Project Details
Year
2024
Location
UCLA, Los Angeles, CA
Category
Academic
Type
Master's Research
Role
Individual Research
Instructor
Laure Michelon
Scale/Area
AI / Machine Learning Architectural Study
Tools Used
Design Narrative
An individual research project exploring the potential of fine-tuning custom LoRA (Low-Rank Adaptation) models on curated architectural datasets using Stable Diffusion 1.5 in ComfyUI. The study investigates two distinct aesthetic outputs — photorealistic architectural rendering and hyper-stylized 'color-pop' visualisations — through systematic manipulation of ControlNet LineArt and Depth inputs, CFG scale, denoise strength, and LoRA weights. Custom Python scripts automate the generation pipeline, enabling rapid iteration across facade studies, entrance views, and pool terraces.
Project Gallery
[ Photorealistic vs. surreal pop-art generation contrast ]
[ AI-generated vermilion canopy pavilion in a desert landscape ]
[ Entrance view — ControlNet LineArt, Depth, realistic, and color-pop outputs ]
[ Front facade — ControlNet wireframe, depth map, and dual render comparison ]
[ Pool facade — water caustics and glass shader translation through diffusion ]
[ Full ComfyUI node graph — CLIP loaders, KSampler, VAE decode, LoRA stack ]
[ V.2 study — red-and-white beach house, hero exterior render ]
[ V.2 study — facade materiality and massing variation ]
[ V.2 study — alternate lighting and site condition pass ]
[ V.2 study — interior-to-exterior threshold exploration ]
[ V.2 study — closing frame of the beach house sequence ]