NVIDIA Cosmos

Multi-Agent Systems 💻 Jupyter Notebook ⚖️ Apache-2.0 🟡 Quiet lately
470 stars

Cosmos-Predict1 is a collection of world foundation models from NVIDIA for predicting future states of visual environments, including diffusion and autoregressive models for Text2World and Video2World generation. It provides pre-trained models, tokenizers, and post-training scripts for Physical AI builders to generate or simulate visual worlds. This repository is no longer actively developed; users are directed to Cosmos 3.

NVIDIA Cosmos demo
🖼️ Screenshot from the project README

✨ Key features

  • Diffusion-based Text2World and Video2World generation
  • Autoregressive Video2World generation
  • Image and video tokenizers (continuous and discrete)
  • Post-training scripts for custom datasets
  • Multi-GPU and multi-node support for inference and training
  • Includes safety filters via Llama Guard 3

🎯 Use cases

  • Generate synthetic visual simulations from text prompts
  • Predict future frames from video prompts
  • Create multi-view clips from single videos
  • Tokenize videos for efficient processing
  • Post-train models for domain-specific world prediction

📦 Installation

🧰 Requirements: Requires NVIDIA GPU and CUDA environment; see INSTALL.md for details. Models are available on Hugging Face.

Please refer to INSTALL.md for general instructions on environment setup.

🚀 Usage

Refer to the examples in the repository for inference and post-training. For example, see Inference with diffusion-based Text2World models.

⚠️ Good to know

This repository is no longer under active development and will receive only limited maintenance updates; future work is focused on Cosmos 3.

❓ FAQ

What is the difference between Cosmos-Predict1 and Cosmos 3?

Cosmos 3 is the next-generation platform with stronger world prediction and unified capabilities, while Cosmos-Predict1 is the older version and is no longer actively developed.

What types of models are included?

Diffusion-based models for Text2World, Video2World, and WorldInterpolator, autoregressive models for Video2World, and image/video tokenizers.

Can I post-train the models with my own data?

Yes, post-training scripts are provided for diffusion and autoregressive models, including multi-node support.

What is the license for the code and models?

Code is under Apache 2.0, models under NVIDIA Open Model License; custom licenses available on request.

📊 Repository

Stars★ 470
Forks🍴 84
Open issues🐛 13
Last commit🕒 Jun 7, 2026
Created📅 Mar 2025
Language💻 Jupyter Notebook
License⚖️ Apache-2.0

🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.