PIU: Proximity-guided Identity Unlearning in ID-Conditioned Diffusion Models
Paper • 2605.22311 • Published
image imagewidth (px) 512 512 | file_name stringlengths 18 18 | cluster_id int64 0 7.49k | cluster_size int64 1 25 |
|---|---|---|---|
celebahq_21253.png | 0 | 6 | |
celebahq_13544.png | 1 | 13 | |
celebahq_13936.png | 2 | 6 | |
celebahq_01743.png | 3 | 9 | |
celebahq_21986.png | 4 | 7 | |
celebahq_12304.png | 5 | 17 | |
celebahq_17828.png | 6 | 11 | |
celebahq_20413.png | 7 | 16 | |
celebahq_20861.png | 8 | 8 | |
celebahq_25898.png | 9 | 15 | |
celebahq_00103.png | 10 | 5 | |
celebahq_11211.png | 3,951 | 1 | |
celebahq_26258.png | 11 | 3 | |
celebahq_23974.png | 3,952 | 1 | |
celebahq_23506.png | 12 | 11 | |
celebahq_18528.png | 13 | 4 | |
celebahq_03016.png | 14 | 8 | |
celebahq_06748.png | 15 | 6 | |
celebahq_19368.png | 3,953 | 1 | |
celebahq_22346.png | 3,954 | 1 | |
celebahq_27418.png | 3,955 | 1 | |
celebahq_10823.png | 3,956 | 1 | |
celebahq_10451.png | 16 | 8 | |
celebahq_07108.png | 17 | 10 | |
celebahq_02656.png | 18 | 11 | |
celebahq_22893.png | 19 | 4 | |
celebahq_20072.png | 20 | 9 | |
celebahq_12765.png | 21 | 6 | |
celebahq_00910.png | 22 | 5 | |
celebahq_00562.png | 3,957 | 1 | |
celebahq_01485.png | 3,958 | 1 | |
celebahq_13125.png | 23 | 17 | |
celebahq_21632.png | 24 | 6 | |
celebahq_01322.png | 3,959 | 1 | |
celebahq_13682.png | 25 | 13 | |
celebahq_21195.png | 3,960 | 1 | |
celebahq_08869.png | 3,961 | 1 | |
celebahq_02590.png | 26 | 10 | |
celebahq_19709.png | 27 | 14 | |
celebahq_10030.png | 3,962 | 1 | |
celebahq_27079.png | 28 | 11 | |
celebahq_22727.png | 29 | 15 | |
celebahq_02237.png | 3,963 | 1 | |
celebahq_07569.png | 30 | 4 | |
celebahq_10797.png | 31 | 8 | |
celebahq_22080.png | 32 | 5 | |
celebahq_23167.png | 33 | 4 | |
celebahq_26639.png | 34 | 22 | |
celebahq_11670.png | 35 | 4 | |
celebahq_18149.png | 36 | 7 | |
celebahq_14689.png | 3,964 | 1 | |
celebahq_06329.png | 37 | 24 | |
celebahq_03477.png | 3,965 | 1 | |
celebahq_03805.png | 3,966 | 1 | |
celebahq_18198.png | 38 | 23 | |
celebahq_03301.png | 39 | 16 | |
celebahq_23611.png | 40 | 8 | |
celebahq_14658.png | 41 | 13 | |
celebahq_11106.png | 3,967 | 1 | |
celebahq_02541.png | 42 | 12 | |
celebahq_02933.png | 43 | 6 | |
celebahq_10746.png | 44 | 7 | |
celebahq_15018.png | 45 | 2 | |
celebahq_22051.png | 46 | 2 | |
celebahq_01826.png | 47 | 8 | |
celebahq_01454.png | 48 | 4 | |
celebahq_13653.png | 3,968 | 1 | |
celebahq_21144.png | 49 | 9 | |
celebahq_05938.png | 50 | 10 | |
celebahq_00214.png | 51 | 8 | |
celebahq_20704.png | 52 | 3 | |
celebahq_12013.png | 53 | 4 | |
celebahq_22397.png | 54 | 10 | |
celebahq_10480.png | 3,969 | 1 | |
celebahq_02120.png | 55 | 7 | |
celebahq_22842.png | 56 | 13 | |
celebahq_22430.png | 57 | 3 | |
celebahq_15479.png | 58 | 2 | |
celebahq_10327.png | 3,970 | 1 | |
celebahq_02687.png | 59 | 2 | |
celebahq_03760.png | 3,971 | 1 | |
celebahq_26289.png | 60 | 20 | |
celebahq_06799.png | 61 | 3 | |
celebahq_11915.png | 62 | 6 | |
celebahq_11567.png | 63 | 6 | |
celebahq_14239.png | 64 | 5 | |
celebahq_23270.png | 65 | 10 | |
celebahq_00675.png | 66 | 12 | |
celebahq_12472.png | 67 | 5 | |
celebahq_12800.png | 3,972 | 1 | |
celebahq_25849.png | 3,973 | 1 | |
celebahq_20365.png | 68 | 14 | |
celebahq_21282.png | 3,974 | 1 | |
celebahq_13595.png | 69 | 4 | |
celebahq_01035.png | 70 | 10 | |
celebahq_21525.png | 71 | 5 | |
celebahq_21957.png | 72 | 2 | |
celebahq_13232.png | 73 | 2 | |
celebahq_01792.png | 74 | 10 | |
celebahq_08235.png | 75 | 7 |
This dataset accompanies PIU: Proximity-guided Identity Unlearning in ID-Conditioned Diffusion Models. It contains 27,996 CelebA-HQ images at 512x512, their cluster annotations, and the paper's ArcFace artifacts.
data/*.parquet: images, filenames, cluster IDs, and cluster sizes.paper_artifacts/embeddings.npy: one embedding per image.paper_artifacts/srk_labels_eps0.35.npy: the identity labels used in the paper.paper_artifacts/srk_centroids_eps0.35.npy: the corresponding identity centroids.The Parquet rows map to the paper arrays by row order; file_name identifies each image. Use srk_labels_eps0.35.npy to reproduce the paper. Parquet cluster_id values are a separate export and should not be assumed to match the paper label IDs.
The labels are automatically inferred with cosine-distance DBSCAN (eps=0.35, min_samples=2). They are not official CelebA-HQ identity labels.
7ecc6a45edfb5483ccf2f7df1035d298ffe7c76b