NVIDIA Cosmos-Predict-1

À propos de NVIDIA

Fondée 1993 · Santa Clara, California, USA

NVIDIA is the dominant supplier of GPUs for AI training and inference and runs a large in-house research organisation across robotics, simulation, and generative modelling. NVIDIA Cosmos was announced at CES 2025 as a family of 'World Foundation Models' (WFMs) for Physical AI - models that predict how the physical world evolves given video, language, and action conditioning. Cosmos is positioned as a developer platform for robotics and autonomous-vehicle teams to generate synthetic training data, run policy evaluations in simulation, and bootstrap Vision-Language-Action (VLA) pipelines. The 'Predict-1' track focuses on diffusion-based video-future prediction conditioned on text and/or first-frame inputs and ships in 7B and 14B parameter variants with open weights under the NVIDIA Open Model License.

Architecture

Diffusion-based world foundation model (text/video-to-video) for Physical AI

Cosmos-Predict-1 is a diffusion world model that predicts future video frames conditioned on text prompts, a starting frame, or short context clips. It uses a 3D causal video tokenizer (Cosmos Tokenizer) to compress video into spatio-temporal latents, then runs a Diffusion Transformer in latent space with cross-attention to text embeddings produced by a T5-XXL encoder. Training data is a curated corpus of ~20 million hours of driving, robotics, and human-activity video, filtered for motion quality, captioning coverage and safety. The model is not itself a VLA controller, but is the world-model backbone of NVIDIA's Cosmos stack: Cosmos-Predict generates rollouts; Cosmos-Reason adds VLM reasoning over predicted futures; and Cosmos-Transfer adapts simulation-to-real video. In a VLA pipeline it provides synthetic 'imagined' trajectories and dense reward / value signals, and is used to evaluate manipulation and driving policies offline at scale.

Paramètres

7B and 14B variants (Predict-1)

Capacités

• Predicts future video conditioned on text, image, or video context
• Two open-weight variants: Predict-1-7B and Predict-1-14B
• Generates physically plausible motion for driving, manipulation, and humanoid scenes
• Integrates with NVIDIA Isaac, Omniverse, and DRIVE pipelines
• Used as synthetic data engine for VLA / autonomy training
• Supports prompt upsampling via Cosmos-Reason VLM
• Cosmos Tokenizer (3D causal VAE) can be reused as a video encoder
• Released alongside Cosmos-Reason and Cosmos-Transfer for full Physical AI stack
• Best for: synthetic data, world-model research, robotics simulation, AV training.

Entraînement et licence

Trained on ~20 million hours of curated physical-world video (driving, robotics manipulation, humanoid / first-person, navigation) sourced from licensed and open datasets, with multi-stage filtering for motion quality, caption alignment and safety. Text conditioning uses T5-XXL embeddings.

Licence: NVIDIA Open Model License - research-only / developer use with restrictions; weights downloadable from Hugging Face and NGC.

Tests de sécurité: NVIDIA applied dataset-level filters for unsafe content and faces, plus output filters and watermarking guidance. The model is positioned as a research artifact and not a deployed user-facing product, so no formal RSP-style policy is published.