After 115K videos in prod, engineer shares Gemini video-understanding gotchas and hacks
TheMoonMidas · x · 2026-09-07
Engineer karolzdeb shared hard-won lessons from running 115K videos through Gemini in production:
- Default 1fps sampling misses details, especially motion; the team uses 10fps, but costs scale linearly with frame rate.
- Videos are downscaled to 480p, so feeding higher-resolution video yields no quality gain.
- Quality hack: send high-resolution screenshots alongside the video file — effective but expensive fast.
- Scene-boundary detection has a dedicated recommended approach (detailed in the original thread).
TheMoonMidas resurfaced the thread with those caveats.
More from coding & agent
- Dev demos AI models handling 3D objects and animations surprisingly well — chongdashu · 2026-10-23
- Jason Kneen burns 104.56B tokens and $24.7K in 304 days of AI coding on Codex and Claude Code — jasonkneen · 2026-10-23
- Typosquatted Blender MCP Repo Hit 958 Stars in Days and Silently Ran Malware — Promptmethus · 2026-10-23
- Benchmarking 56 fresh dependency bugs: MCP routing lifts Claude Code accuracy 68% to 91% — Available_Side_7134 · 2026-10-23
- Swapping only the eval harness moves results 30%: designing agent evals like a scientist — AI Engineer · 2026-10-23
- Med student builds a whole-library RAG with Cognee; single-book ingestion still freezes his machine — shrike293103 · 2026-10-23