CUTLOBBY clip studio // local render
Video stays local
CutLobby Season 01 // Local Lobby No upload · Solo queue

Cut the hype.
Skip the timestamps.

Drop a full VOD. CutLobby reads audio transients, motion and scene cuts, ranks every candidate window against gaming clip formats, and hands you vertical cuts with hooks — decoded and exported inside this tab.

Target feed
02 — Cut room 9:16 · 1080p export
No source loaded Load a VOD and scan it. Your first cut lands here.
00:00 / 00:00
03 — Overlay pick a moment

Export records the 9:16 canvas in real time (a 24s cut takes ~24s) as WebM · 1080×1920. Hooks are burned in. Edit sheet gives you the timecodes for Premiere or CapCut.

05 — Hype map waiting for source
hype proposed cut selected
Click to scrub · bands are ranked windows
A loudness + motion curve is built for the whole file, then cut into format-sized windows. —
Under the hood

How the cutter thinks

1. Audio, in 20 ms frames. RMS loudness catches reactions and gunfire; positive energy jumps isolate transients; high-band energy catches impacts; mid-band energy is a voice/comms proxy. Each feature is normalised per VOD.

2. Motion, where it matters. Frames are sampled on a coarse seek grid, with denser samples around the loudest audio hotspots. Frame differencing gives motion density, while large pixel changes flag scene cuts.

3. Salience, not just volume. Each second is compared with its own 10–20 second baseline, so a quiet game with one explosion outranks constant loud music. Onset, impact, motion, cuts and the comms proxy are fused into the hype curve; rising-energy builds are scored separately.

4. Cut construction. Peaks become format-sized windows. Boundaries are snapped at 0.25s to quiet audio and low motion, with lead and tail room around the peak. Each drafted cut has 0.1s In/Out nudges; dead air is penalised, and overlap plus peak-distance guards keep the draft diverse.

5. Scoring. Peak salience, opening jump, build, impact, motion, comms, loopability, clean edges, length fit and trend-format fit combine into the 0–98 score. Score tiers: Banger 88+, Heater 78+, Solid 68+.

6. Qwen3.5 director. The local engine still finds and times every candidate window. Qwen3.5 is enabled by default: a compact list — timestamps, signal scores, format and your creator context — goes to the configured model. It picks the final cuts, rescores them and writes the hook/title. The default free Qwen3.5 director needs no key; other providers may. No video or audio is sent. If the key is missing, the provider blocks browser requests, the free slot is cooling down, or the call fails, the local ranking is used.

Trend Radar is a curated playbook. The weights are authored format priors and your own boosts/keywords — not live metrics from any platform. Use it to express taste, not to predict a specific algorithm.

Limits worth knowing:

  • No speech-to-text in this build; the comms signal is a mid-band energy proxy. Overlays are hook text you or the model write, not auto-transcribed dialogue.
  • LLM mode sends candidate metadata, not the VOD. The model sees signal shape, not gameplay frames or dialogue, so its titles and ranking are only as good as those summaries.
  • The API key is stored in this browser's localStorage and sent only to the API base you enter. It is not included in edit sheets or exports; use a restricted key.
  • The default free director uses Qwen3.5-9B on OVHcloud AI Endpoints. It needs no key, is rate-limited (about two requests per minute), and is best-effort; candidate metadata goes to that endpoint. Pollinations' GPT-OSS 20B is available as a fallback free provider. Video and audio never leave the browser. Ollama keeps inference fully local.
  • Long files are sampled, not frame-perfect; a 3-hour VOD gets coarser motion resolution.
  • Audio decode is skipped on very large files or unsupported codecs, and the scan falls back to motion only.
  • Exports are WebM (VP9/Opus). TikTok/Shorts accept WebM uploads; convert to MP4 if your pipeline demands it.
LLM director

Let a model pick the cuts

The local engine finds and times every candidate. With the director on, the top windows go to an OpenAI-compatible model, which picks the final set, rescores them and writes the hooks. The video stays in this browser.

Free Qwen3.5 uses Qwen3.5-9B on OVHcloud AI Endpoints: no key, no cost, roughly two requests per minute, best-effort. Candidate metadata leaves this browser for that request; video and audio never do. Pollinations' GPT-OSS 20B is available as a fallback; choose Ollama for a fully local model.

Keys stay in this browser. Keys are only needed for keyed providers; they are stored in localStorage and sent only to the API base you enter. They are never included in edit sheets or exports. If the provider blocks browser requests (CORS), the scan falls back to local ranking. For Ollama, allow the app origin with OLLAMA_ORIGINS=*.