SARE-evaluation / README.md
juyeon0332's picture
Add README.md
698ed63 verified
|
Raw History Blame Contribute Delete
3.17 kB
metadata
pretty_name: SARE Evaluation Data
task_categories:
  - image-classification
tags:
  - deepfake-detection
  - synthetic-image-detection
  - evaluation

SARE Evaluation Data

Release status: Complete. All archives listed in manifest.json have been uploaded and verified.

Prepared evaluation inputs for SARE on GenImage, CommunityForensics, Forensynths, and Chameleon. Maintained by Jaehong Park and Juyeon Kang.

Only files selected by the SARE test loaders are included. GenImage uses its val split for evaluation. No training split is included. Each sample requires the image in wo_prompt/.../crop/..., the paired image in w_prompt_blip/.../crop/..., and the reconstruction in w_prompt_blip/.../inpainting/....

Download and extract

Archives are independent, uncompressed tar files of approximately 4 GiB each, grouped under archives/<dataset>/. Download all parts for each dataset you intend to evaluate. All parts are required for its full evaluation. The image files are already compressed; tar preserves their bytes.

From the root of your local SARE code repository:

hf download SARE-DB/SARE-evaluation --repo-type dataset \
  --include "archives/Chameleon/*" --include "manifest.json" \
  --include "SHA256SUMS" --local-dir downloads/SARE-evaluation

# Verify the archives you downloaded.
(cd downloads/SARE-evaluation && sha256sum --ignore-missing -c SHA256SUMS)

# Extract each downloaded archive from the SARE repository root.
for archive in downloads/SARE-evaluation/archives/Chameleon/*.tar; do
  tar -xf "$archive"
done

Replace Chameleon with Forensynths, CommunityForensics, or Genimage to download another dataset. During upload, use only datasets listed in manifest.json under completed_datasets. Download the manifest and checksum list again when downloading a newly completed dataset.

Archive members begin with dataset/, so extraction creates:

dataset/
  Chameleon/Chameleon_recon/{wo_prompt,w_prompt_blip}/
  Forensynths/Forensynths_recon/{wo_prompt,w_prompt_blip}/
  CommunityForensics/CommunityForensics_recon/{wo_prompt,w_prompt_blip}/
  Genimage/Genimage_recon/{wo_prompt,w_prompt_blip}/

Keep these directory names and their capitalization. Do not extract inside an existing dataset/ directory, as that adds an extra level. The archives contain regular image files, not the original server's symlinks. Model checkpoints and DINOv3 backbone weights are not included here. Use the SARE code repository's Installation and Inference instructions.

Provenance and terms

This is a prepared evaluation collection derived from the named benchmarks, not a claim of authorship over their original images. Original benchmark and image rights remain with their respective owners. No new blanket license is asserted over these third-party materials. Consult the original benchmark terms before reuse or redistribution.

manifest.json records archive sizes, SHA-256 checksums, and file counts. inventory/<dataset>.jsonl records the relative path, source byte size, and archive assignment of each included file.