PAIDF AnomalyGen is a diffusion-based synthetic data generation pipeline that manufactures photorealistic, mask-aligned anomaly images for industrial visual inspection, even when only a small set of real examples are available rather than a large labeled dataset. Built on NVIDIA Cosmos-Predict2 Text2Image, it fine-tunes on a few real defect examples so a new defect class can be learned quickly, keeping the heavy diffusion backbone frozen while training only lightweight added modules. At generation time the model paints a defect of a specified texture and type into a clean reference image at the exact location marked by a placement mask, producing photorealistic anomalies that inherit the surrounding product texture and arrive paired with the masks and labels that downstream training needs. The pipeline offers three simple modes—fine-tune then generate, fine-tune only, or generate from an existing checkpoint—so teams can scale a defect library from a handful of real samples to thousands of labeled synthetic ones.

A defining feature of AnomalyGen is its broad model flexibility paired with built-in, automated quality control. It supports both the smaller and larger Cosmos-Predict2 Text2Image backbones and offers three mask-placement strategies that govern where a defect can appear—across the whole image, within a text-prompted region of interest, or constrained to a precise component shape—giving fine spatial control for repeated components, restricted regions, and shape-sensitive defects. Output quality is governed by an automated similarity score that measures how closely each generated defect matches the real defect set, complemented by secondary diagnostics for added confidence. To push quality higher, an agent-steered per-sample search refines generation settings over multiple rounds and concentrates its effort on the weakest samples, while a final quality gate automatically discards low-scoring results and regenerates replacements so each run delivers a full batch at the requested count. Every run emits the synthetic images alongside per-sample quality scores, generation parameters, and aggregate evaluation logs, organized for easy review and reuse.

AnomalyGen is delivered as an agent-orchestrated skill rather than a network service, so teams control it in plain language. A user simply asks for what they need—fine-tune on a dataset, or generate a set of anomaly images—and the agent gathers the required parameters up front and drives the full fine-tune-and-generate workflow end to end, with no manual command wrangling. For reproducible deployment it ships containerized images, including a fully air-gapped variant with model checkpoints built in and a hardened product image that runs in a locked-down, non-privileged configuration behind a preflight safety check. A companion browser-based interface is the recommended way to design and validate mask-placement strategies before launching batch runs. Generated datasets are built for downstream consumption and include standard bounding boxes, masks, and captions so the synthetic anomalies can directly train downstream detection and inspection models.

Governing terms

Our containers are governed by the NVIDIA Software License Agreement and Product-Specific Terms for NVIDIA AI Products.