Visual quality control in 1 second of training, 65ms per frame, on laptop CPU only
Embarrassed_Bison527 · reddit · 2026-10-08
A developer demoed a fully local visual quality-control tool: point a webcam at parts, show 30 good and 30 bad examples, hit Learn, and it sorts new parts live in 65ms per frame on a laptop CPU — no GPU, no cloud, nothing leaves the machine.
The trick is that no big model is trained: a frozen SigLIP2 image encoder turns frames into descriptions, and only a thin classifier is trained on top. Retraining takes a second whenever the part changes, and text queries like "what colour is this?" work zero-shot. Compared to calling a vision API, cost, latency and privacy all favor this setup. The trained classifier clearly beats plain similarity matching, though it only hits 81% on a deliberately faint defect. The author is candid that accuracy numbers come from a single photo session, so they measure consistency, not robustness to different lighting or cameras. Built on the open-source indecis library.
More from coding & agent
- AI agents made writing code cheap, but reviewing generated code is a different job — alex_verem · 2026-10-08
- CADFather: Autonomous Agentic System Reconstructs Parametric CAD from 3D Meshes — Gennadiy Savrasov · 2026-10-08
- UniSkill: Actor-Aligned Skill Proposals Hit 98.4% Success on ALFWorld — Yifei Lu · 2026-10-08
- Client banned AI from payments data; 15 months later they asked for the AI setup — alex_verem · 2026-10-08
- Open-Source FrameForge Chains Prompts on MiniMax h3 for Long Combat Videos — Super_Range45 · 2026-10-08
- Aseprite MCP Rebuild: Grid Loop + Pixel-Art Knowledge Base Boosts Output Quality — Complex-Log-635 · 2026-10-08