Industrial safety models need examples of people standing near moving machines, yet the most valuable scenes are too dangerous to stage and too rare to collect naturally. An AWS demonstration uses image generation to add synthetic people to real equipment photographs, producing training data without placing anyone in harm’s way.
The pipeline runs Qwen-Image-Edit-2509 on Amazon SageMaker to insert people in locations such as tracks, vehicle paths and equipment edges. Editing an existing photograph preserves the real background, lighting and machinery labels. Amazon Rekognition then finds the inserted person and converts the bounding box into the format used to train an object detector.
AWS reports an improvement of up to 160% in mean average precision at a 50% overlap threshold. The process also avoids manual annotation, which the article estimates can cost $3 to $5 per image. The test used four Nvidia A10G GPUs for image generation and an 80% confidence threshold for automatic labels.
Synthetic imagery can still contain visual artifacts, biased placements or missed labels. A large percentage gain may also start from a weak baseline and does not guarantee safe performance on a tractor, forklift or railcar. The method is best used to supplement difficult real examples, followed by validation on cameras and conditions matching the deployed equipment.