I had a folder of fifty-ish clips from a trip and zero interest in editing them. Quick Cuts is the thing I built so I’d stop staring at the folder.
It transcribes each clip, plans cuts that drop silences and fillers and re-takes, optionally stabilizes with Gyroflow and denoises, then renders the whole thing as one concatenation. Output: one watchable rough cut. Forked from Louise de Sadeleer’s cut-video and rebuilt around the hardware I actually own.
The encoder path was the annoying engineering problem. The upstream assumed NVENC; my main machine is a Strix Halo iGPU. So Quick Cuts probes ffmpeg at startup and picks, in order: AMD AMF (h264_amf + d3d11va), NVIDIA NVENC (h264_nvenc + cuda), then libx264 — the fallback that works everywhere and ruins your battery. Sounded like a half-day refactor; wasn’t. Encoder availability isn’t one boolean: detection has to handle the symbol being linked but the runtime failing to allocate the device. Lots of that.
The part that mattered more was preserving comedic pauses. Killing silences sounds easy — VAD, threshold, snip. But the half-second pauses around jokes sit below the same threshold as the silences, and cutting them makes the result feel like a robot stomped on it. Smarter VAD didn’t fix it; transcription confidence did. A confident-empty stretch is silence. A low-confidence-empty stretch is usually a pause around something the model heard but couldn’t transcribe — which is where the laughs are.
One thing I got wrong first time: I used the small faster-whisper model because it was fast, and the cuts felt off without my knowing why. A larger model fixed it. Cut quality is downstream of transcript quality — shaving a minute on transcribe costs five re-running cuts on bad words.
I built the GUI first, then wrapped the same pipeline as a Claude Code skill on a whim. Now I use the skill more than the GUI. “Drop folder, vibe tight, quality medium, render” is faster to type than to click, and the review step is unnecessary for most jobs.
Whole-clip slow motion landed this week. B-roll mode and proper baked captions are queued.