Skip to page content
SAVRN
Search Plan a project

Dataset · Image segmentation

gap-sam-coco-cache

by Truong KhangTruong/gap-sam-coco-cache

Pre-computed High-Dimensional Forensic Latents for ultra-fast Diffusion-Diff forensics training. - imgid (string): Image identifier. - label (int64): Class label (0: authentic, 1: fully synthetic, 2: tampered/inpainted).

Rows—
Configurations—
Size33.5 MB
Licenseapache-2.0
AccessPublicly accessible
Monthly Downloads—

Dataset Card

By Truong, published under apache-2.0, revision e66d7b4c5c48.

Pre-computed High-Dimensional Forensic Latents for ultra-fast Diffusion-Diff forensics training.

Dataset Details

  • Source Dataset: KhangTruong/COCO-inpainted
  • Model Extractor: gap_sam
  • Cached Channels: 256
  • Latent Resolution: 32x32 (input 256x256)
  • Format: Parquet with zstd compression
  • Status: in_progress

Schema / Columns

  • img_id (string): Image identifier.
  • label (int64): Class label (0: authentic, 1: fully synthetic, 2: tampered/inpainted).
  • z_high_dim (binary): Pre-computed 256-channel fp16 latent tensor bytes.
  • mask (binary): Ground-truth binary manipulation mask encoded as PNG bytes.
  • channels (int64): Total channel count (256).
  • latent_h (int64): Latent height (32).
  • latent_w (int64): Latent width (32).

High-Speed Training with SID-UNet

sid-train --config configs/experiments/diffusion_diff_minimized/default.yaml \
          --cached-hf-repo KhangTruong/gap-sam-coco-cache

Details

Repository
KhangTruong/gap-sam-coco-cache
Publisher
Truong
Task category
Image segmentation
Tags
synthetic-image-detection, diffusion-diff, forensic-latents
Size category
Not stated by the source
Languages
Not stated by the source
Revision
e66d7b4c5c48907de471e8e8727bac126b011f97
Last updated
2026-10-10

Files

23 files, 33.5 MB in total.

Data21 files · 33.5 MB
Documentation1 file · 1.3 KB
Repository1 file · 2.5 KB
Every file
FileTypeSizeSHA-256
cache_info.jsonData515 B—
train/train-00000.parquetData1.7 MBb8fe74447a8a
train/train-00001.parquetData1.7 MB817fd202dfbb
train/train-00002.parquetData1.7 MBd812f60d95cd
train/train-00003.parquetData1.7 MBbfef785e4587
train/train-00004.parquetData1.7 MB7f450456f81e
train/train-00005.parquetData1.7 MBa5a4988e9d0a
train/train-00006.parquetData1.7 MBd88c6fcad88a
train/train-00007.parquetData1.7 MB382ef8a0c166
train/train-00008.parquetData1.7 MB2bf390c5ccfb
train/train-00009.parquetData1.7 MB28fe921cb3ad
train/train-00010.parquetData1.6 MBa733a3353cc3
train/train-00011.parquetData1.7 MBadd4a989c5e0
train/train-00012.parquetData1.7 MBd2265d68c40d
train/train-00013.parquetData1.7 MB77bed818943d
train/train-00014.parquetData1.7 MBe4c057cf35e8
train/train-00015.parquetData1.7 MB6e43e3a6eab6
train/train-00016.parquetData1.7 MBb0dd3853b2fd
train/train-00017.parquetData1.7 MB14c9fece8c6e
train/train-00018.parquetData1.7 MB1ada0b865641
train/train-00019.parquetData1.7 MBb8bb2c2631f7
README.mdDocumentation1.3 KB—
.gitattributesRepository2.5 KB—

License and Download

License
apache-2.0
Access
No access gate
Download from Truong

Released by Truong through its official repository on Hugging Face. Read the license.