logo-color
Contact Us
logo-white
Contact Us
logo-color

Musashi AI Adopts NVIDIA Cosmos Synthetic Data Generation to Accelerate AI Visual Inspection Deployment

WATERLOO, Ontario — June 30th, 2026

Musashi AI North America, developer of the Cendiant® AI-powered visual inspection platform, today announced it is working with NVIDIA to integrate NVIDIA Defect Image Generation skill powered by NVIDIA Cosmos world foundation model for synthetic data generation (SDG) for industrial defect detection. Initial results from Musashi AI’s in-house evaluation indicate that synthetic defect images generated from a limited set of real production samples can credibly reproduce challenging defect types — a capability that could meaningfully reduce the time and data required to deploy high-accuracy inspection models in production environments.

A persistent obstacle in deploying AI-based visual inspection is the scarcity of real defect data. Many defect types occur rarely, vary widely in appearance, and can take months to accumulate in sufficient volume to train reliable models — extending deployment timelines. Synthetic data generation addresses this by using a small number of real defect examples as seed data to generate large volumes of realistic, varied training images, without waiting for defects to occur on the production line.

Evaluation Using Musashi Production Data

For its initial evaluation, Musashi AI trained NVIDIA Cosmos models on defect image data from Musashi’s own production operations, spanning eight defect classes including porosity, cracks, scratches, and dents. Training ran on Musashi AI’s on-premises infrastructure using four RTX PRO 6000 Blackwell Max-Q Workstation Edition GPUs.

In engineering review, both models produced realistic synthetic defects, with the larger model showing modest advantages on harder cases such as dents and cracks. The compact 2-billion parameter model offered a particularly attractive balance of quality and training time.

An initial focus was generational data transfer within a single application: whether defect data from a customer’s existing Gen 1 Cendiant system can generate equivalent training data for that same application on the higher-resolution Gen 2 platform. Several Musashi AI customers operate both generations, and carrying existing defect knowledge forward, rather than restarting data collection, would accelerate Gen 2 deployments and improve time-to-value.

Real production image (left) alongside a synthetic crack generated by NVIDIA Defect Image Generation skill (right). Source: Musashi AI internal evaluation.
Real production image (left) alongside a synthetic crack generated by NVIDIA Defect Image Generation skill (right). Source: Musashi AI internal evaluation.

Data Governance

All images used in this evaluation originate from Musashi’s internal production operations. In customer deployments, each customer’s inspection data is used solely for that customer’s own applications and remains protected under existing confidentiality agreements. Synthetic data generation does not involve combining or transferring data between customers; rather, it reduces the amount of data any individual customer needs to collect in the first place.

Next Phase: Rare Defects and Edge Cases

The next phase will apply synthetic data generation to rare edge cases and boundary conditions — defect types for which little or no real-world data exists, and where obtaining representative samples can be impractical or impossible. Musashi AI sees significant potential for SDG to improve model robustness in exactly these situations.

“Data availability, not algorithm capability, is often the gating factor in bringing AI inspection into production. Reducing the data burden on our customers is one of the most meaningful ways we can accelerate the adoption of automated inspection while lowering upfront costs,” said Edward van Amstel, Managing Director, Musashi AI North America. “Our early work with NVIDIA’s synthetic data generation models reflects how we are investing in the underlying AI capabilities that make the Cendiant platform faster to deploy and stronger over time.”

Musashi AI is continuing the evaluation and will share further findings as the work progresses. The company expects synthetic data generation to play a growing role in its AI development workflows and future product capabilities.

About Musashi AI

Musashi AI North America is a hardware- and software-focused technology company that builds and develops AI-powered machine vision solutions for quality assurance in manufacturing environments. Based in Waterloo, Ontario, the Musashi Technical Centre employs a multidisciplinary team of research, development, and applications engineers focused on advancing automated inspection technology through innovative hardware, software, and artificial intelligence solutions.

