soruly/trace.moe
Trace.moe
This project is an anime scene reverse image search engine that matches a screenshot to the exact anime episode and timestamp. It is designed as a self-hosted search service that can be deployed using Docker containers and pre-indexed databases, enabling private operation on local or custom infrastructure.
At its core, the system extracts visual features from frames using a convolutional neural network trained on anime imagery. Query images provided via URL are processed through the same feature extraction pipeline, and an approximate nearest neighbor search matches the query against millions of precomputed frame features for sub-second results. The matched frame is then translated into the corresponding episode, offset, and scene metadata.
The service is delivered as a Docker Compose orchestration that configures multiple containers for the search, API, and database layers. A precomputed database dump is loaded on startup to initialize the search service without requiring full indexing. An API endpoint accepts an image URL as a query parameter and returns source information, including a sample clip.