Hi-Lite

Turn long gameplay, podcasts or sports videos into a highlight reel. No account, no sign-up, and your video files never leave your computer.

Hi-Lite transcribes your video's commentary on-device (faster-whisper), cuts one clip per spoken segment, lets you pick the good ones in a web UI, and exports a stitched highlight reel. There is also an optional Auto-Generate feature that has an LLM score the clips for you, using a backend you choose: a model running on your own machine (fully offline), or an OpenAI-compatible API with your own key. See What leaves your machine.

Install (Windows)

Download the latest installer from Releases. No Python, Docker, or Whisper install needed — everything is bundled. On first use the Whisper model (~500 MB) downloads automatically.

How it works

1. Upload a video — audio is extracted and transcribed locally with word-level timestamps
2. Review — one clip per transcript segment, previewed in the browser
3. Select clips manually — fully offline — or hit Auto-Generate to have an LLM score every segment against your own criteria ("clutch rounds and funny moments only"). Scoring runs on a local model or your own API account; see Auto-Generate setup
4. Export — clips are stitched into a single mp4 you can download

What leaves your machine

What you do	What goes over the network
Upload, transcribe, cut clips, review, select manually, export	Nothing
First transcription after install	One-time Whisper model download from Hugging Face (~500 MB)
Auto-Generate with a local model	Nothing — scoring runs on your own machine
Auto-Generate with an API backend	The transcript text of your video, sent to the provider you configured with your own key

Your video and audio files are never uploaded, under any feature. Transcription happens on your own CPU/GPU via faster-whisper. The in-app AI Settings panel always shows which backend is active and what (if anything) it sends.

Auto-Generate setup

Auto-Generate needs an LLM to score your clips. Two supported ways to provide one — if neither is configured the button is disabled and everything else keeps working:

Local model (nothing leaves your machine). Run Ollama (or any server exposing OpenAI-compatible /v1 routes, e.g. llama.cpp) and pull a small instruct model:

ollama pull qwen2.5:7b-instruct   # current default; configurable in AI Settings

Hi-Lite auto-detects a server at 127.0.0.1:11434 and prefers it over any API.

OpenAI-compatible API (your provider, your account, your terms). Open AI Settings in the app, enter a base URL, model, and API key — works with OpenAI, OpenRouter, Together, a self-hosted gateway, or anything else speaking /v1/chat/completions. The key is stored on your machine in the app's data directory, never in the browser, and is only ever sent to the base URL you configured.

These two backends — a local model and an OpenAI-compatible API — are the ones Hi-Lite ships with built-in support for.

Testing

Other providers can be wired up for development or educational purposes; see "Custom third-party providers" in CONTRIBUTING.md for the walkthrough. You are responsible for complying with the terms of service of anything you integrate manually.

Quick start for developers

# Backend (Flask, :8000)
pip install -r requirements.txt
python3 web_interface.py

# Frontend (Vite, :5173 — proxies API calls to :8000)
cd web && npm install && npm run dev

Requires Python 3.12+, Node 20, and ffmpeg on PATH. A Docker dev environment is also available — see CONTRIBUTING.md for that, the Electron shell, and Windows packaging.

Configuration

Env var	Default	Purpose
WHISPER_MODEL	small	Whisper model size (tiny for quick tests)
HI_LITE_PRE_BUFFER / HI_LITE_POST_BUFFER	0.5 / 0.75	Seconds of padding around each clip
HI_LITE_PORT	8000	Backend port
HI_LITE_DATA_DIR	repo root	Where uploads, clips, and models live
HI_LITE_LLM_BACKEND	auto	Force the Auto-Generate backend: local or api
HI_LITE_LLM_BASE_URL / HI_LITE_LLM_API_KEY / HI_LITE_LLM_MODEL	OpenAI / — / gpt-4o-mini	API backend settings (also editable in-app)
HI_LITE_LLM_LOCAL_URL / HI_LITE_LLM_LOCAL_MODEL	http://127.0.0.1:11434 / qwen2.5:7b-instruct	Local backend settings

Contributing

Issues and PRs welcome — CONTRIBUTING.md has the architecture tour, dev environments, and the Windows-packaging gotchas that have bitten us before.

License

Copyright (C) 2025 seanng-1

Hi-Lite is free software licensed under GPL-3.0-or-later. It comes with ABSOLUTELY NO WARRANTY, to the extent permitted by law.

The Windows installer bundles ffmpeg and other third-party software — see THIRD-PARTY-LICENSES.md for their licenses and for the written offer of corresponding source.