Jianing Luo
↗

Bio-ComfyUI: A wasp nests-inspired 'Structure-Code DNA' in biomimicry design and fabrication

Type
AI & Generative Design Research
Year
2025 – 2026
Keywords
ComfyUI
Karamba3D
Generative Biomimicry
Structure-Code DNA
AI-Assisted Architectural Design
Structural Optimisation
Finite Element Analysis (FEA)

Abstract

Bio-ComfyUI is a methodological workflow that encodes a "Structure-Code DNA" — a generative logic linking material behaviour, biological growth patterns, and AI-driven form synthesis. The system integrates Finite Element Analysis (FEA), parametric modelling, and multi-conditional diffusion control within ComfyUI to translate structural intelligence into a reproducible design language.

Through four conditioning layers — structural skeleton, topological nodes, semantic bitmap, and stylistic reference — Bio-ComfyUI enables AI to learn and reproduce architectural grammar grounded in both engineering logic and aesthetic formation. The workflow closes the loop from AI generation to FEA validation and robotic fabrication, producing morphologies that are visually expressive yet structurally and materially feasible.

Physical prototypes confirm that the encoded DNA carries embedded structural intelligence, allowing design, analysis, and making to co-evolve within one adaptive system. The research repositions generative AI as an active morphogenetic collaborator, advancing computational biomimicry toward an integrated, self-learning design-to-fabrication ecology.

Read the full paper ↗

Architectural-scale generative design using Bio-ComfyUI
Figure 1 — Biological Reference & Structure-Code DNA
Figure 1 — Biological Reference & Structure-Code DNA
Wasp-nest morphology, biomimetic precedents (Gaudí’s Sagrada Família and the Biorock pavilion), and the pose-and-segmentation logic that encodes form as a Structure-Code DNA. Nature’s self-organised intelligence becomes the seed grammar for the workflow.
Figure 2 — Multi-Condition Generative Framework
Figure 2 — Multi-Condition Generative Framework
The end-to-end frame: typological optimisation (GA + FEA in Karamba3D), semantic training, and stylistic control via ControlNet, ComfyUI, and Hunyuan 3D — moving from form-finding through training iteration to 3D reconstruction.
Figure 3 — Four-Layer Conditioning Logic
Figure 3 — Four-Layer Conditioning Logic
(a) Structural skeleton from Karamba3D; (b) topological nodes; (c) semantic bitmap; (d) stylistic reference. Together these four channels form the conditioning inputs that steer the diffusion process toward structurally valid morphologies.
Figure 4 — Structural Skeleton Dataset
Figure 4 — Structural Skeleton Dataset
FEA-informed structural skeletons generated in Grasshopper / Karamba3D. This curated set teaches the LoRA network the load-path logic underlying mechanically plausible forms.
Figure 5 — Topological Nodes
Figure 5 — Topological Nodes
Manually annotated nodal hierarchies marking intersections and force aggregation. These act as biological growth nodes that govern branching behaviour and load distribution.
Figure 6 — Semantic Bitmap & Ground Truth
Figure 6 — Semantic Bitmap & Ground Truth
Colour-coded segmentation distinguishing core, skin, and canopy regions, giving the diffusion model spatial density cues for differentiated material articulation.
Figure 7 — Hunyuan AI in the ComfyUI Workflow
Figure 7 — Hunyuan AI in the ComfyUI Workflow
The integrated 2D-to-3D module: multi-conditioned diffusion outputs are reconstructed into coherent 3D mesh geometry, closing the loop toward structural and fabrication analysis.
Figure 8 — Prototype, FEA & Fabrication
Figure 8 — Prototype, FEA & Fabrication
Generated prototype, FEA validation, and robotic-extrusion / sand-based 3D-printed 1:8 models. Structural members thicken in high-stress zones exactly as predicted, confirming the Structure-Code DNA feedback loop.
Figure 9 — Architectural-Scale Generation
Figure 9 — Architectural-Scale Generation
Large-scale verification: the workflow translates AI-generated morphologies into a structurally coherent, fabrication-ready architectural form, where dense columns, porous skins, and sheltering canopies emerge from a single generative logic.
← Back to Research