Early evaluation shows synthetic defect data generated with NVIDIA Defect Image Generation skill powered by NVIDIA Cosmos can recreate real-world manufacturing defects, with the potential to reduce data collection requirements and shorten customer deployment timelines for the Cendiant® inspection platform, accelerating return on investment.

WATERLOO, Ontario — June 30th, 2026

Musashi AI North America, developer of the Cendiant® AI-powered visual inspection platform, today announced it is working with NVIDIA to integrate NVIDIA Defect Image Generation skill powered by NVIDIA Cosmos world foundation model for synthetic data generation (SDG) for industrial defect detection. Initial results from Musashi AI’s in-house evaluation indicate that synthetic defect images generated from a limited set of real production samples can credibly reproduce challenging defect types — a capability that could meaningfully reduce the time and data required to deploy high-accuracy inspection models in production environments.

A persistent obstacle in deploying AI-based visual inspection is the scarcity of real defect data. Many defect types occur rarely, vary widely in appearance, and can take months to accumulate in sufficient volume to train reliable models — extending deployment timelines. Synthetic data generation addresses this by using a small number of real defect examples as seed data to generate large volumes of realistic, varied training images, without waiting for defects to occur on the production line.

Evaluation Using Musashi Production Data

For its initial evaluation, Musashi AI trained NVIDIA Cosmos models on defect image data from Musashi’s own production operations, spanning eight defect classes including porosity, cracks, scratches, and dents. Training ran on Musashi AI’s on-premises infrastructure using four RTX PRO 6000 Blackwell Max-Q Workstation Edition GPUs.

In engineering review, both models produced realistic synthetic defects, with the larger model showing modest advantages on harder cases such as dents and cracks. The compact 2-billion parameter model offered a particularly attractive balance of quality and training time.

An initial focus was generational data transfer within a single application: whether defect data from a customer’s existing Gen 1 Cendiant system can generate equivalent training data for that same application on the higher-resolution Gen 2 platform. Several Musashi AI customers operate both generations, and carrying existing defect knowledge forward, rather than restarting data collection, would accelerate Gen 2 deployments and improve time-to-value.

Real production image (left) alongside a synthetic crack generated by NVIDIA Defect Image Generation skill (right). Source: Musashi AI internal evaluation.
Real production image (left) alongside a synthetic crack generated by NVIDIA Defect Image Generation skill (right). Source: Musashi AI internal evaluation.

Data Governance

All images used in this evaluation originate from Musashi’s internal production operations. In customer deployments, each customer’s inspection data is used solely for that customer’s own applications and remains protected under existing confidentiality agreements. Synthetic data generation does not involve combining or transferring data between customers; rather, it reduces the amount of data any individual customer needs to collect in the first place.

Next Phase: Rare Defects and Edge Cases

The next phase will apply synthetic data generation to rare edge cases and boundary conditions — defect types for which little or no real-world data exists, and where obtaining representative samples can be impractical or impossible. Musashi AI sees significant potential for SDG to improve model robustness in exactly these situations.

“Data availability, not algorithm capability, is often the gating factor in bringing AI inspection into production. Reducing the data burden on our customers is one of the most meaningful ways we can accelerate the adoption of automated inspection while lowering upfront costs,” said Edward van Amstel, Managing Director, Musashi AI North America. “Our early work with NVIDIA’s synthetic data generation models reflects how we are investing in the underlying AI capabilities that make the Cendiant platform faster to deploy and stronger over time.”

Musashi AI is continuing the evaluation and will share further findings as the work progresses. The company expects synthetic data generation to play a growing role in its AI development workflows and future product capabilities.

About Musashi AI

Musashi AI North America is a hardware- and software-focused technology company that builds and develops AI-powered machine vision solutions for quality assurance in manufacturing environments. Based in Waterloo, Ontario, the Musashi Technical Centre employs a multidisciplinary team of research, development, and applications engineers focused on advancing automated inspection technology through innovative hardware, software, and artificial intelligence solutions.

logo-color