Artificial Context — Brad Machado

Type: Graduate design studio — final
Role: Group project
Tools: Grasshopper · Flux.1-dev LoRA · Karamba3D · Wallacei X · Ladybug · MATLAB · Blender
Status: Completed studio project

Biomorphic massing study render in blue, magenta, and gold, with tangled tube-like structural forms. — Fig. 01 — Massing study render, primary material palette.

01 / Process

Method

A Low-Rank Adaptation model was trained on site imagery gathered in Suseong-gu, Daegu, and applied to Flux.1-dev in ComfyUI, producing latent-space views of a city the base model had never seen. Those images are read as depth maps through the Grasshopper Image Sampler to generate topography, sectioned into 2D genotypes, and screened structurally in Karamba3D. Surviving genotypes are stacked, rotated, and lofted by Python script into volumetric phenotypes, optimized in Wallacei X and evaluated for solar response in Ladybug; a neural network trained in MATLAB on the resulting Pareto fronts. Facade, glazing, and structural shell were resolved in Blender.

Two white 3D-printed physical study models of paired gate-like forms, photographed against a blue-white background. — Fig. 02 — 3D-printed study models, paired forms.

Grasshopper visual-scripting canvas showing the facade and structural-system generation logic. — Fig. 03 — Grasshopper definition, facade and structural logic.

Thin colorful wireframe render of the massing form on a black background, showing internal structural lines. — Fig. 04 — Wireframe massing study, structural diagram.

Photorealistic render of two people walking beneath the glossy purple tubular structure at ground level. — Fig. 05 — Occupied render, ground-level view.

Close-up sunset render of teal and magenta tubular forms framing an interior threshold, with figures inside. — Fig. 06 — Occupied render, interior threshold.

02 / Outcome

Result

The first critique asked the only question that matters here — do these images contain context? They did not, at first: the initial LoRA was over-weighted and had to be recalibrated before its output read as Daegu rather than as the base model’s prior. The MATLAB network trained to R = 0.93 across all sets, and the selected forms were carried into a parametric model. Context arrived as a statistical artifact of the training set rather than as a reading of the site, which is the argument the project is making.

Exterior render of the white biomorphic structure against a clear blue sky, black openings punched through its skin. — Fig. 07 — Exterior render, urban context.

Next Project Cyprus Cove →