Object detectors often have trouble spotting rare objects because these categories have very few training images. This long-tail problem limits their real-world use.
Tail-SD uses a (an AI that generates images from random noise) to create new training pictures for rare objects. It accepts text instructions, object layouts (boxes or masks), or a mix of both, so users can control what appears and where. The model learns from all input types at once, improving quality and diversity.
On the LVIS dataset, rare-category detection accuracy rose by 4.89% for boxes and 5.68% for masks. This shows that controllable synthetic data can help detectors learn rare objects without collecting more real-world examples.