Built for Performance.
Engineered for Precision.
This isn't a cloud wrapper or a basic FFmpeg script. It's a kinetic orchestration engine with chunked rendering, 768-dimensional vector search, and a full cinematic VFX pipeline - running entirely on your metal.
Your Footage, Semantically Understood
Onset Engine uses OpenCLIP to build a deep visual fingerprint of every clip in your library. This isn't keyword tagging - it's semantic understanding. The engine knows the difference between "a character standing calmly" and "an explosive fight scene" without you labeling a single frame.
- ✓ Automatic scene type and mood classification
- ✓ Semantic search - find clips by describing what's in them
- ✓ Few-shot subjects - tag ~5 clips of anything (a person, a vehicle, a shot type) and it finds the rest, with a review queue that sharpens each pass
- ✓ Motion scoring for dynamic content detection
Every Beat Mapped. Every Drop Detected.
The audio engine uses librosa to decompose your music track into beats, onset-strength envelopes, and energy curves. Cuts don't just land on beats - they accelerate during buildups and breathe during interludes.
- ✓ Beat detection and BPM analysis via librosa
- ✓ Continuous energy mapping (0.0–1.0) with smooth pacing interpolation
- ✓ Drop detection - finds the 30-second peak energy window
- ✓ Silent tail trimming - auto-trims fade-outs so videos don't play over silence
- ✓ Voice activity detection for smart audio ducking
- ✓ onset-aware cut snapping (±200ms musical precision)
Memory-Safe Chunked Rendering
Render 10-hour compilations without crashing. The EDL-based pipeline processes clips in chunks of 10–20, flushing RAM between each batch. Final output is stitched via FFmpeg with music and bumpers.
- ✓ Two-phase architecture: Selection (~30s) → Render (chunked)
- ✓ Multithreaded MoviePy rendering (3–4× speedup)
- ✓ Maximum quality: NVENC p6/CQ17 with full prestige VFX
- ✓ Narrative arc modes: Buildup → Drop, Flat, or Follow Music
Cinematic Effects, Zero Manual Work
Nine built-in style presets - each with signature cinematic effects powered by AI depth estimation, optical flow, and foreground extraction. From raw cuts to full Hollywood prestige with halation, anamorphic streaks, luma-matte transitions, and beat-reactive effects. Build your own presets and share them as JSON.
- ✓ Face-aware smart cropping
- ✓ AI cutout pop transitions + luma-matte brightness reveals
- ✓ MiDaS depth parallax and rack focus effects
- ✓ Film halation, anamorphic streaks, echo trails, VHS tracking
- ✓ Optical-flow motion streaks and speed freeze effects
- ✓ Beat-reactive bloom, light leaks, and bass blur pulse
- ✓ Dissolve, luma-matte, whip pan, cutout pop, and zoom transitions
- ✓ Cinematic letterboxing and color grading
Content-Agnostic by Design
Onset Engine is not hard-coded for any content type. The JSON Driver System lets you define what "intensity" means for any genre - anime, gaming, nature, automotive. v3 drivers support subject tags, mood filters, min-rating gates, and contrastive scoring that measures tier specificity, not just raw CLIP similarity.
{
"meta": { "name": "AMV", "version": "3.0" },
"tiers": {
"1_LOW": {
"descriptions": ["character standing calmly"],
"moods": ["serene"]
},
"4_MAX": {
"descriptions": ["massive energy explosion"],
"moods": ["epic"],
"min_rating": 4
}
}
}
One Click. Full Video.
Describe what you want in plain text - "calm landscapes during quiet parts, explosive action on the drops." Hit 🚀 Generate & Render. Onset Engine ingests your footage, analyzes every beat, selects clips by semantic match, and renders a final video with full VFX - no manual timeline work.
- ✓ Text-prompted clip direction - describe low and high energy in natural language
- ✓ Auto-generates a 4-tier semantic driver from your descriptions
- ✓ Glass Engine live timeline - watch clips placed in real-time
- ✓ Vision Matrix - live 3×3 AI analysis grid during ingest
Live Video Mixing. Zero Latency.
