Low-Cost Video Captioning on Hugging Face: Timestamped Scripts Using Marlin-2B
vanstriendaniel · x · 2026-07-30
Hugging Face shared a practical video processing script that uses the Marlin-2B vision-language model to batch generate timestamped captions and scene descriptions for videos.
- How it works: Running via vLLM, the script splits videos into 60s chunks and outputs a parquet dataset containing scene descriptions and second-precise event timeframes.
- Low cost: Processing one hour of video footage costs only about $0.05 on a single A10G GPU.
- Features: Supports resumable processing (skipping completed rows) and temporal grounding mode (using the --find argument to locate when a specific event occurs).
- Execution: Can be run directly via Hub URL in HF Jobs without local code cloning.
More from coding & agent
- Agent Memory Matters: Intelligence ≠ Memory, AI That Remembers Everything — lxfater · 2026-07-30
- Cartoon Studio: Open-Source App Turns Scripts into 2D Animated MP4s — tom_doerr · 2026-07-30
- Building a Fully Automated AI Video Editor Inside Claude Code — Aiden_Tech_Ai · 2026-07-30
- Introducing NISPO: Open-Source IUPAC Naming Tool Built via Agentic Self-Improvement — charlieharris01 · 2026-07-30
- Using Claude Opus 5 for Massive Crowd Rendering in the Browser — willeastcott · 2026-07-30
- PLOS Medicine Perspective: How to Benchmark Medical AI Agents Beyond Final Answers — MihaelaVDS · 2026-07-30