techjam-aigc/wildfake-eval-subset

WildFake Eval Subset

Reference benchmark for the AIGC-detection track, repackaged from WildFake as parquet so it loads in one line. Four configs: the spec-faithful set, plus three that remove artifacts which make the spec-faithful set trivially gameable.

Demonstration purposes only. Do not train on any config here. These exist so you can sanity-check a model and track iterative improvements. They do not contribute to the final score, and the final test set is drawn from the same corpus — training on this leaks.

Start here

Access. This repo is private to the techjam-aigc org. If load_dataset 401s, you are either not a member or not logged in — ask an org admin for an invite, then hf auth login.

Quick start.

from datasets import load_dataset

ds = load_dataset("techjam-aigc/wildfake-eval-subset", "laion_matched", split="validation")
# configs: default | normalized | laion_matched | cross_generator

Three rules.

1. Do not train on any of this. Not the images, not a subset, not "just for augmentation". The final test set comes from the same corpus, so training here leaks and your real score will not survive it.
2. Report which config your number came from. "0.98 AUC" is meaningless without it — the same model can score 1.00 on default and 0.75 on laion_matched.
3. Never report accuracy alone on default. It is 36% real / 64% fake, so predicting "fake" for everything scores 64%. Use AUC or balanced accuracy.

Sanity-check yourself before you believe a good number. Run this first:

def cheat(img):
    return 0 if img.size == (200, 200) else 1

On default that scores AUC 1.000 with no model at all. If your detector is near 1.00 on default but falls to ~0.75 on laion_matched, it learned image size, not detection.

Suggested reporting template.

config	AUC	balanced acc	notes
default			spec compliance only — expect ~1.0, it means little
laion_matched			the number to actually compare on
cross_generator, per source			does it hold past DALL·E 3?

Known issue worth escalating. In default, every real image is 200x200 and no fake image is, so the two classes are perfectly separable without looking at content. This is upstream WildFake preprocessing. If the final test set shares it, the leaderboard will rank resolution detectors rather than AIGC detectors — worth raising with the organizers before tuning against it.

Which config to use

config	rows	contents	resolution	use it for
default	13,841	4,998 COCO val2017 + 8,843 DALL·E 3	as upstream	matching the official spec exactly
normalized	13,841	same images	200x200	the same benchmark without the size giveaway
laion_matched	7,652	3,826 LAION-5B + 3,826 DALL·E 3, both natively >=1024px	512x512	the most meaningful number
cross_generator	5,494	1,500 LAION + DALL·E 3, Midjourney v5, SDXL, GigaGAN	256x256	does it generalize past DALL·E?

from datasets import load_dataset

ds = load_dataset("techjam-aigc/wildfake-eval-subset", "laion_matched", split="validation")
# omit the config name to get `default`

If you only run one, run laion_matched. default is reported for spec compliance, but see below for why its headline number means nothing on its own.

Read this before trusting any score

The classes in default are 100% separable by image size, no model required:

	count	dimensions
COCO val2017 (real)	4,998	every image is exactly 200x200
DALL·E 3 (fake)	8,843	none is 200x200; min side-max 346, max 3056

def cheat(img):
    return 0 if img.size == (200, 200) else 1   # AUC 1.000, learns nothing

This is upstream WildFake preprocessing — its COCO copies are downscaled to 200x200 while the DALL·E 3 images keep native resolution — not an artifact of this repackaging. The same applies to everything in WildFake's Typical trees; the Advanced trees keep native resolution.

How much shortcut survives in each config

Single trivial features, no learning. 1.000 = perfect shortcut, 0.500 = no signal:

feature	default	normalized	laion_matched	cross_generator
image size	1.000	0.500	0.500	0.500
mean luminance	—	0.529	0.734	0.701
recompressed bytes	—	0.602	0.696	0.565
saturation	—	0.582	0.615	0.609
high-freq energy	—	0.580	0.561	0.513
Laplacian variance	—	0.569	0.526	0.640

Two honest observations:

laion_matched has a stronger trivial leak than normalized (0.734 vs 0.602), which is counterintuitive. LAION web photos differ from DALL·E generations in brightness and saturation more than COCO photos do, and the aggressive 200x200 downscale partly washes that out. The difference in kind still matters: 1.000 from pixel dimensions is a pure artifact with no relationship to the task, whereas ~0.73 from brightness is a genuine stylistic difference between web imagery and AI generations — closer to real signal, though a model leaning on it will not survive a distribution shift.

