Corvi 2022 — DMimageDetection (mirror)
Unmodified mirror of the two Corvi 2022 detector weights from grip-unina/DMimageDetection, hosted here for reliable consumption from a Hugging Face Space.
Grag2021_progan/model_epoch_best.pth # ResNet-50 stride1, trained on ProGAN outputs
Grag2021_latent/model_epoch_best.pth # ResNet-50 stride1, trained on latent-diffusion outputs
Each file is ~282 MB. Architecture: ResNet-50 with stride0=1,
gap_size=1, num_classes=1. Output is a per-image logit; positive
= fake. ImageNet normalization, no resize / crop required.
Reference
R. Corvi, D. Cozzolino, G. Zingarini, G. Poggi, K. Nagano, L. Verdoliva. On the detection of synthetic images generated by diffusion models. ICASSP 2023. https://doi.org/10.1109/ICASSP49357.2023.10095167
License
Apache 2.0 — same as upstream. Weights are byte-identical to the ones
distributed at the upstream Google Drive link
(https://drive.google.com/file/d/1sAoAuOGCWS4dAMBhDkRHgBf4SgBgvkVf).
Sanity numbers
On 5 random uniform PIL noise images of varied sizes (a basic end-to-end-wiring check, not a benchmark):
| Model | Mean logit | Std |
|---|---|---|
| Grag2021_progan | -4.73 | 0.21 |
| Grag2021_latent | -4.84 | 0.26 |
Negative means "real / not synthetic", which is the expected response to structureless noise.
Cross-architecture observations (GPT-Image-2, 2026-04)
We measured these weights on GPT-Image-2 (autoregressive) outputs in a diagnostic study; full numbers and methodology are in the companion Space:
| Detector | Overall AUC | TPR @ logit=0 | Weak class |
|---|---|---|---|
| Grag2021_latent | 0.968 | 0.790 | doc (AUC 0.886) |
| Grag2021_progan | 0.773 | 0.097 | doc (AUC 0.560 — near random) |
FPR @ logit=0 ≤ 0.4% on real (FFHQ-500 + COCO-500), so the detectors
remain well-calibrated on natural-photo real reference.