Loading…
A personal research project that quantitatively analyzes why some YouTube music videos travel and others are buried, running 100% locally. An extreme-group design compares the top 25 and bottom 25 by views-per-subscriber across 50 Vocaloid MVs, replicated on 50 K-pop MVs. Effect sizes are reported as Cliff's δ for cuts per minute (23.9 vs 5.5), median shot length (1.5s vs 6.2s) and whether the thumbnail carries text (9/25 vs 24/25) — and the signals that do not separate the groups (lyric themes, sentiment, song length, upload time) are written up separately as genre convention rather than quietly dropped. Collection runs on yt-dlp, scene boundaries on PySceneDetect, audio on librosa, scene tagging and hard-sub OCR on a local VLM (Ollama qwen2.5-vl), and lyric transcription on demucs plus faster-whisper, with GPU stages serialized so they never overlap inside 16 GB of VRAM. One cycle-one finding — that a premiere release helps — vanished under a within-channel matched reanalysis of 55 pairs (Wilcoxon p=0.46), so it is recorded as channel-level confounding and the project publishes the reversal of its own result. An IMRaD paper draft is complete; the code opens once data reliability is further validated.