Additive manufacturing (AM) of wood composites enables the reuse of sawdust, a byproduct of the wood industry, and offers potential reductions in construction-related carbon emissions, supporting a circular economy. Although material researchers have examined the material formulation of 3D printed wood, focusing on binder selection and mechanical characterization, there is a gap in exploring the design possibilities of 3D-printed wood. In contrast to traditional orthogonality of wood structures, this paper presents a systematic workflow for exploring topologically optimized (TO) structural wood-composite slabs for 3D printing, benchmarked against cross-laminated timber (CLT). Although the current mechanical properties of 3D printed wood are inferior to CLT, this study evaluates structural adequacy under present conditions and explores the role of TO in achieving material efficiency. To validate the simulation approach, we recreated the Gatty Wool Factory slab designed by Nervi in a CAD environment (Rhino) and analyzed its deformation and stress behavior in ANSYS, establishing a baseline for slab design and structural performance analysis. Next, a CLT slab was modeled using the same dimensions and boundary conditions as a benchmark. Methyl-cellulose (MC) was selected as the binder for the 3D printed wood-composite slab, and a set of topology-optimized wood-MC slabs was generated using the Topos plug-in in Grasshopper and evaluated in ANSYS. Several variants demonstrated structural performance comparable to the CLT benchmark. The lightest design has an average depth of 303mm, using almost twice the mass of CLT but less than half the material of an equivalent concrete slab. Its cradle-to-gate global warming potential (GWP) is estimated to be 1.3 times that of CLT and 0.45 times that of concrete, indicating that 3D-printed wood composites can provide a structurally viable and environmentally competitive alternative. This study contributes a validated methodology for linking material innovation, computational design, and structural benchmarking in support of sustainable construction.

Generative diffusion models have rapidly entered architectural practice, producing visually persuasive imagery through statistical inference over vast datasets. Yet these models remain fundamentally representational systems whose outputs frequently lack disciplinary specificity, tectonic logic, or environmental reasoning. This paper introduces a methodological framework for fine-tuning generative AI within architectural design workflows through curated datasets, Low-Rank Adaptation (LoRA) training, and Midjourney AI mood-boards. Rather than treating AI as a mere image generator, the research positions it as a specialized design collaborator trained through domain-specific visual knowledge. The study proposes a multi-scale generative workflow in which architectural design tasks are structured across four levels of spatial reasoning: material scale, modular unit scale, architectural element scale, and spatial system scale. At each stage, generative models are guided through curated datasets and fine-tuned models to ensure that AI outputs progressively transition from material articulation to spatial organization. The framework was implemented within an experimental elective course in architectural design, where students developed Midjourney mood-board datasets, trained LoRA models, and employed Midjourney-based generative processes to produce design artifacts across these scales. The pedagogical experiment demonstrates that fine-tuned generative AI can move beyond stylistic speculation toward structured design reasoning when guided through scale-aware workflows. The study contributes to emerging discourse on AI in architectural design by proposing a model of computational authorship grounded in dataset curation, model calibration, and multi-scale spatial reasoning, reframing the architect as an orchestrator of generative intelligence.

Deep Pixels presents a two-phase design research project that tests the use of generative AI for design ideation on building surfaces, particularly facades. This exploratory work asks how AI-generated images relate to the production of geometry that can be aligned with real buildings, rationalized, and translated into physical prototypes, with the goal of outlining how AI might contribute to a conceptual framework of facade tectonics.

Phase 1 focuses on 2D-to-3D translation through two methods that are directly compared. The first tests a NeRF and Gaussian splat point-cloud reconstruction pipeline capable of producing meshes from AI-generated video inputs. Although conceptually robust, this pipeline introduces frequent reconstruction instabilities such as irregular point densities, topological discontinuities, and view-dependent artifacts. The second method employs a direct image-to-mesh workflow using Midjourney for image generation and ComfyUI with Hunyuan 3D-2.0 for mesh synthesis. The comparative evaluation shows that the mesh-native workflow is more stable and computationally efficient and becomes the primary pipeline for further exploration.

Phase 2 applies these insights at the architectural scale, focusing on facade design. ChatGPT structures context-specific prompts that encode structural information, site conditions, and references to Detroit-based artists. These prompts are combined with elevation and section drawings in Midjourney to generate oriented facade images. Selected images are converted to meshes, registered to the building grid, panelized, thickened, and 3D-printed as sectional fragments, while attachment strategies are examined through exploded axonometric diagrams.

The results indicate that generative AI can serve as a meaningful collaborator in facade design when paired with mesh-native workflows, contextual prompts, and physical prototyping. As these tools advance and gain greater capacity to interpret multi-material assemblies, their collaborative potential will grow. The tested pipeline shows the potential for GenAI to contribute to facade design in ways that are both conceptually and materially coherent.