# NVIDIA Details Two-Prompt Cosmos Workflow for Physical AI Data

By Simon Yoon

Canonical URL: https://www.tokenpost.com/news/technology/61053
Published: 2026-11-19T02:47:23.000Z
Updated: 2026-11-19T02:47:23.000Z
Section: Technology

> The Nov. 18 walkthrough adds estimates for autonomous-driving clips and defect-image dataset assembly. Its cost comparison pairs a generated clip with a road-collected scenario, which may not be equivalent.

NVIDIA has added performance estimates to a two-prompt walkthrough for Cosmos Data Factory, a toolkit for preparing training data for robots and autonomous vehicles using developers’ own datasets.

The first prompt directs a coding agent to clone and configure the workflow orchestration tools. A second prompt selects a blueprint and runs it against the developer’s data. Five reference blueprints cover autonomous-vehicle dataset generation, image attribute augmentation, video augmentation, event-video generation and defect-image generation. The workflows use six composable skills.

Examples include finding similar scenes in driving footage and adding weather variations, generating event videos from a source image, and creating defect images for inspection training. NVIDIA’s Cosmos Evaluator checks generated video for physical plausibility, and some workflows automatically label their outputs.

For its benchmark, NVIDIA used a cloud server with eight NVIDIA RTX PRO 6000 Blackwell Series GPUs. The company estimated the server cost at about $40 per server-hour, or $960 per day, and the server generated about 130 autonomous-driving edge-case clips daily.

NVIDIA priced road collection at $125 per road-collected scenario. The comparison pairs a generated clip with a road-collected scenario; the measures are not established as equivalent units. Separately, NVIDIA’s example reduced defect-image dataset assembly from two months to one day, a 60-fold speedup.

The cost estimates cover GPU compute. Source-data preparation, setup, agent use, storage and human review may add expenses. The blueprints can run on users’ GPUs on premises or in the cloud, but require infrastructure such as Kubernetes, object storage and a container registry. GPU needs vary by workflow.

NVIDIA outlined the five blueprints in an Oct. 21 Cosmos update, covered in [an earlier story on the toolkit](<https://www.tokenpost.com/news/technology/38125>). The Nov. 18 walkthrough adds instructions for launching the workflows with a coding agent and NVIDIA’s cost and assembly-time estimates.



## Links in this article

- [an earlier story on the toolkit](https://www.tokenpost.com/news/technology/38125)
