Ultralytics and NVIDIA: YOLO26x matches full-data accuracy on fewer labels
Ultralytics and NVIDIA say an agentic training loop combined with synthetic data from Cosmos Data Factory let Ultralytics YOLO26x match the accuracy of full-dataset training on 11 of 19 public industrial-inspection datasets. According to Ultralytics, 53% of the labeled images on those 11 datasets were never needed.
- Ultralytics YOLO26x trained with NVIDIA's agentic pipeline matched full-dataset accuracy on 11 of 19 public inspection datasets.
- On those 11 datasets, 53% of the labeled images were never needed, according to Ultralytics.
- Training was faster on 10 datasets, with a 69 to 74% reduction on DAGM, but slower on nine others.
- Cosmos Data Factory defects raised [email protected] by 0.11 to 0.32 on DAGM classes, with DAGM class 9 rising from 0.650 to 0.966.
- Of 11,448 generated images, only 1,481 were kept for training.
How the training loop picks its data
The approach rests on a simple premise from Ultralytics: more data is not the answer, the right data is. Defect-detection models often handle common defects well but miss the rare scratch, unusual lighting or odd camera angle. The usual remedy, collecting and labeling more of everything, is described as slow and expensive, and it mostly adds images the model already handles well.
NVIDIA's agentic data-enhanced training loop wraps the training process and decides what data the model needs next. According to the post, it does three things:
- Mines real data from existing image pools to fill specific gaps in the training set.
- Generates synthetic data with Cosmos Data Factory where real examples do not exist.
- Labels automatically through a component called AnomalyGen, so synthetic images need no manual annotation before retraining.
Cosmos Data Factory is NVIDIA's data curation, processing and generation platform, built on its Cosmos world models. Ultralytics YOLO is then retrained on the combined real and synthetic set using the standard Ultralytics training pipeline.
Test setup and main results
The Ultralytics machine learning team ran the pipeline on 19 public industrial-inspection datasets covering steel, PCBs, fabric, metal surfaces and consumer products, about 91,000 training images in total. Every run used one 8-GPU host. The baseline and the Cosmos Data Factory model used an identical recipe: Ultralytics YOLO26x, 100 epochs, the same optimizer, learning rate and seed, scored on the same held-out test set. Only the training images differed.
On 11 of the 19 datasets, the agentic pipeline matched the model trained on the full dataset. On the DAGM datasets it reached the same [email protected] on a tenth of the labels, and on two classes it finished slightly ahead. On pcb-defect it dropped 7,000 of 20,924 images, with an accuracy difference four times smaller than run-to-run noise, according to Ultralytics.
Training time also fell on 10 of the 19 datasets. On DAGM it dropped from 13 to 24 minutes down to 4 to 7 minutes, a 69 to 74% reduction. Larger datasets saw smaller savings: 19% on VisA and 12% on metal-surface-defect. On the other nine datasets the pipeline took longer than a single full-data run, because each curated round retrains from scratch and the extra rounds outweigh the savings when the curated subset is not much smaller.
Generated defects and stricter localization
A third test isolated synthetic data on datasets with polygon annotations. Defects generated with Cosmos Data Factory and labeled by AnomalyGen were compared against a model trained on real images only. The metric was [email protected], which counts a prediction only if its box overlaps the true defect by at least 75%, so it measures how precisely a defect is located.
Adding generated defects raised [email protected] by 0.11 to 0.32 on the DAGM classes, and recall improved on every dataset tested. The largest jump was DAGM class 9, from 0.650 to 0.966. On Real-IAD the gain was marginal at +0.008.
Synthetic data was tightly limited. Across 18 generation rounds the pipeline generated 11,448 images, and only 1,481, about one in eight, were kept. The team tested caps equal to 25% or 100% of the real images in the training set, so the cap, not the generator's output, set how much synthetic data a model saw.
What it means for inspection teams
Ultralytics argues that for teams already training Ultralytics YOLO, the approach fits the existing workflow: the same models and the same training recipe, with better-targeted data going in. For teams paying for annotation, the company calls the labeling savings the headline result, since most of the labeling budget on these datasets went into images that did not change the outcome.
The results come with limits stated in the post: the accuracy match held on 11 of 19 datasets, speed gains were not universal, and the benefit of synthetic data varied by dataset. The figures come from Ultralytics' own tests on public benchmarks.
