Archviz AI
Abstract
An AI-powered platform that combines ComfyUI workflows with custom LoRA models trained on precise architectural representation diagrams. Users generate publication-ready architectural diagrams almost instantly and can edit individual elements directly — powered entirely by generative AI.
Context / Problem
Architectural representation is a discipline with its own visual language: axonometrics, exploded diagrams, section perspectives, analytical overlays. Generic image-generation models fail to produce these with the precision architects require. Archviz AI closes this gap by training custom LoRA adapters on curated datasets of high-quality architectural drawings, grounding the generative process in the conventions of the discipline.
Methodology
LoRAs are trained per representation type (section, axonometric, diagram, render) using annotated architectural drawings. ComfyUI orchestrates multi-step pipelines — sketch-to-diagram, style transfer, element inpainting — exposed through a clean web interface. ControlNet conditioning on line-work ensures geometric fidelity while the LoRA steers aesthetic style. Users can select, regenerate, or edit individual diagram elements without re-running the full pipeline.