We are excited to share a critical resource for anyone working in Physical AI: the NVIDIA Cosmos Cookbook!
This cookbook is your essential, practical guide to the Cosmos Open Models—providing step-by-step workflows, technical recipes, and concrete, reproducible examples for building, adapting, and deploying world foundation models (WFMs). Our goal is to empower you to quickly reproduce successful model deployments and customize them for your unique domain challenges.
Dive into a few high-impact recipes available this week:
🦋 Overcoming Data Scarcity in Biodiversity (Featuring Voxel51)
Our partner Voxel51 demonstrates how to leverage Cosmos Transfer 2.5 to address data scarcity and domain gaps. This recipe takes a handful of existing moth images and transforms them into rich, field-like agricultural scenes, perfect for training biodiversity AI.
⚽ Sports Video Generation with LoRA Post Training
Generate photorealistic, physically accurate, and rule-aware sports videos with smooth player motions. This recipe showcases a powerful LoRA post-training workflow on Cosmos Predict 2.5 for domain-specific video generation.
- 📖 Full recipe: Predict for Sports Video Generation - Cosmos Cookbook
🚦 Synthetic Data Generation for Traffic Scenarios (End-to-End Workflow)
Explore a complete, end-to-end workflow designed to generate high-fidelity, photorealistic synthetic data specifically for complex traffic scenarios and smart city applications.
- 📖 Full recipe: Synthetic Data Generation for Smart Cities - Cosmos Cookbook
⚙️ Physical Plausibility Prediction (Physics-Savvy Critic)
Learn how to post-train Cosmos Reason 1 into a sophisticated physics-savvy critic. This recipe allows your model to judge whether generated videos obey real-world physical laws, serving as a powerful quality control filter for synthetic data.
- 📖 Full recipe: Reason for Physical Plausibility Check - Cosmos Cookbook
We encourage you to clone the repository, run the examples, and contribute your own innovative recipes to the community!