HN Showcase
Back to browse
GitHub Repository

CPU-optimised inference wrapper for Meta's SAM 3 — inference pipeline with chunked video processing, automatic object detection, and IoU-based tracking.

20 starsJupyter Notebook

SAM3-CPU – Run Segment Anything on CPU with memory-aware video chunking

by judlaw·Apr 1, 2026·6 points·1 comment
Visit ProjectView on HN

AI Analysis

●●SolidBig BrainNiche Gem

Memory-aware video chunking with IoU tracking lets SAM 3 run without GPU limits.

Strengths
  • •IoU-based mask remapping maintains object IDs across memory-constrained video chunks.
  • •Zero VRAM flag allows inference on shared servers where GPUs are occupied.
  • •Unified CLI and Python API supports text, point, box, and mask prompts.
Weaknesses
  • •CPU inference remains significantly slower than GPU regardless of memory optimization.
  • •Requires HuggingFace authentication and model access approval before running local code.
Category
AI/ML
Target Audience

Computer vision engineers, developers without GPU access

Similar To

Hugging Face pipelines · Replicate · Standard SAM wrappers

Similar Projects

Finance●●Solid

We Made Nasdaq Parsing Even Faster (and More Reliable)

They stopped pretending chunking at arbitrary byte offsets was fine and instead scan once to build message boundaries, then binary-search for clean split points — that simple change eliminates the OOM-by-design scenario. Couple that with SIMD-aware prefetch tuning (different distances for AVX2 vs AVX-512) and you get practical microarch-aware engineering, not just benchmark stunts; I want this shipped as a library or tool so other firms can stop reinventing the same footguns.

WizardryNiche Gem
sundancegh
204mo ago