Show HN: Gemini can now natively embed video, so I built sub-second video search

Original Article ↗ Hacker News Discussion ↗

Native video embeddings in Google’s Gemini model are enabling sub-second semantic search over raw video, demonstrated by an open-source project that lets users query dashcam or security footage with natural language like “car with a bike rack that cut me off at night.” Commenters explore practical applications such as home security, wildlife and social media monitoring, video editing, and ad detection, while noting current cost and performance constraints and the lack of equivalent local, open-weight models. Many also warn that as the technology becomes cheaper and more pervasive, it could supercharge state and corporate surveillance, turning ubiquitous cameras into an easily searchable panopticon.

Use Cases and Applications

  • Dashcam and home security footage highlighted as primary use cases (e.g., quickly finding specific incidents, pets escaping, or falls).
  • Proposed for home monitoring, trail and game cams (“find all deer encounters”), and commercial surveillance (retail, worker, state).
  • Suggested for social media/product monitoring (brand tracking across TikTok/Instagram), porn indexing, ad detection/removal, and automatic alerts as a “virtual security guard.”
  • Video editing ideas: search-and-cut features (“remove all scenes with X”) via EDL export or NLE plugins.

Technical Behavior and Quality

  • Gemini Embedding 2 can embed video directly: no transcription or captions required.
  • Embeddings capture visible text (signs, captions) and audio features (e.g., “someone yelling”), though audio was not fully tested by the project author.
  • Temporal structure is respected; not just per-frame CLIP-style averaging.
  • Video is chunked with a configurable overlap (default 5s) to avoid missing events at boundaries; no formal benchmarks yet.
  • Retrieval quality is good but often requires specific queries; more detailed descriptions yield better matches.
  • Currently no confidence thresholding; system returns the “closest” clip even if no good match exists.

Cost, Scale, and Local Alternatives

  • Gemini pricing: ~1 frame/sec, ~$0.00079/frame, ~$2.84 per hour of indexed footage under default settings.
  • Some debate/misunderstanding in the thread about effective cost per hour; resolved by clarifying Gemini’s internal 1 fps tokenization.
  • Cost currently limits continuous real-time indexing; could be trivial for governments or wealthy orgs at scale.
  • Several participants seek open-weight or local video embedding models; CLIP-based and Qwen VL embedding mentioned, plus Intel/OpenVINO tooling, but none clearly match Gemini’s temporal video embedding out of the box.

Privacy, Surveillance, and Dystopia Concerns

  • Strong concerns about panopticon-like surveillance once it is cheap to index every public and private camera feed.
  • Discussion of law-enforcement and municipal systems (ALPR, Fusus, civilian camera integration, facial recognition vans).
  • Embeddings enable searching for descriptions (“tall man in trench coat”) rather than just faces, raising tracking concerns.
  • Some see this as inevitable tech progression; others argue for regulation, pausing AI, or keeping processing local to mitigate risks.