Akashic Pulse v1.0

Introduction

AkashicPulse v1.0 is a fine-tuned model based on RouWei, designed to produce high-quality text-to-image outputs. It has undergone a series of merging and finetuning steps to ensure superior performance.

Architecture

AkashicPulse is built on the Illustrious-based RouWei model, incorporating advanced techniques to enhance image generation quality.

Training

The training process involved multiple steps:

• Step 1: CyberFix Treatment

  • Application of specialized techniques to improve the base model.

• Step 2: Training New Concepts

  • Dataset size: ~10,000 images
  • Hardware: 2xA100 80GB GPUs
  • Optimizer: AdaFactor
  • Learning Rates: 7.5e-6 (Unet), 3.75e-6 (Text Encoder)
  • Batch Size: 16, Gradient Accumulation: 3
  • Loss: Huber with SNR Schedule

• Step 3: Finetuning I

  • Dataset size: ~4,500 images
  • Hardware: 1xA100 80GB GPU
  • Optimizer: AdaFactor
  • Learning Rate: 3e-6 (Unet)
  • Batch Size: 16, Gradient Accumulation: 3
  • Loss: Huber with SNR Schedule

• Step 4: Finetuning II

  • Dataset size: ~4,500 images
  • Hardware: 1xA100 80GB GPU
  • Optimizer: AdaFactor
  • Learning Rate: 3e-6 (Unet)
  • Batch Size: 48, Gradient Accumulation: 1
  • Loss: L2 with Noise Offset of 0.0357

Guide: Running Locally

1. Setup Environment

   • Ensure you have a Python environment set up with necessary dependencies (e.g., PyTorch).

2. Download Model

   • Access the model from the Hugging Face repository.

3. Run Inference

   • Utilize the recommended prompting format for best results:
     • Example Prompt: 1girl, green hair, sweater, looking at viewer, upper body, beanie, outdoors, night, turtleneck, masterpiece, best quality
     • Negative Prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, signature, watermark, username, blurry

4. Recommended Hardware

   • Cloud GPUs such as NVIDIA A100 are suggested for optimal performance.

License

AkashicPulse v1.0 is distributed under the Fair AI Public License 1.0-SD. More details can be found here.