Dataset / WildFake /Fake /README.md
Edward0707's picture
Initial commit
1d2e6f2
|
Raw History Blame Contribute Delete
2.86 kB
# WildFake/Fake
This prefix contains 24,982 images from WildFake with
`label=1` (`fake`).
**These JPEG bytes are not the source bytes.** The source images are lossless
PNG, and shipping them as PNG next to JPEG real images would hand any classifier
the file format as a perfect shortcut. Each image was therefore encoded exactly
once from its PNG master, using the single JPEG profile carried by the
real images these fakes train against -- `Complementary` and `NTIRE/real`, both
of which declare profile `d39db51fdcf2ed8c0f2148e77eb00b5f48d4993e8fa2d5e501a28188e1fd9f44`
with `all_rows_match: true`. That profile was originally measured on
`NTIRE/Fake`, which is why the source file is named for it; the target here is
the real side. Encoding replays the qtable set value for value (reverse-lookup:
IJG quality 95) at 4:2:0 chroma, baseline. The lossless masters were never
JPEG-compressed, so every row carries exactly one JPEG generation.
The source file is a marginal and says so -- it omits component selectors,
subsampling, and progressive mode. Subsampling was confirmed separately with the
NTIRE side; progressive, Huffman optimisation, JFIF density and ICC were left at
the encoder's defaults and are **not** verified against the target. Only qtables
and subsampling change decoded pixels, so the pixel-level alignment is exact and
the unverified fields affect file bytes only. `summary.json` records the split.
## Selection
label=1; per-generator quota; PNG master re-encoded to the single NTIRE fake profile
- `SD-original`: 6,000
- `Midjourney`: 5,988
- `SD-personalized`: 4,000
- `SD-adaptor`: 4,000
- `Imagen`: 3,000
- `GigaGAN`: 1,994
Total: 24,982 images.
## Dimensions
The `NTIRE/*` and `Complementary` prefixes filter on `width > 512 AND height > 512`.
This prefix uses a short-edge bound of `>= 512` instead, so images that are exactly
512 on one side are kept: short edge min 512, median 512, max 2048,
with 17,603 images (70.5%) at exactly 512. Train with native-scale random
cropping rather than resizing -- resampling is itself a signal.
## Load with Datasets
```python
from datasets import load_dataset
dataset = load_dataset(
"TechJam2026-Jamlai-Bench/TechJam2026_Jamlai_Bench",
data_files={"train": "WildFake/Fake/train-*.parquet"},
)
```
The Parquet schema is identical to the `NTIRE/*` prefixes: the same sixteen columns in
the same order, so the prefixes concatenate. Everything specific to this set --
generator, source archive, the JPEG profile actually applied -- lives in
`selection_manifest.csv` instead, which also records every source row, output Parquet
row, image dimensions, and SHA-256 digest. `summary.json` records the selection rule,
per-file checksums, and verification results.
Before making a public redistribution, verify that the source dataset license and
attribution terms permit it.