Zero Diffusion

Introduction

ZeroDiffusion is a text-to-image model developed to serve as a robust training base for other models and to facilitate research utilizing zero terminal Signal-to-Noise Ratio (SNR). It includes both a base model and an inpainting variant.

Architecture

• ZeroDiffusion-Base v0.9: A foundational model trained on approximately 20 million samples with zero terminal SNR.
• ZeroDiffusion-Inpaint v0.9: An experimental fine-tuned version of the stable-diffusion-inpainting model, derived from a merge of ZeroDiffusion 0.9.

Training

The models were trained using Google's TPU Research Cloud program, focusing on achieving zero terminal SNR. This setup aims to give researchers a clean model base for further exploration and adaptation.

Guide: Running Locally

1. Model Setup:

   • Download the ZeroDiffusion model files.
   • Obtain the corresponding YAML configuration file and place it in the same directory as the model.

2. Environment Requirements:

   • Install necessary dependencies and configure the environment, possibly using A1111's webui or similar interfaces.
   • Implement CFG rescale using the plugin available at CFG_Rescale_webui.

3. Execution:

   • Ensure the web UI is set to v-prediction mode for optimal performance.

4. Hardware Suggestion:

   • Utilize cloud GPUs like those offered by Google Cloud or AWS for efficient model execution and training.

License

The ZeroDiffusion models are released under the CreativeML OpenRAIL-M license.