A simple PoC (Proof of Concept) NSFW Detector API Server using NsfwSpy.js model.

nsfw-detector-api is a core dependency of nostr-filter-relay.

Published docker image is available in ghcr.io. Run it instantly:

docker run --init --rm -p 8081:8081 ghcr.io/atrifat/nsfw-detector-api:main

or using docker image with CUDA drivers included (marked with cuda suffix in the tag name) and NVIDIA docker runtime (Container Toolkit) support:

docker run --gpus all --init --rm -p 8081:8081 ghcr.io/atrifat/nsfw-detector-api:main-cuda

You can also clone this repository to run or modify it locally

git clone https://github.com/atrifat/nsfw-detector-api
cd nsfw-detector-api

install its dependencies

npm install

and run it using command

npm run start

or run it using node command directly

node src/index.mjs

If you want to test the API server, you can use GUI tools like Postman or using curl.

Send request by using image URL:

curl --header "Content-Type: application/json" \
  --header "Authorization: Bearer myapitokenchangethislater" \
  --request POST \
  --data '{"url":"https://example.org/image.jpg"}' \
  http://localhost:8081/predict

or send request by using base64 string of the image:

curl --header "Content-Type: application/json" \
  --header "Authorization: Bearer myapitokenchangethislater" \
  --request POST \
  --data '{"data":"base64stringhere"}' \
  http://localhost:8081/predict_data

The output is JSON which consists of four predicted classes (based on NsfwSpy.js) as follows:

{
    "data": {
        "hentai": 0.00016754239914007485,
        "neutral": 0.9930612444877625,
        "pornography": 0.0058021554723382,
        "sexy": 0.0009690204169601202,
        "predictedLabel": "neutral"
    }
}

Run unit and integration tests using the following command:

npm run test

Run end-to-end tests using the following command:

npm run test:e2e

The following environment variables can be used to configure the API server:

• PORT: (Optional) The port the server listens on (default: 8081).
• API_TOKEN: (Optional) The API token for authentication (default: myapitokenchangethislater).
• ENABLE_API_TOKEN: (Optional) Enable or disable API token authentication (default: false).
• ENABLE_CONTENT_TYPE_CHECK: (Optional) Ensure content type check via header request (default: false).
• FFMPEG_PATH: (Optional) Set to other path for ffmpeg installed in system (example: /usr/bin/ffmpeg) otherwise automatically inferred from ffmpeg.path pre-installed dependency
• IMG_DOWNLOAD_PATH: (Optional) Directory to store temporary files (default: /tmp/).
• MAX_VIDEO_SIZE_MB: (Optional) Maximum size of video for classification in MB (default: 100).
• CACHE_DURATION_IN_SECONDS: (Optional) Duration of classification cache in seconds (default: 86400).
• MAX_CACHE_ITEM_NUM: (Optional) Maximum number of items in classification cache (default: 200000).
• REQUEST_TIMEOUT_IN_SECONDS: (Optional) Request timeout for downloading image or checking image header in seconds (default: 60).
• USER_AGENT: (Optional) User agent for downloading files (default: 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/135.0.0.0 Safari/537.36').
• WORKER_POOL_MIN_WORKERS: (Optional) Minimum number of workers in the worker pool for NSFW detection and image processing (default: 2).
• WORKER_POOL_MAX_WORKERS: (Optional) Maximum number of workers in the worker pool. Falls back to WORKER_POOL_MIN_WORKERS if not set, then to 2.
• VIDEO_PROCESSING_CONCURRENCY: (Optional) Set concurrency for video processing tasks (default: 10).
• MUTEX_CACHE_MAX_ITEM_NUM: (Optional) Set maximum number of item in mutex cache (default: 5000).
• MUTEX_CACHE_TTL_IN_SECONDS: (Optional) Set duration of mutex cache in seconds (default: 600).

MIT

• Rif'at Ahdi Ramadhani (atrifat)
• d00M_L0rDz (NsfwSpy)