Starsector Portraits

Introduction

The Starsector Portraits project by Severian-Void provides various models trained to generate images in the style of the Starsector video game. These models are fine-tuned to enhance mod makers' capabilities in creating character portraits with specific styles and features using Stable Diffusion technology.

Architecture

The models utilize Stable Diffusion pipelines, specifically through diffusers. They include a series of versioned models and hypernetworks that are incrementally improved through fine-tuning and various training strategies, such as prior preservation loss.

Training

The models are trained using a labeled dataset of 556 images, with additional versions incorporating prior preservation loss using a set of 8,000 portrait images. Different model versions focus on various aspects such as style, clothing, and faction-specific attributes. Hypernetworks are also fine-tuned to improve the quality and accuracy of the generated portraits.

Guide: Running Locally

1. Setup Environment: Ensure your system has Python installed. Set up a virtual environment and install the necessary packages, including diffusers and other dependencies.

2. Download Model: Clone the model repository from Hugging Face and navigate to the appropriate model version folder.

3. Execute Model: Run the image generation script with desired prompts. Use keywords such as "starsectorportrait" to activate model-specific features.

4. Hardware Recommendations: For efficient performance, especially with larger model versions, consider using cloud GPUs available through AWS, Google Cloud, or Azure.

License

The models are licensed under the CreativeML OpenRAIL-M license. Users must comply with the terms and conditions specified under this license, particularly when distributing or using the models for commercial purposes.