Evolutionary optimisation + generative AI · UCL Digital Ecologies · 2025
Reimagining
Jeddah.
Performance-driven urban form for a hot, sprawling city — visualised by a model that knows what Hejazi architecture looks like.
Evolutionary optimisation + generative AI · UCL Digital Ecologies · 2025
Performance-driven urban form for a hot, sprawling city — visualised by a model that knows what Hejazi architecture looks like.
01 · Genetic algorithm
A plot-threshold gene grows streets, parcels, land use and volumes. NSGA-II searches it.
Site
Parcels
Land use
Simulation 1 — Pareto-front phenotypes
Evolution across generations
Three selected individuals and their objective ranks
Layout → massing
Best solution — hierarchical streets, varied heights
Phenotype catalogue
02 · AI visualisation
A captioned dataset of coral-stone houses, Shibam towers and modern-traditional façades, fine-tuned into a LoRA. Segmentation and depth from the massing condition every image.
Dataset — classes × views
Palette extracted from references
Five checkpoints — 4 to 29 epochs
Conditioning — segmentation + depth from the massing
Without / with LoRA — same prompt, same seed
Denoising, frame by frame
Denoising — second seed
03 · Image to 3D
Hunyuan3D 2.0 lifts generated façades into textured meshes; octree resolution benchmarked against time.
Input
Normals
Textured mesh
Turntable — 01, single image
Turntable — 02
Turntable — 03
Turntable — multi-image cluster
Octree 32 → 640 — chunks and seconds
Multi-image workflow
04 · Ray-cast texture projection
A ray from each camera to each vertex; if nothing occludes it, the pixel colour is written to the vertex. Written in Python inside Grasshopper.
Five cameras, five images
Vertex ownership by camera
100 quads
500 quads
1,000 quads
Retextured — camera 1
Camera 2
Camera 3
05 · Film
Final film — from GA layout to textured 3D to AI walkthrough
Clip — image-to-video from a LoRA render
Clip — plaza at dusk