Indexing 669 GB of GoPro Videos with Local ML

Ilias Haddad built an open-source, local-first pipeline to make 628 GoPro cycling videos (669 GB, 15+ hours) semantically searchable — all running on an M1 Max with no cloud dependencies. The resulting project, edit-mind (1.6k GitHub stars), combines Whisper transcription, YOLO object detection, DeepFace facial recognition, OCR, and Qwen2.5-VL scene captioning into a five-stage pipeline backed by ChromaDB vector search and PostgreSQL. Total compute: 67 hours. The result: natural-language search over personal video that sends clips straight to DaVinci Resolve.

Key Quotes

"The Docker version couldn't access the M1 Max GPU to utilize its power."

The MPS (Metal Performance Shaders) backend fails inside containers, forcing CPU fallback. Haddad had to build a native macOS desktop app to get GPU acceleration. This is the local-ML tax: the convenience of containerization breaks at the GPU boundary on Apple Silicon.

Running this project over an NVIDIA GPU like RTX 3060 … I was able to get faster results than running it over my M1 Max.

But the M1 Max has a countervailing advantage: unified memory lets it hold all four models (Whisper ~3GB, YOLOv8x ~250MB, DeepFace ~1GB, Qwen2.5-VL-7B ~14GB) simultaneously — something the 12GB RTX 3060 physically can't do. It's the fewer-GPUs-larger-memory tradeoff vs. dense-compute. The M1 Max wins on capability bandwidth (what you can do in one pass); the RTX wins on throughput (how fast each pass runs).

Whisper transcription: 25h 13m (37.3% of total). Frame analysis: 24h 56m (36.8% of total).

The two most expensive stages consume nearly three-quarters of the pipeline. Everything else — embedding generation, scene structuring, text indexing — is rounding error by comparison. The lesson: if you're building something like this, throw your optimization effort at transcription and frame analysis. Embedding is cheap.

Key Themes

• #tool — edit-mind: open-source local video indexing with multimodal AI analysis
• #pattern — Local-first ML pipeline: the five-stage extract→classify→caption→embed→search architecture generalizes beyond video
• #concept — Multimodal semantic search: text, visual, and audio embeddings as parallel retrieval paths over the same content
• #tool — Qwen2.5-VL-7B-Instruct: vision-language model used as a scene captioner, with multi-frame context (5 frames at 720p) for action understanding
• #pattern — GPU-in-container gap on Apple Silicon: MPS backend failure forces native desktop builds. The containerization promise breaks at the ML boundary on Macs
• #concept — Unified memory as ML superpower: the M1 Max's architectural advantage isn't raw FLOPS, it's the ability to keep all models resident simultaneously

Critical Analysis

The real insight isn't the tech — it's the category. Haddad is solving a problem that millions of people have (mountains of unsearchable personal video) with tools that didn't exist five years ago (local Whisper, local VLMs, local vector DBs). The fact that a single developer on a laptop can build a functional video search engine is genuinely new. This was cloud-only territory until very recently.

The pipeline is a commodity; the integration is the product. None of the individual components (Whisper, YOLO, DeepFace, ChromaDB) are novel. The value is in wiring them together into a coherent local-first experience — and that's exactly what edit-mind does. This is the "dark factory" pattern applied to personal media: the pipeline DOT file is the valuable artifact, not any single model.

The Docker-on-Mac GPU gap is a real constraint on local ML tools. Haddad isn't the first to hit this and won't be the last. Until Apple Silicon GPU access from containers is solved (podman/runkit/vllm-metal is promising but immature), local-first ML tools face an ugly choice: ship a desktop app or accept CPU-only performance. The container-native path is simply not ready.

67 hours of compute for 15 hours of video is simultaneously impressive and unusable. It's impressive that it works at all on consumer hardware. But 0.22× realtime means you can't index as you shoot — this is a batch overnight/weekend job. For the hobbyist with an existing library, that's fine. For anyone hoping to make this part of a daily workflow, the economics don't close yet. Hardware will fix this (M4/M5, better GPU access), but we're not there.

The unresolved question that matters most: incremental indexing. Haddad's blog doesn't clarify whether adding a new video reuses existing embeddings or triggers a full re-index. For a tool positioned as a "video knowledge base," this is the difference between a one-off experiment and a living system. If every new GoPro ride requires re-processing the entire library, the tool breaks at scale no matter how fast the hardware gets.

DaVinci Resolve integration is the killer feature nobody talks about. The ability to search semantically and send clips directly to an NLE timeline transforms this from a curiosity into a tool. Most video indexing projects stop at "here are your search results." edit-mind closes the loop: find → edit. That's the difference between a demo and a workflow.

Related Pages

• Local and Open Source Inference — Hub page for running models on your own hardware
• LocalAI — Open-source local AI stack with composable gRPC architecture
• Datacenter GPU in a Gaming PC — The other end of the local-ML hardware spectrum: £200 V100 in a gaming rig
• Playing with Vision Embeddings — What vision embeddings actually encode; relevant to understanding edit-mind's visual embedding layer
• QMD — Local CLI search with hybrid BM25+vector+LLM re-ranking; same "search your own stuff" philosophy, different domain
• Understand-Anything — Knowledge graphs from codebases; same "index local content with AI" pattern
• docmason — Local knowledge base from office documents; parallel approach for documents instead of video
• graphify — Codebase-to-knowledge-graph; the code-domain analog of video-to-knowledge-graph
• Process Flow — Choreography-as-a-service; relevant if edit-mind's pipeline were decomposed into discrete stages
• Smart Models Dumb Pipes — The pipeline is the dumb pipe; the models are the smart components

Sources: I indexed 669 GB of my GoPro videos using my M1 Max computer and local ML models Last updated: 2026-06-15