I indexed 669 GB of my GoPro videos using my M1 Max computer and local ML models

Hacker News Discussion ↗

Indexing hundreds of gigabytes of personal GoPro footage with local machine‑learning models on an M1 Max MacBook shows that AI-powered video search, face recognition, and semantic querying can be done privately without cloud services. Commenters compare this DIY pipeline to built‑in features in tools like DaVinci Resolve, debate hardware trade‑offs between Apple Silicon, NVIDIA GPUs, and cloud instances, and explore use cases ranging from personal media organization to automated highlight reels. The thread also surfaces concerns about speed, practicality for non‑experts, and the broader privacy implications of entrusting family videos to big tech platforms.

Post format and access

  • Several commenters say this should have been a “Show HN” and note difficulty editing the title.
  • One person reports the main site briefly returning Cloudflare errors; an archive link is shared, later the site works again.

Use cases and extensions

  • Many are excited about local, open pipelines for organizing large personal media collections (videos, photos, documents).
  • A recurring question: could the same approach be used to index porn collections; people discuss model safety filters, abliterated models, LoRA finetuning, and how easy it is to bypass content restrictions with multi-turn prompting.
  • Others focus on family content: hopes for automatic “memories” and year-in-review compilations, combining photos, videos, and music.

Technical pipeline and models

  • Core flow (as discussed): extract scenes at ~1 fps, downscale frames (e.g., 720p), run face/object/text detection, transcription (Whisper), and visual description (Qwen2.5-VL variants).
  • Outputs go into a vector DB plus SQL for semantic search, RAG, and querying by text, screenshot, or audio.
  • One user notes Whisper can hallucinate when fed non-speech (e.g., moaning, slapping); another suggests Parakeet-style models that filter non-voice sounds.
  • Some want true video-clip embeddings, not just frame-level, to better capture actions.

Hardware performance and acceleration

  • Discussion compares M1 Max to 11th gen i9 and Snapdragon X Elite: similar CPU scores, but Apple’s unified memory and bandwidth (and local “AI accelerator”) are seen as major advantages for these workloads.
  • RTX GPUs (e.g., 3060, 5090) are expected to be significantly faster than M1 Max for indexing.
  • People suggest pay-as-you-go GPU providers (Runpod, vast.ai) to speed up large jobs while keeping models local-ish.

Existing tools and integration

  • DaVinci Resolve Studio and Adobe Premiere are mentioned as having built-in or cloud-based AI indexing; DaVinci’s AI runs locally but reportedly lacks full face tagging.
  • Third-party tools like Jumper, Immich, and other local video-indexing projects are suggested, some with NLE integrations and APIs.
  • There’s interest in containerized GPU access on Apple Silicon (podman + Mesa, vLLM-metal via Docker).

Skepticism, usability, and alternatives

  • Some question the example highlight reels as underwhelming given the volume of footage, wondering if the tech is mature enough.
  • A contrasting “simple” workflow is proposed: use GoPro’s built-in “HiLight Tag” while recording, then manually cut those marked segments later.
  • Others argue that while manual tagging is simpler, the ML pipeline enables retroactive search, multi-modal queries, and broader use cases beyond highlights.