video-analyze
Installation
SKILL.md
Video Analysis with Nemotron 3 Nano Omni
Analyze videos, audio, images, and documents using NVIDIA's Nemotron 3 Nano Omni multimodal model.
Inputs the skill accepts
| Input | What happens |
|---|---|
*.mp4 / .mov / .webm |
One Omni call with the full video |
*.mp3 / .wav / .m4a |
One Omni call with audio |
*.png / .jpg / .webp |
One Omni call with the image |
Directory of PNGs (e.g. *.pdf-pages/) |
If ≤ 8 pages, one Omni call. If > 8 pages, batched (8 pages per call) + synthesis — Omni's per-request image cap is 8. |
Directory of MP4 chunks (e.g. *-chunks/) |
Per-chunk Omni call + final synthesis (medium-video path, 2-30 min) |
Long-video bundle dir (e.g. *-longvideo/ containing audio.mp3 + frames/ + manifest.json) |
Transcribe audio + send full transcript with 8 keyframes + user's question in one Omni call (long-video path, > 30 min) |
The chunked directory path expects chunk_001.mp4, chunk_002.mp4, … and a chunks.json manifest produced by the host-side chunk-upload.sh helper. If chunks.json is missing, the skill falls back to probing each chunk's duration to derive timestamps.