arjunsingh-neeved/FFHQ-Dataset-FaceCrops

FFHQ single-face crops (SCRFD-10G)

Face crops of juhi-neeved/FFHQ-Dataset (revision a96f3a94e7), same columns and labels (image, binary_label, class_id, binary_name, class_name; all rows real / 0).

Preprocessing

1. Decode, convert to RGB.
2. Detect faces with SCRFD-10G (det_10g.onnx, ONNX Runtime CUDA, 2 x T4, batch 8/GPU) at analysis sizes 128, 320, 640 px.
3. Keep a row only if exactly one face with confidence >= 0.8 is found at >= 2 sizes and those boxes agree with IoU >= 0.75.
4. Single person only: rows where any size finds 2+ faces (score >= 0.8) are removed.
5. Take the highest-confidence box, add 15% padding per side, clip to the image, crop the original.
6. Save as lossless PNG at the crop's native size (no resizing, no alignment). Labels unchanged.

53,012 of 70,000 rows kept.

outcome	rows
kept	53,012
single_face_at_too_few_scales	16,021
no_face	526
multiple_faces	376
scales_disagree_iou	65

Per-image decisions (scores per size, IoU, detected box, crop box, source row) are in preprocessing/manifest.parquet; the code is in preprocessing/.

FFHQ is by NVIDIA (Karras et al., 2019), released under CC BY-NC-SA 4.0; individual images carry their original Flickr licenses.