auraflow v0.3 pvc style lora

Introduction

The AuraFlow V0.3 PVC Style LoRA is designed for generating PVC style images using text-to-image diffusion models. It integrates with ComfyUI and may not be compatible with other WebUIs.

Architecture

The model is based on the AuraFlow-v0.3-bnb-nf4 architecture with LoRA (Low-Rank Adaptation) applied to specific attention and MLP layers. The model employs bitsandbytes NF4 quantization, targeting specific keys while excluding others like "text_encoder" and "vae."

Training

The model was trained using the auraflow-ft tool on an RTX 4070Ti Super GPU for about 62 hours. Training utilized 7,452 images with danbooru-style tags, employing techniques such as gradient checkpointing and RAdamScheduleFree optimizer. The LoRA settings include a rank of 16 and an alpha of 1, with a learning rate of 1e-3.

Guide: Running Locally

1. Prerequisites: Ensure you have a compatible environment with tools like ComfyUI and auraflow-ft.
2. Download Model: Obtain the Safetensors format weights from the Files & versions tab on the Hugging Face model page.
3. Install Dependencies: Set up necessary libraries and dependencies as per your environment's needs.
4. Run Inference: Use the model with pvc, pvc figure, figma, or nendoroid as trigger words in your prompts for optimal output.
5. Cloud GPUs: For large-scale inference, consider using cloud GPUs such as AWS EC2 P3 instances or Google Cloud TPUs.

License

The model is licensed under Apache-2.0, allowing for broad use and distribution with attribution.