Cosmos 1.0 Guardrail

Introduction

Cosmos-1.0 Guardrail is a content safety model developed by NVIDIA, designed for generating physics-aware videos and world states for physical AI development. It consists of four main components: an LLM for content safety, a blocklist for filtering, a video content safety filter, and a face blur filter.

Architecture

• Aegis: Built on Llama 2 backbone, this component is fine-tuned for content safety.
• Video Content Safety Filter: Uses an MLP backbone with SigLIP embeddings to classify video frames as safe or unsafe.
• Face Blur Filter: Employs RetinaFace for identifying and pixelating facial features.

Training

The training process for Cosmos-1.0 Guardrail involves the use of NVIDIA's Aegis Content Safety Dataset, encompassing a wide range of critical safety risk categories. The model is intended for use in content moderation, ensuring the safety of input prompts and generated videos.

Guide: Running Locally

1. Setup: Clone the Cosmos repository from GitHub.
2. Installation: Install required dependencies as listed in the repository.
3. Configuration: Configure the model for your specific use case by setting the appropriate parameters.
4. Execution: Run the model using an appropriate runtime environment.
5. Hardware Recommendations: Utilize cloud GPUs such as NVIDIA Ampere, Hopper, or Blackwell for optimal performance.

License

Cosmos-1.0 Guardrail is released under the NVIDIA Open Model License. This license allows for commercial use and creation of derivative models, with NVIDIA not claiming ownership over outputs generated using the models. However, bypassing any technical limitations or safety mechanisms will result in the termination of rights under this license.