This repository contains code and data used to train a Lora (Low-Rank Adaptation, an efficient finetuning approach for large Diffusion models) on a dataset of Roman mosaics images. The majority of the images were sourced from the Museo Nazionale Romano - Palazzo Massimo in Rome, some of them are my own photos.

To use the training script, you need to clone the Diffusers library into the folder of this repository. The Diffusers library provides necessary functionality for the training process.

1. Clone this repository to your local machine.
2. Clone the Diffusers library into the repository folder.
3. Run the training script to train the Lora model on the Roman mosaics dataset.
4. Convert the resulting Lora model to a WebUI lora using the convert_to_webui.py script. (A cell in the notebook is dedicated to this step, but it is not necessary to run it.)

The resulting Lora model has been converted for use in the SD WebUI by automatic1111. The file that is usable as a WebUI lora is roman_mosaics_webui.safetensors.

For more information, please refer to the HuggingFace repo of the Lora.