hdcli/copyrightgpt-v1

CopyrightGPT

A fast first-draft video dataset + fingerprint store, built to answer one question: has this video (or something very similar to it) been seen before?

Each ingested video gets a YouTube-style random id and a folder ("bin") of binary data files. A perceptual hash (pHash) is computed for sampled frames, so a new video can be checked against everything already stored to flag likely duplicates / re-uploads.

This is an early, intentionally simple draft — perceptual-hash frame matching only, no audio fingerprinting yet, no ML model. It works, but it's not production-grade copyright detection.

Layout

copyrightgpt-v1/
  idgen.py        # generates YouTube-style 11-char ids
  fingerprint.py   # samples video frames + computes pHash fingerprints
  storage.py        # reads/writes the dataset/ bin+patch binary format
  ingest.py           # add a video to the dataset
  match.py              # compare a query video's fingerprint against the dataset
  cli.py                  # command-line interface (ingest / check / list)
  hf_sync.py                # push this project + dataset/ to this HF repo
  dataset/
    <id>/
      main.bin               # metadata: id, title, source path, duration, patch count
      patch-0.dat              # frame hashes from the first ingest of this id
      patch-1.dat, ...           # more patches appended on later re-ingests

main.bin and patch-N.dat are custom binary formats (via Python's struct) — not pickle, not JSON — length-prefixed strings and packed (timestamp, hash) records.

Usage

pip install -r requirements.txt

python cli.py ingest path/to/video.mp4 --title "My Video"
python cli.py check path/to/new_video.mp4
python cli.py list

check reports a match_ratio (fraction of the query video's sampled frames that have a close pHash match somewhere in a stored video) and flags is_likely_copyrighted: true above a configurable threshold (default 0.5).

Known limitations (first draft)

• Frame matching is unordered nearest-hash — won't reliably catch heavily cropped, rotated, or edited clips, or locate where a short clip appears inside a longer video.
• No audio fingerprinting.
• Single-process; no batching/indexing for large datasets yet.