Boreal-Flux-Dev-V2 Model Documentation

Introduction

Boreal-Flux-Dev-V2 is a text-to-image model designed with specific enhancements over previous versions. It operates using the LoRA (Low-Rank Adaptation) technique, tailored for generating photorealistic images.

Architecture

The model is built upon the "black-forest-labs/FLUX.1-dev" base model. It employs a unique LoRA approach to adjust its performance and handle dataset variations. This version addresses previous issues, such as the latent shift dot problem, and supports dynamic thresholding with high negative guidance.

Training

The model was trained with a novel approach to its dataset, intended to mitigate prior issues and enhance image realism. The training parameters were adjusted, aiming for balance but resulting in reduced creativity compared to earlier iterations.

Guide: Running Locally

To run Boreal-Flux-Dev-V2 locally:

1. Download the Model Weights:

   • Access the Safetensors format weights from the Files & Versions tab on the model's page.

2. Setup Environment:

   • Install necessary libraries such as PyTorch and Hugging Face Transformers.
   • Ensure that you have Python set up on your system.

3. Running the Model:

   • Load the model using your preferred framework or environment.
   • For optimal performance, use a cloud GPU service like AWS EC2, Google Cloud, or Azure to handle the computational demands.

4. Image Generation:

   • Use "photo" as a trigger word in prompts to initiate image generation.
   • Consider adjusting the strength settings below 1.0 to avoid overtraining artifacts.

License

The Boreal-Flux-Dev-V2 model is available under the license specified in the repository. Please check the repository for detailed licensing information.