Spun off from the hashing discussion in #4.
The file fingerprint planned in #4 (size + hashed head/mid/tail chunks) handles exact identity — dedup and tracking files as they move. This issue is the separate, fuzzier problem: detecting clips that are visually near-duplicates — the same scene shot several times — so culling can surface "keep the best of these N takes."
Not file hashing. Needs its own approach, e.g.:
- Perceptual hashes over the sampled frames we already extract
- Or embedding similarity (vision/face embeddings are already computed per clip)
- Grouping by location + timestamp proximity as a cheap pre-filter
Open questions:
- Similarity threshold / how to cluster takes
- How it surfaces — a query flag (
--near-dupes), a section in the folder summary, or both
- Whether it runs during indexing or as a separate pass over existing sidecars
No work scheduled yet — tracking only.
Spun off from the hashing discussion in #4.
The file fingerprint planned in #4 (size + hashed head/mid/tail chunks) handles exact identity — dedup and tracking files as they move. This issue is the separate, fuzzier problem: detecting clips that are visually near-duplicates — the same scene shot several times — so culling can surface "keep the best of these N takes."
Not file hashing. Needs its own approach, e.g.:
Open questions:
--near-dupes), a section in the folder summary, or bothNo work scheduled yet — tracking only.