Results¶

Quantitative and qualitative results demonstrating the effectiveness of CIA for synthetic data augmentation.

Pipeline Overview¶

CIA pipeline overview

The CIA framework generates high-fidelity synthetic images by extracting structural conditions from real images and using them to guide Stable Diffusion + ControlNet generation.

Quantitative Results¶

COCO People Detection¶

COCO all samplings

Synthetic augmentation improves object detection on COCO people across multiple sampling strategies.

Training Loss¶

COCO training loss

Models trained with CIA-augmented data converge effectively.

Flickr30K¶

Flickr30K all samplings

Consistent improvements observed on the Flickr30K entity detection task.

Quality Assessment¶

IQI measure

Synthetic images generated by CIA maintain high perceptual quality as measured by image quality assessment metrics.

Sampling Strategies¶

Random sampling

Comparison of different synthetic image selection strategies. Filtering by PTD scores (top-k) selects the most distribution-faithful samples, improving downstream task performance.