Rekognition draws the boxes for the people AWS's diffusion model inserts into real scenes

A structured prompt tells Qwen-Image-Edit-2509 where to put the person: on tracks, on top of equipment, hanging from an edge, or standing in a vehicle's path ^[12]. The generator does not return coordinates. Amazon Rekognition's DetectLabels API reads the finished image and produces the bounding box for the person that was just inserted ^[4]. Rekognition never saw the prompt.

Label quality on the hardest examples therefore depends on a second detector finding a person in exactly the poses that make detection hard. Those are the poses the dataset exists to cover ^[10]. An inserted worker that DetectLabels misses drops out of the training set, or comes back with a loose box, and the pipeline has no annotator in the loop to catch either case ^[4].

Editing real photographs instead of rendering whole scenes is the strongest choice in the design. The insertion preserves background, lighting and scale ^[3]. AWS's reason is domain gap, which the post defines as the performance drop when a model trained on one distribution, such as fully synthetic scenes, is applied to real-world images ^[6]. Existing equipment annotations also stay valid, so the machinery in each scene does not have to be relabeled ^[5].

At the rates the post quotes, a 100,000-image tail set costs $300,000 to $500,000 and takes 50 days of one annotation team ^[15]. AWS puts the requirement at thousands to millions of images ^[8]. A million images comes to $3 million to $5 million and 500 team-days ^[16]. The generation side is one ml.g5.12xlarge instance with four NVIDIA A10G GPUs and 96 GB of VRAM, 24 GB per GPU ^[11][17].

The insertion spec in the walkthrough randomizes gender for demographic diversity ^[12]. Age and body size are a different distribution, and randomizing gender does not produce them. The post opens on a child near a moving implement, though, and says that placing vulnerable populations including children near active machinery for photography is unsafe, unethical and often impractical ^[10][13].

An improvement of 160 percent is a multiplier of 2.6 on whatever the starting mAP50 was ^[2][14]. For it to transfer, your baseline has to be as starved in the tail classes as theirs, because a low starting number makes a large percentage cheap to earn. The evaluation set has to hold the hazardous scenarios too, or the gain lands where nothing is measuring it. And the detector has to sit in the same weight class as an edge model. That is the post's own reason every training example counts disproportionately: the camera is mounted on the tractor, forklift or railcar, and that caps the size of the model it can run ^[9].