EDM2 Synthetic Skin Lesion Pool with Per-Image Quality Scores

275,472 synthetic dermoscopic images (512 x 512 PNG) of the seven ISIC-2018 diagnostic classes (AKIEC, BCC, BKL, DF, MEL, NV, VASC), generated with EDM2 trained on the ISIC-2018 training data (task 3), together with per-image quality scores. The images come from 16 generation configurations (4 training checkpoints x 4 classifier-free guidance scales), 17,217 images each. Files: 16 zip archives (ckpt_<K>k_cfg_<X>.zip), metadata.csv (one row per image, with the generation parameters, 14 per-image metric scores and the similarity to the nearest real training image), config_metrics.csv (four distributional metrics for each configuration), README.md (file layout and column descriptions) and SHA256SUMS. The images are synthetic and are for research use only, not for clinical use. The similarity column shows proximity to training images and is not a privacy guarantee.