Amazon Rekognition Content Moderation vs LocalMod

Amazon Rekognition Content Moderation uses machine learning and managed APIs to automate moderation of images and videos. It detects categories including explicit or suggestive content, violence, drugs, tobacco, alcohol, hate symbols, gambling, and disturbing content. Results include labels, confidence scores, and video timestamps, which can support granular rules based on a hierarchical taxonomy. Customers can train a Custom Moderation adapter for image analysis with as few as 20 annotated images. Low-confidence predictions can be sent to Amazon Augmented AI for review by internal teams or third-party reviewers. For stored video, StartContentModeration analyzes content in an Amazon S3 bucket and publishes completion status to an Amazon SNS topic. AWS charges according to images or video duration processed, with no minimum fees or upfront commitments. The Free Tier lasts 12 months from account creation and includes 60 video-analysis minutes per month and two custom-adapter training hours monthly. Images and videos passed to the service may be stored and used to improve it unless customers opt out; stored content is encrypted at rest using AWS KMS.

LocalMod is a free content moderation API for text and images that runs on Linux within a self-hosted setup. The project says it can run offline, keeping user data within the user's infrastructure. Text classifiers check for PII, toxicity, prompt injection, spam, and NSFW text; the PII classifier covers emails, phone numbers, Social Security numbers, and credit card numbers. Image moderation detects explicit or adult images. API endpoints support single and batch text analysis, PII redaction, and image analysis from a URL or uploaded file. The README documents Python use, running an API server, and Docker deployment, and includes an example Discord bot for real-time text and image moderation with auto-delete, timeouts, and logging. Offline operation requires downloading the models once and setting LOCALMOD_OFFLINE=1; the documented download is about 3.5 GB and includes the image model. The README recommends a GPU for production workloads and lists CPU and GPU latency estimates by classifier. LocalMod is MIT licensed.