route96 NSFW detector (ViT 224)

Fine-tune of Falconsai/nsfw_image_detection (ViT-B/16, 224px) for the route96 media server. Labels: normal, nsfw.

Training data

About 12.7k images from nostr.download, labeled by a vision LLM (Qwen 3.5) and then corrected by hand.

Positives: images where the LLM tags include an explicit term (nudity, genitals, sex acts, bondage and so on), plus images I confirmed by eye.
Negatives: ordinary photos and art, plus screenshots, UI captures, documents and memes as hard negatives.
Images where the LLM only used suggestive tags (lingerie, underwear, bikini) were left out, so the boundary for those is undefined.
Any negative that this model, Falconsai or nostrbuild scored 0.5 or higher was reviewed by hand, and the explicit ones were relabeled.

Training: 5 epochs, AdamW, lr 2e-5, class-balanced sampling, bf16, 224px. The weights from epoch 4 are published.

Evaluation

Scores are at the default threshold of 0.5. Models with more than two classes count the sexual classes as NSFW. AP is average precision, so it does not depend on the threshold.

nostr.download held-out set (1,908 images, 209 NSFW)

Images from the same source as the training data, held out of training.

Model	AP	Precision	Recall	F1	False positives	Misses
route96-nsfw-vit224 (this model)	0.984	0.917	0.957	0.937	18	9
`Falconsai/nsfw_image_detection`	0.921	0.850	0.871	0.861	32	27
`nostrbuild/nsfw-image-detection-large`	0.973	0.735	0.981	0.840	74	4
`carbon225/vit-base-patch16-224-hentai`	0.963	0.773	0.947	0.852	58	11
`umairrkhn/fine-tuned-nsfw-classification`	0.913	0.958	0.756	0.845	7	51
`giacomoarienti/nsfw-classifier`	0.865	0.791	0.871	0.829	48	27
`AdamCodd/vit-base-nsfw-detector`	0.922	0.690	0.957	0.802	90	9
`LukeJacob2023/nsfw-image-detector`	0.815	0.635	0.914	0.749	110	18
`Parveshiiii/SafeScan`	0.855	0.674	0.900	0.770	91	21
`JanadaSroor/vit-nsfw-classifier`	0.775	0.271	0.995	0.426	559	1

Public mix (6,592 images, 3,592 NSFW)

Not seen in training: markandey14/nsfw_detection_xl (sfw and nsfw), jiangchengchengNLP/cashbox-nsfw-image-level (L3 and L4 as NSFW, L1 as normal) and DarkyMan/nsfw-image-classification. These sets define NSFW differently from the training data, so read this table loosely.

Model	AP	Precision	Recall	F1	False positives	Misses
route96-nsfw-vit224 (this model)	0.976	0.938	0.927	0.932	222	261
`Falconsai/nsfw_image_detection`	0.972	0.922	0.905	0.913	275	343
`nostrbuild/nsfw-image-detection-large`	0.979	0.809	0.978	0.886	827	78
`carbon225/vit-base-patch16-224-hentai`	0.968	0.855	0.941	0.896	573	211
`umairrkhn/fine-tuned-nsfw-classification`	0.963	0.976	0.743	0.843	67	924
`giacomoarienti/nsfw-classifier`	0.951	0.888	0.882	0.885	398	423
`AdamCodd/vit-base-nsfw-detector`	0.952	0.834	0.954	0.890	681	165
`LukeJacob2023/nsfw-image-detector`	0.890	0.827	0.923	0.872	694	278
`Parveshiiii/SafeScan`	0.940	0.829	0.946	0.884	700	193
`JanadaSroor/vit-nsfw-classifier`	0.869	0.652	0.995	0.788	1909	19

nostrbuild/nsfw-image-detection-large has the best AP on the public mix and catches more NSFW images, but it needs a threshold near 0.95 to avoid flagging many safe ones. It is a FocalNet with a non-commercial license (CC BY-NC-SA 4.0).

Training and evaluation scripts: https://github.com/v0l/route96/tree/main/training

v0l/route96-nsfw-vit224

route96 NSFW detector (ViT 224)

Training data

Evaluation

nostr.download held-out set (1,908 images, 209 NSFW)

Public mix (6,592 images, 3,592 NSFW)