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

Ilias Haddad is a long-distance cyclist from Morocco who built edit-mind, an open-source, local-first video indexing system to make his 2,207 GoPro cycling videos searchable by content.

The Problem

After years of cycling with a GoPro, Haddad had accumulated 2,207 videos totaling 669 GB and 15+ hours of footage. The files were organized only by filename and date — there was no way to search for specific moments, faces, or scenes without manually scrubbing through everything.

The Pipeline

1. Audio extraction via ffmpeg → Whisper (large-v3) for transcription
2. Frame extraction at 1fps via ffmpeg → YOLOv8 (object detection) + DeepFace (facial recognition against custom dataset) + PaddleOCR/Tesseract (on-screen text) + shot-type classifier (wide/close/selfie)
3. Qwen2.5-VL-7B-Instruct generates scene descriptions (captioning) — optional "Advanced mode"
4. Three parallel embedding types — text, visual, audio — loaded into local vector database (ChromaDB)
5. Chat agent enables natural-language search, can send matched clips to DaVinci Resolve timeline

Performance

• Input: 628 videos processed, 668.68 GB, 15h 13m 18s duration
• Total compute: 67h 40m 42s (0.22× realtime)
• Frames analyzed: 57,537 (at 1fps, downscaled to 720p)

Stage breakdown: | Stage | Time | % | |---|---|---| | Whisper transcription | 25h 13m | 37.3% | | Frame analysis (YOLO + DeepFace + OCR + captioning) | 24h 56m | 36.8% | | Visual embedding | 11h 49m | 17.5% | | Audio embedding | 4h 18m | 6.4% | | Scene structuring | 48m | 1.2% | | Text embedding | 37m | 0.9% |

Whisper + frame analysis consumed 74% of total time. Per-frame processing: ~4.2 seconds.

Key Engineering Decisions

Desktop App Over Docker

Docker couldn't access the M1 Max GPU (MPS backend fails inside containers). Native macOS desktop app required for GPU acceleration.

M1 Max vs. RTX 3060

RTX 3060 (12GB VRAM) was faster for raw compute, but M1 Max's unified memory (32-64GB) could hold all four models simultaneously — something the 12GB card physically cannot do.

Advanced Mode

Enabling Qwen2.5-VL-7B for scene captioning slows indexing but dramatically improves semantic search for complex queries like "bicycle + mountain pass."

The Open-Source Project: Edit Mind

• GitHub: IliasHad/edit-mind — 1.6k stars, MIT license
• Monorepo: pnpm workspaces (TypeScript 89%, Python 9.8%)
• Stack: React/Vite/Electron frontend, Node.js/Express/BullMQ backend, Python AI layer (OpenCV, PyTorch, Whisper, DeepFace)
• DB: ChromaDB (vectors) + PostgreSQL/Prisma (relational)
• Deployment: Docker Compose + native desktop app
• Latest release: v0.22.0 (2026-05-18)

HN Discussion Highlights

• DaVinci Resolve 21 has built-in AI IntelliSearch that processes locally (no face tagging yet)
• Adobe Premiere Pro equivalent processes in the cloud
• Multiple similar projects mentioned (Framedex)
• Docker + Apple GPU possible via podman/runkit/vllm-metal (Haddad planned to add support)
• Qwen2.5-VL can analyze multiple frames (5 frames at 720p) to understand actions like falling

Unresolved Questions

• License ambiguity (GitHub API returns other — non-standard SPDX)
• No quantified RTX 3060 speedup multiplier
• No isolated timing for Qwen2.5-VL-7B advanced mode
• DaVinci Resolve integration mechanism not fully explained
• Incremental indexing behavior unclear (reuse existing embeddings or full re-index?)

Source reconstructed from: HN discussion (item 48528029), GitHub repo IliasHad/edit-mind, and Chinese technical analysis on cnblogs.com/ninghg/p/20526919. Original URL returned HTTP 503 on all fetch attempts (2026-06-15).