Just type in what features your apartment has, in normal language. ✨
“two bedrooms and two bathrooms”
“three bedrooms with a kitchen adjacent to the dining room”
This demo application has only been trained on basic styles of rooms so far. If you get too complicated the results will not respond well. This will improve with further training.
This application allows simple natural language prompts to generate infinite, never-seen-before architecture layouts.
Layouts are composed using semantic information such as:
• Typology: "a house with two bedrooms and two bathrooms"
• Enumeration: "a house with five rooms"
• Adjacency: "the kitchen is adjacent to a bedroom", "the living room is not adjacent to the bathroom"
• Location: "a house with a bedroom in the north east side"
Layouts are produced complete with coordinate and labeling metadata, allowing for use in 3D rendering programs.
Architext runs a purposely-trained transformer model. Specifically, a fine-tuned GPT-J 6B model was used to produce millions of synthetic layouts (training data). This model was trained on TPUs courtesy of Google. A GPT-Neo 160M model was fine-tuned on this data and powers the above demo running on GPU-time donated by Huggingface.
Our research is supported with Cloud TPUs from Google's TPU Research Cloud (TRC).