Snap Scene

Find and extract any video moment in a snap — sub-second, lossless 4K clipping powered by on-device AI on UGREEN NAS.

Comment

Inspiration

As drone pilots, action-cam videographers, and family archivists, we store hundreds of gigabytes of raw 4K footage on our NAS. Yet, video remains a frustrating "black box." Finding a single 10-second memorable moment (like a river crossing or a child blowing birthday candles) over SMB/WebDAV network streaming means endless scrubbing, playback stutter, and wasted hours.

Cloud AI tools cannot solve this: uploading a 50GB raw video is impractical, expensive, and risks family privacy. We asked: Why can't our UGREEN AI NAS act as an intelligent, local video editing assistant that finds and slices scenes in the blink of an eye? That inspired SnapScene.

What it does

SnapScene is a semantic video scene finder and zero-reencode clipper built for UGREEN AI NAS. By typing a natural language prompt into the UGOS AI Assistant (e.g., "Find the river crossing from last weekend's trip and clip it"), SnapScene:

  1. Discovers & Inspects: Uses file-agent to scan video containers and inspect stream metadata.
  2. Understands Multimodal Cues: Coordinates photo-agent (on-device vision models for visual landmarks) and knowledge-agent (audio transcription for spoken dialogue) to correlate exact scene boundaries.
  3. Slices in ~40ms (Zero-Reencode): Extracts lossless 4K clips using native FFmpeg stream copying (-c copy) with I-Frame Snapping (no frozen/black frames) and FastStart metadata for instant playback.
  4. 100% Safe & Private: Original footage is never modified. All processing stays strictly on-device.

How we built it

  • Uliya Skill Standard v2.0: Built a strictly compliant orchestration package (SKILL.md and linked references in references/) validated with zero errors via the official UGREEN check-skill.mjs.
  • Zero-Reencode Clipper Engine: Implemented an FFmpeg stream-copy pipeline in Python that slices video at the container level without pixel re-encoding, keeping NAS CPU usage below 3%.
  • Intelligent I-Frame Snapping: Used ffprobe packet inspection to snap start boundaries to the nearest preceding keyframe, eliminating initial decode lag and black frames.
  • FastStart Optimization: Embedded -movflags +faststart to move the moov atom to the beginning of the file, allowing immediate web streaming in UGOS.
  • Lightweight Indexing: Created a periodic thumbnail extraction mechanism generating tiny 320px keyframes (~15KB) indexed in SQLite (snapscene.db).

Challenges we ran into

  • Keyframe Boundary Glitches: Naive stream copying often starts with grey or frozen frames if the slice point isn't aligned to an I-frame. We engineered dynamic packet PTS lookback inspection to cleanly snap to keyframes.
  • Strict Hackathon Packaging Rules: Round 1 strictly prohibits scripts or binaries inside the submission package. We designed a pure Uliya agentic delegation workflow in SKILL.md while keeping the benchmark prototype cleanly separated.
  • NAS Thermal & CPU Constraints: Heavy video transcode would choke a NAS. Zero-reencode ensures even 4K 60fps files are processed in tens of milliseconds with near-zero CPU footprint.

Accomplishments that we're proud of

  • Sub-Second Speed: Achieved a benchmarked slicing speed of ~39ms – 71ms for lossless 4K extractions.
  • 100% Privacy & Zero Cloud Egress: All visual and audio analysis is orchestrated locally on UGOS.
  • Static Check Compliance: Passed official Uliya static validation with 0 errors and 0 warnings.
  • Non-Destructive Guardrails: Enforced mandatory human-in-the-loop review for staging and file protection.

What we learned

  • The power of UGREEN UGOS multi-agent delegation contracts (file-agent, photo-agent, knowledge-agent).
  • Deep container-level video engineering (GOP structures, PTS/DTS timing, and moov atom placement).

What's next for SnapScene

  • Physical NAS Testing: Deploy on the assigned UGREEN AI NAS iDX6011 Pro to benchmark I/O throughput across RAID storage pools.
  • Multimodal Confidence Scoring: Implement weighted ranking combining visual features, dialogue transcripts, and geolocation metadata.
  • Multi-Stream Concurrency: Stress-test concurrent extraction requests on multi-core NAS hardware.

Built With

Share this project:

Updates

Submission history