B R I A 4 B Adapt

Introduction

BRIA-4B-Adapt is a text-to-image model featuring 4 billion parameters, designed for commercial fine-tuning. It offers comprehensive legal liability coverage, ensuring ethical use of content without copyright infringements. The model is customizable, supporting various business needs while maintaining high alignment with prompted styles.

Architecture

BRIA-4B-Adapt is a Latent Flow-Matching Text-to-Image Model developed by BRIA AI. It is trained on a professional-grade, licensed dataset, ensuring compliance with commercial licensing terms and conditions. The model supports extensive customization through access to source code and weights.

Training

The model can be fine-tuned using a LoRA (Low-Rank Adaptation) method, which allows specific style or character training. When training a new LoRA:

• Use diverse yet consistent images to generalize across domains.
• Maintain a minimum image resolution of 1024x1024.
• Use concise and descriptive captions, ideally less than 128 tokens.
• Consider hyperparameters like "rank" and "max_train_steps" to fine-tune model details.

Guide: Running Locally

1. Installation:

   • Install dependencies with:

     pip install -qr https://huggingface.co/briaai/BRIA-4B-Adapt/resolve/main/requirements.txt
     

   • Download required scripts using hf_hub_download.

2. Training:

   • Use the provided script train_lora.py to fine-tune the model on your dataset.

3. Inference:

   • Load the trained LoRA into the Bria Pipeline and generate images using prompts.
   • Adjust parameters like prompt, negative prompt, and resolution.

4. Hardware Suggestions:

   • Utilize cloud GPUs such as those offered by AWS, Google Cloud, or Azure to handle intensive computations.

License

BRIA-4B-Adapt requires a commercial license for access and use. The model is covered under specific terms and conditions detailed here. Full legal liability coverage is provided, ensuring compliance with copyright and privacy regulations.