Launch a gapless, hardware-decoded MPV session that mixes clips to your music in real-time. Keyboard-driven energy tiers let you ride the mix like a DJ - force transitions on beats, lock styles mid-set, and cue CLIP text queries on the fly.
- ✓ Gapless MPV playback with hardware decoding
- ✓ Energy tier locking - force calm, medium, or intense footage
- ✓ Live CLIP text query - type a phrase, instantly match clips
- ✓ Session recording - save as EDL, render to MP4 later
- ✓ Library-aware: uses driver scoring, ratings, and collections
Describe It. Don't Configure It.
You don't need to write JSON to tell Onset Engine what you want. Two text fields in Studio Mode - "During quiet parts, focus on..." and "On the heavy drops, focus on..." - auto-generate a full 4-tier semantic driver behind the scenes. Power users can still open the Driver Wizard or load a custom JSON.
- ✓ Natural language input - no schema knowledge required
- ✓ Auto-generates 1_LOW → 4_MAX tiers from your descriptions
- ✓ Priority chain: custom driver JSON → text descriptions → motion fallback
- ✓ Vision query - filter footage at ingest time by describing what you want
- ✓ Crate Digger - rate clips while you wait for the render
Fix the Last 5%. Manually.
The AI gets 95% of the edit perfect. The Curator gives you manual tools to fix the rest - swap a bad clip, nudge timing by one frame, or re-roll until the cut feels right. Available for Core and Studio users.
- ✓ Swap - click any clip, search your entire library by text
- ✓ Micro-slip ±1 frame - nudge clip start without moving the cut point
- ✓ Pin & Lock - protect clips from re-rolls
- ✓ OTIO Export - open your timeline in Premiere Pro, DaVinci Resolve (Studio)
Pro Integrations
Onset Engine fits into your existing workflow, not the other way around.
📐 OpenTimelineIO Export (Studio)
Export your AI-generated timeline as a .otio file. Open it directly in DaVinci Resolve, Premiere Pro, or any OTIO-compatible NLE. Full editorial control, zero lock-in.
📱 Triptych Mode (Studio)
Automatically generate 3-panel side-by-side vertical layouts (1920×1080). Perfect for TikTok, Reels, and Shorts - the viral format that keeps audiences watching.
🎧 Audio Ducking
Voice activity detection automatically ducks the background music during dialogue segments. Mix original clip audio alongside your chosen track with precision.
📁 Pointer-Only Ingest (Core+)
Index 4TB of video without duplicating a single file. The Skip Clip Cutting option creates virtual pointers in the database. Zero extra disk space.
📂 Source Collections
Create named subsets of your library. Scope renders, DJ sessions, and autopilot runs to specific collections - perfect for organizing projects by client, genre, or shoot date.
⭐ Unified Tagging
Quality ratings, mood and scene type classification, and few-shot subject propagation. Tag 5 clips, the AI finds the other 800.
🥁 Beat Grid Export
Export any track's detected beats as an Audacity label track, CSV, or timing file. Use Onset's beat detection in Audacity, Resolve, or your own tooling - even if you render elsewhere. Not license-gated.
📈 External Energy Curves
Drive an edit's intensity from a file instead of the music. Supply a simple time,intensity CSV - a hand-drawn build, a sensor trace, a telemetry log - and the edit follows your curve while cuts stay locked to the beat. Blend it against the track or replace it outright.
🗂️ Media Library Import
Already catalogued your footage in a self-hosted media manager? Import those subjects and tags directly over localhost, complete with visual fingerprints - so your existing organization drives editing from day one instead of being retyped.
🕶️ Privacy Mode
Blur every thumbnail for screen sharing, neutralize the window title, and hide the app with a hotkey. Optional PIN at launch. For editing client work in a coffee shop, or anything you'd rather not have on screen during a call.
See It in Action
Download the free demo and watch Onset Engine analyze your footage locally.