The project implies a platform that can later replace an architectural firm. The user sets a description of the city to be built. The platform draws a small concept of the city (e.g. a waterfront or a futuristic city), based on which a presentation is made to the customer.

Infrastructure Deployment: Create and configure a computational server for model inference and model training.

Software Environment Setup: Prepare and configure the software environment for model inference and model training.

Data preparation: Generating training datasets for general city plans (Aerial View) and close building plans, preparing data for preferred styles.

Model training and testing: Implementation of LoRA approach for training of auxiliary neural networks, training of test LoRA modules on trained data and testing of modules performance for different base models.

Integration with Automatic WebUI: Integrate models with Automatic WebUI software to provide a user-friendly interface for working with models and implementing additional functions.

Infrastructure deployment: A Linux-based compute server with an Nvidia Tesla video adapter was used to infer models. A Nvidia Tesla video adapter was used to train the models.

Software environment setup: For inference, an environment based on Python 3.11, Pytorch 2.0.0, bitsandbytes and xformers libraries were used. The kohya-ss software was used to implement the LoRA training. The LoRA (Low-Rank Adaptation) method allowed us to efficiently train small auxiliary neural networks, which are then connected to the main Stable Diffusion model and guide the generation process. This made it possible to flexibly switch such modules for specific tasks without having to re-train the entire model. Automatic WebUI software was used as the interface for working with the models. This software has an open API, which allowed deep integration and implementation of additional functions.

Data preparation: Training datasets were generated for general city plans (Aerial View) and close building plans. The data was trained for preferred styles.

Model training and testing: LoRA approach was implemented to train auxiliary neural networks. Trained LoRA test modules for the trained data. Testing of the modules’ performance with respect to different bases was carried out testing the performance of the modules with respect to different underlying models.

A Linux-based compute server with an Nvidia Tesla T4 or A4000 video adapter for inference models was successfully configured.

A software environment based on Python 3.11, Pytorch 2.0.0, bitsandbytes and xformers libraries was successfully prepared.

Training datasets were generated for general city plans (Aerial View) and close building plans.

The test LoRA modules were trained on the trained data.

Integration and use of the open API Automatic WebUI was realized to provide a user-friendly interface for working with models and implementing additional functions.