require substantially less support material than baseline outputs (top).

Of this research extend across industries that rely on 3D printing, from custom manufacturing and healthcare to education and consumer products. By reducing support material waste, SEG can lower costs and environmental impact, while shorter print times increase throughput. The framework maintains high fidelity to input prompts, ensuring that designs remain visually appealing and functionally sound. This alignment of AI creativity with physical constraints represents a step toward more practical and sustainable digital fabrication, where the transition from screen to object becomes smoother and less resource-intensive.

Despite its advancements, the paper notes several limitations. SEG's support simulation relies on the manifold assumption, which may introduce biases with non-manifold meshes or floating facets. primarily targets Fused Filament Fabrication (FFF) printers with a standard 45-degree self-support angle, though it is flexible enough to be extended to other technologies like SLA and SLS. Future work could explore optimizing upright orientation and self-support angles for different printing s, and addressing geometry assurance in 3D generation models to further improve reliability. These limitations highlight ongoing s in bridging AI-generated designs with real-world manufacturing constraints.