On Replicate an instrument for automatic detection of NSFW content in videos has been launched

Developers presented video-nsfw-filter — a specialized tool based on the Stable Diffusion architecture, designed for automatic detection of prohibited content in video streams. The system analyzes the video sequence by frame-by-frame processing, assigning each fragment a probability coefficient of the presence of NSFW materials. The user receives a structured JSON report at the output, containing summary estimates for each checked frame, as well as optional screenshots with recorded violations. The technical implementation of the model is based on the sample sampling method: the algorithm checks not every frame of the video, but separate fragments, which allows optimizing computational costs when processing long videos. It is important to consider that the current version of the tool focuses exclusively on the visual sequence and ignores the audio track. The tool is available for integration through the Replicate platform, which allows developers to embed automatic moderation into their own video services or content quality control systems. Using the model requires API requests to the platform's cloud infrastructure. Such solutions automate the process of pre-moderation of user content, reducing the load on manual verification services. The use of Stable Diffusion as a classifier provides high accuracy of pattern recognition, adapted to the specifics of video data.