Bioinspired123D: Generative 3D Modeling System for Bioinspired Structures

Abstract

Generative AI has made rapid progress in text, image, and video synthesis, yet text-to-3D modeling for scientific design remains particularly challenging due to limited controllability and high computational cost. Most existing 3D generative methods rely on meshes, voxels, or point clouds which can be costly to train and difficult to control. We introduce Bioinspired123D, a lightweight and modular code-as-geometry pipeline that generates fabricable 3D structures directly through parametric programs rather than dense visual representations. At the core of Bioinspired123D is Bioinspired3D, a compact language model finetuned to translate natural language design cues into Blender Python scripts encoding smooth, biologically inspired geometries. We curate a domain-specific dataset of over 4,000 bioinspired and geometric design scripts spanning helical, cellular, and tubular motifs with parametric variability. The dataset is expanded and validated through an automated LLM-driven, Blender-based quality control pipeline. Bioinspired3D is then embedded in a graph-based agentic framework that integrates multimodal retrieval-augmented generation and a vision-language model critic to iteratively evaluate, critique, and repair generated scripts. We evaluate performance on a new benchmark for 3D geometry script generation and show that Bioinspired123D demonstrates a near fourfold improvement over its non-finetuned base model, while also outperforming substantially larger state-of-the-art language models despite using far fewer parameters and compute. By prioritizing code-as-geometry representations, Bioinspired123D enables compute-efficient, controllable, and interpretable text-to-3D generation, lowering barriers to AI driven scientific discovery in materials and structural design.

Figures (6)

• Figure 1: Study overview. a) End to end translation from 1D text descriptions to 3D geometric structures by coupling BioinspiredLLM with the Bioinspired123D pipeline. b) Overview of the Bioinspired123D system, centered on Bioinspired3D, a language model finetuned on a curated dataset of Blender Python scripts and natural language prompts, evaluated using a custom 3D benchmark and integrated within a multimodal, graph based agentic framework. c) Detailed view of Bioinspired3D, highlighting its input representation as natural language prompts and its output representation as executable Blender Python scripts. Generated scripts are extracted and validated by execution within a Blender subprocess.
• Figure 2: Bioinspired dataset overview. a) Common biological structural design elements, including cellular, fibrous, layered, helical, and tubular motifs, and their correspondence to the three parametric classes of bioinspired 3D structures introduced in this work. b) Representative examples of 3D structures illustrating variation across key geometric parameters for each class.
• Figure 3: a) Data processing pipeline powered by LLM distillation, consisting of three phases: diversification of base scripts into multiple coherent variants, embedded reasoning in which each script is packaged as a narrative step by step example, and validation via headless Blender subprocess execution to ensure successful runtime and correct geometry based on visual inspection of rendered outputs. b) Embedding space visualization of base and diversified instruction script pairs generated through the dataset pipeline. c) Instruction dataset generation process. For each script, a base identifier is retrieved to determine the corresponding bioinspired material class (helical, tubular, or cellular). A shape phrase is constructed by sampling from a class specific Bioinspired Word Bank, which is then combined with a randomly sampled instruction template and a primitive word bank containing varied verbs and grammatical forms to generate natural language instructions.
• Figure 4: a) Benchmark performance across models, including the base model, the final finetuned checkpoint Bioinspired3D, with and without RAG. b) Benchmark performance across models, separated by benchmark question difficulty level, along with a stacked bar plot displaying the benchmark composition showing strong concentration of "hard" leveled questions. c) Examples of successful generations for "hard" level questions in which the prompts utilize both new phrasing and request creative extrapolated ideas that go beyond the dataset. d) Depiction of the Bioinspired123D graph-based agentic system. e) Benchmark performance across models, now including Bioinspired123D and state of the art models, GPT-4o-mini and GPT-5-mini, with breakdown by difficulty level.
• Figure 5: a) Examples of evaluation prompts passed to Bioinspired123D with progressive renders from the agent system, refining the structures chronologically from left to right. b) Photographs of 3D printed samples of the novel AI generated 3D structures.