iso Cities

Introduction

The isoCities v1 model is a specialized adaptation of the Stable Diffusion 1.5 model, designed to generate detailed isometric city illustrations. It was created by merging an isometric city model with Stable Diffusion 1.5 using Automatic1111's checkpoint merger tool.

Architecture

The model leverages the Stable Diffusion framework to provide enhanced capabilities for creating digital art representations of isometric cities. It is trained to respond to prompts like "an illustration of an isometric city, digital art, highly detailed."

Training

The model was developed using a WSL2 Ubuntu machine and trained with DreamBooth. Key hyperparameters include:

• Prior Preservation: Enabled
• Loss Weight: 1.0
• Prompts: "an illustration of an isometric city, digital art, highly detailed"
• Resolution: 512
• Batch Size: 1
• Mixed Precision: FP16
• Learning Rate: 5e-6
• Scheduler: Constant
• Max Train Steps: 1500

Sample images from the training process showcase the model's capabilities.

Guide: Running Locally

1. Setup Environment: Ensure that your environment has access to a machine with GPU capabilities for optimal performance.
2. Download Model Checkpoint: Obtain the model checkpoint from the provided link here.
3. Install Dependencies: Install necessary libraries such as PyTorch and Hugging Face's Transformers.
4. Load Model: Use the downloaded checkpoint to load the model in your script.
5. Run Inference: Provide text prompts as input to generate isometric city images.

For improved performance, consider using cloud GPUs from providers like AWS, Google Cloud, or Azure.

License

The isoCities v1 model is licensed under the CreativeML OpenRAIL-M license. Key points include:

• Prohibition against generating illegal or harmful content.
• Freedom to use generated outputs; however, users are accountable for their usage.
• Redistribution of model weights and commercial use are allowed, provided the same license terms are included. Users must read the full license for complete details.