Retro Pixel Flux Lo R A

Introduction

The Retro-Pixel-Flux-LoRA model by prithivMLmods is a text-to-image model utilizing the LoRA method for generating pixel art images. It supports creating images with a retro, pixelated aesthetic using prompts.

Architecture

The model is based on the "black-forest-labs/FLUX.1-dev" base model and utilizes LoRA (Low-Rank Adaptation) for fine-tuning. Key parameters include:

• Learning Rate Scheduler: Constant
• Optimizer: AdamW
• Network Dim: 64
• Network Alpha: 32
• Epochs: 15
• Image Dimensions: Optimally 1024 x 1024

Training

The model is trained with 16 high-resolution images and uses parameters like multires noise and noise offset for improved image quality. It is still in the training phase, which means some artifacts may appear, and the performance might vary.

Guide: Running Locally

1. Setup Environment:

   • Ensure you have torch and pipelines installed.

2. Initialize Model:

   import torch
   from pipelines import DiffusionPipeline
   
   base_model = "black-forest-labs/FLUX.1-dev"
   pipe = DiffusionPipeline.from_pretrained(base_model, torch_dtype=torch.bfloat16)
   
   lora_repo = "prithivMLmods/Retro-Pixel-Flux-LoRA"
   trigger_word = "Retro Pixel"  
   pipe.load_lora_weights(lora_repo)
   

3. Device Configuration:

   device = torch.device("cuda")
   pipe.to(device)
   

4. Weights Download:

   • Weights are available in Safetensors format. Download them from the Files & versions tab.

Cloud GPUs: To handle large models efficiently, consider using cloud GPU services like AWS EC2, Google Cloud, or Azure.

License

The Retro-Pixel-Flux-LoRA model is licensed under the CreativeML OpenRAIL-M license.