marco-willi/clip-cues-artifacts

CLIP-Cues — frozen input snapshot

Cached CLIP features and the checksummed inputs behind "Synthetic Image Detection with CLIP: Understanding and Assessing Predictive Cues" (Willi, Mathys & Graber). This repository holds no images: it is the set of frozen arrays that lets every number and figure in the paper be reproduced on a CPU in minutes, without re-extracting features from 33k photographs.

Code: https://github.com/marco-willi/clip-cues · Image datasets: synthclic · synthbuster-plus · cnnspot-small

Use it

git clone https://github.com/marco-willi/clip-cues.git && cd clip-cues
uv sync --extra all
make finalexp-fetch     # downloads this repo and verifies every file
make finalexp-all       # F1-F7 end to end, a few minutes on CPU

finalexp-fetch checks each file against release_manifest.json before anything loads it, so a truncated download or a swapped file fails loudly instead of silently changing a result.

Contents (289 MB, 20 artifacts)

path	what
embeddings/pooler_l14/	frozen CLIP ViT-L/14-336 pooler_output, 1024-d — SynthCLIC (10815), SynthBuster+ (13999), CNNSpot (8000). The space the canonical detector D_h lives in
embeddings/projected_derived/	the same three corpora in the 768-d shared image–text space, after the single visual projection W_p
embeddings/cue_scores/	per-image × per-cue cosines against the 168 antonym cue directions
projection/	W_p, the 768×1024 CLIP visual projection matrix
vocabularies/	the 168 antonym cue directions (.npz) and their term list (.csv)
checkpoints/	the three published heads used as bridge targets
rankings/	detector-score rankings, so the montage figures rebuild without shipping pixels
reference/	regression anchors and the derived-vs-cached crosscheck

Governance files travel with the data: MANIFEST.md (every artifact with its sha256, shape, declared vector space, origin and consumers), manifest.json (the machine-readable form), and EXCLUDED.md — every plausible-but-wrong neighbour of a snapshot artifact, with the reason it is not used.

Format

Every array is an object-free .npz: it loads with allow_pickle=False, so nothing here executes code on download, and it is portable across Python versions. Embedding frames pack the metadata columns alongside the matrix (embeddings, columns, col__*):

import numpy as np
z = np.load("embeddings/pooler_l14/synthclic.npz", allow_pickle=False)
emb = z["embeddings"]                       # (10815, 1024) float32
split = z["col__split"]                     # train / validation / test
label = z["col__label"]                     # 1 = synthetic

release_manifest.json records two hashes per artifact: sha256 is the file as built — the provenance anchor every run_meta.json in the paper's experiment records cites — and release_sha256 is the distributed file, which is what a download is verified against. They differ exactly where converted is true.

Three things to carry into any table you build from this

1. "mAP" is overloaded. The code's default is pooled AP over all test images; the paper's tables report per-generator mean AP. They are different numbers — state which one you mean.
2. CNNSpot has two evaluation frames. The arrays here are the 4,000-image / 20-generator evaluation frame, a strict subset of the 108,310-image benchmark test split used by the appendix per-generator table.
3. CNNSpot's source column is an evaluation subset, not a provenance. A real photograph filed under progan means the real half of the ProGAN evaluation subset — not that ProGAN produced it.

What is deliberately absent

Text embeddings written before 2026-07-17 are in a double-projected ("W²") space — a fixed bug — and are shape-identical to the correct ones, so they cannot be told apart by inspection. They are excluded here, and EXCLUDED.md names each one. Also absent: raw images (see the dataset repositories above), third-party detector weights, and the B/16–B/32 backbone caches.

Citation

@article{willi2026synthetic,
  title={Synthetic Image Detection with CLIP: Understanding and Assessing Predictive Cues},
  author={Willi, Marco and Mathys, Melanie and Graber, Michael},
  year={2026}
}

The MIT license covers these derived arrays and the accompanying code. The photographs they were computed from remain subject to their own licenses.