CremyX Vision
Under maintenancePricing
Pay per usage
Face detection + tracking (SCRFD ONNX), torch-free active-speaker detection (mouth motion x audio envelope) and auto-reframe: returns normalized 9:16 crop keyframes ready for an editplan.
faces = detect + track faces. reframe = the same, plus normalized crop keyframes for a target aspect ratio (editplan-ready). asd = active-speaker detection (mouth motion x audio envelope, no torch needed): per-track speech scores + speaking segments; pass target_aspect to also get crop keyframes that follow whoever is speaking. diarize = visual diarization: the same ASD pass, returned as speaker turns (spk_) with speaker_count/coverage/degraded — use it when audio diarization returns 0 speakers. capabilities = return supported ops/features without running any work (no source needed).
https:// URL, or kv:/ from an earlier run (e.g. the 480p proxy from cremyx-media-tools — analysing the proxy is much cheaper than the master).
Vision options include sampling/detection controls plus embeddings (default true) and embedding_samples (default 5, max 8). Face embeddings are 512-d normalized ArcFace vectors computed only from representative aligned samples, without a second video decode.
{ signed_upload_url } to PUT the result JSON straight into your own storage.
on_success = keep only result artifacts in the run's key-value store. always = also drop them after the signed upload. off = keep everything for debugging.