Benchmark your model against this table. If your AUC on default is ~1.00 but drops to ~0.75 on laion_matched, you have learned the shortcut, not the task.

Config details

default — spec-faithful

Exactly the official demo subset: real_coco.csv filtered to /val2017/ (4,998) and all of dalle3.csv, i.e. WildFake DALLE3 with IsAdvanced=1 (8,843). Bytes are untouched — no resizing, re-encoding, or filtering. Classes are unbalanced (36% real / 64% fake), so report AUC or balanced accuracy rather than raw accuracy.

normalized — size shortcut removed, cheaply

The same 13,841 images, each center-cropped to a square and resized to 200x200, re-encoded at JPEG q92. Removes size as a cue but destroys high-frequency detail — the very signal a forensic detector should use. Treat it as a smoke test.

laion_matched — the fair comparison

Native-resolution pairing is not achievable here: LAION clusters at 800x800 and DALL·E 3 at 1024x1024, giving only 66 exact (w,h) matches across 14,000 sampled LAION images. So instead both classes are restricted to natively >=1024px images and put through one identical downscale to 512x512. Both classes therefore start large and receive the same resampling, unlike default where the reals were pre-destroyed and the fakes were not. LAION-5B is also the training distribution for these models, making it the apt real counterpart. Balanced 50/50.

cross_generator — generalization probe

Every fake in the other configs is DALL·E 3, so a strong score there says nothing about other generators. This config holds 1,500 LAION reals against four generators, all through an identical pipeline at 256x256:

source	label	n
laion5b	0	1,500
dalle3	1	1,000
midjourney_v5	1	999
sdxl	1	1,000
gigagan	1	995

DALL·E 3 is included as a same-pipeline reference point, so the drop from DALL·E to the others is measurable within one config:

ds = load_dataset("techjam-aigc/wildfake-eval-subset", "cross_generator", split="validation")
real = ds.filter(lambda x: x["label"] == 0)
for gen in ["dalle3", "midjourney_v5", "sdxl", "gigagan"]:
    fake = ds.filter(lambda x: x["source"] == gen)
    ...  # score real vs fake, compare across generators

256x256 rather than 512 is deliberate: text-to-image GANs output natively smaller than diffusion models (GigaGAN is 512x512), so a 512 target would have left GigaGAN as the only un-resampled source and turned the GAN probe into a resampling detector.

Fields

Identical across all configs.

• image — the image, embedded in the parquet shards.
• label — ClassLabel, 0=real, 1=fake.
• source — origin, e.g. coco_val2017, dalle3_advanced, laion5b, midjourney_v5, sdxl, gigagan.
• orig_path — path within the upstream WildFake archive, for tracing a row back to source.
• id — "{source}/{basename}".

Rows in every config are shuffled with a fixed seed (0), so a truncated or streamed read still sees every class. Streaming works if you don't want the ~3 GB default locally:

ds = load_dataset("techjam-aigc/wildfake-eval-subset", "laion_matched",
                  split="validation", streaming=True)

Provenance

Built by range-reading the upstream ModelScope archives over HTTP — parsing each Zip64 central directory and fetching only the needed byte spans, rather than downloading ~28 GB of zips. Every extracted member was CRC-verified against its central-directory entry.

Sources: Images/Real/coco.zip, Images/Real/laion5b.zip, Images/Diffusion_based/DALLE.zip, Images/Diffusion_based/Midjourney/Advanced/part_1.zip, Images/Diffusion_based/SD/originalSD/Advanced/part_1.zip, Images/GAN_based.zip.

Known upstream defect: 5 of 1,000 sampled GigaGAN PNGs are undecodable. They pass the zip CRC — the bytes match the archive exactly — but fail to parse, so they are dropped (hence 995).

License / attribution

Upstream WildFake terms apply (research / non-commercial). Underlying images retain their own terms: COCO under the COCO terms of use, LAION-5B under its own license, and generated images under their respective providers' terms. Redistributed here for benchmark use within the org.

@article{hong2024wildfake,
  title={WildFake: A Large-scale Challenging Dataset for AI-Generated Images Detection},
  author={Hong, Yan and Zhang, Jianfu},
  journal={arXiv preprint arXiv:2402.11843},
  year={2024}
}