Video Nsfw Detection
Video NSFW Detection checks a video for explicit content and returns one label, normal or nsfw, using one of two open ViT image classifiers.
Overview
Video NSFW Detection is a Wiro moderation tool. It samples frames across your video, runs each one through an NSFW image classifier and returns one word for the whole clip: normal or nsfw. There is no threshold to tune and no score to parse, so an upload filter or a review step can act on the answer directly.
You choose between two classifiers. Both are Vision Transformer (ViT) models fine-tuned for NSFW detection and published under Apache-2.0.
- nsfw_image_detection by Falconsai is the default. It starts from google/vit-base-patch16-224-in21k and was fine-tuned on about 80,000 images in two classes, normal and nsfw. Falconsai reports an evaluation accuracy of 0.980.
- vit-base-nsfw-detector by AdamCodd starts from google/vit-base-patch16-384 and was fine-tuned on around 25,000 images, drawings and photos included. Its author reports 0.9654 accuracy and 0.9948 AUC, and describes it as deliberately strict: images that show cleavage or a lot of skin are classed as NSFW.
What you can build
- Pre-publish checks on user-uploaded videos in a social or community app.
- A check in front of a video generation workflow that holds clips labeled nsfw for review before they are shown.
- Audits of an existing video library that flag clips for human review.
- Routing rules that send nsfw clips to an age-gated area instead of the main feed.
Inputs
A video in MP4, WebM or MOV, either uploaded as a file or given as a public URL, and the classifier to use. If you do not pick one, nsfw_image_detection runs.
Outputs
A text file with a single label, normal or nsfw. Whichever classifier you pick, the label uses these two words. The samples on this page show one result of each.
Recommended settings
Switch to vit-base-nsfw-detector when revealing clothing should count as NSFW, or when your videos contain drawings, which were part of its training data.
Limitations
- Two classes only. The result does not say what was found, where in the video it appears or how confident the model is.
- Both classifiers judge still frames, so the label reflects what is visible on screen. Audio is not analyzed.
- AdamCodd reports that vit-base-nsfw-detector does worse on AI-generated images, with 0.86 accuracy on them, because its training set contained none.
- Both classifiers make mistakes in both directions: expect some false positives and some missed clips.
Safety & compliance
Automated labels are one layer of moderation. Keep a person in the loop for appeals and for decisions that remove content or restrict accounts.
API quick start
Run video-nsfw-detection with a single API call.
{
"model": 183168,
"inputVideo": "https://your-cdn.com/input.mp4"
}curl -X POST "https://api.wiro.ai/v1/Run/wiro/video-nsfw-detection" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"model": 183168,
"inputVideo": "https://your-cdn.com/input.mp4"
}
JSONOther models you might like
Explore similar video generation models.

