The Manual Way
Most videographers organize footage with some combination of:
- Folder naming:
/2024/Japan_Trip/Day_3_Tokyo/— works until you need to find a sunset across all trips - Manual tagging in a DAM: Tools like Eagle, Kyno, or Silverstack require you to watch and tag every clip by hand
- Spreadsheets: Some editors maintain clip databases in Excel. This works surprisingly well but scales terribly
For a library of 10,000 clips, manual tagging at 30 seconds per clip = 83 hours of pure data entry. That's two full work weeks of watching clips and typing tags. Nobody does this. The footage stays unorganized.
The Cloud Way
Google Photos and Frame.io offer AI-powered search, but require uploading everything. For a 5TB library on external drives, that's:
- Upload time: 5TB at 100 Mbps = ~4.6 days of continuous upload
- Storage cost: Google One 5TB ≈ $25/month ongoing
- Privacy: Your footage is stored on and analyzed by Google/Adobe servers
- Lock-in: Tags and search are only available through the cloud platform
The Shortcut: Local CLIP Embeddings
Onset Engine runs OpenCLIP ViT-L/14 on your local GPU during ingest. Every clip gets a 768-dimensional embedding — a mathematical fingerprint of its visual content — stored in a local SQLite database.
- Semantic search: Type "red sports car drifting" and find every matching clip across every drive by cosine similarity — even if the filename is
DJI_0847.MP4 - Auto-classification: Scene type (close-up, wide, aerial, POV), mood (epic, melancholic, serene, tense), and motion score computed automatically
- Few-shot tagging: Tag 5 clips as "Goku" and the engine finds the other 800 automatically via embedding similarity
- Pointer-only mode: Index 5TB without copying a single file. The database stores references, not copies
- Portable database: The SQLite file is a single portable file you can back up, move, or query from scripts
Processing time: ~3 minutes per hour of 4K footage on an RTX 3070. Your entire library, searchable, in an afternoon.
The Power of Accumulated Metadata
Every time you ingest footage, the database grows. After 6 months of ingesting project footage, you have a searchable archive spanning every shoot, every trip, every client. Semantic search works across the entire library — no manual organization required.
Need all your sunset shots for a compilation? Type "golden hour sunset." Need every clip of a specific person? Use few-shot tagging to propagate from 5 labeled examples. Need high-energy action clips for a phonk edit? Filter by motion score > 0.7.
The database persists. The metadata compounds. Your footage library becomes an asset instead of a liability.