EcoDiff Pruned Models

Introduction

EcoDiff Pruned Models is a project focused on providing pruned versions of popular text-to-image models using the EcoDiff technique. This involves reducing the model complexity by selectively removing neurons, resulting in more efficient models with a 20% pruning ratio.

Architecture

The pruned models are based on two primary architectures:

• FLUX.1-schnell by Black Forest Labs
• Stable Diffusion XL Base 1.0 by Stability AI

Models are stored in pickle files to facilitate in-place neuron removal, leveraging the diffusers library for model loading.

Training

The training process involves pruning the models using the EcoDiff method, which systematically removes 20% of the neurons to enhance efficiency while maintaining performance.

Guide: Running Locally

To run these models locally, follow these steps:

1. Clone the Repository: Download the models from the Hugging Face repository.
2. Install Requirements: Ensure that all necessary Python packages and dependencies are installed.
3. Load the Model: Use the diffusers library to load the pruned model.
4. Run Inference: Execute the model to generate images from text prompts.

For optimal performance, it is recommended to use cloud GPUs, such as those provided by AWS, Google Cloud Platform, or Azure.

License

The EcoDiff Pruned Models are released under the Apache 2.0 License, allowing for both personal and commercial use with appropriate credit.