POWERED BY DESERT ANT LABS
From explanation to a finished cut.
Local transcription → filler review → word alignment → highlight selection → audio cleanup → exact-length export. Each failure is visible. Laya can select text excerpts, not manufacture missing speech.
Pipeline map
Each stage runs in its own sandboxed frame, one at a time. Select any stage to run it alone, with everything it depends on, or with everything after it. Drag cards and panels to rearrange them, or use Alt + arrow keys. Dashed edges are the Laya failover path.
1 · Choose your material
Original files are never modified.
2 · Analyze and handle failures
The full on-device run downloads large original model files. Keep this tab open. Models run one at a time to contain memory use. A failed transcription needs imported words or manual review; it is never invented.
Ready
Transcript and model results
3 · Review and splice voice-over
The worked examples have curated timelines, original-speed narration, and silent diagram holds. Edit the JSON if needed; the exporter validates every boundary and frame. No synthetic narration is supplied.
Recording requires microphone permission. Only silent slots are replaced; longer recordings are rejected.
4 · Export and verify
MP4 export uses your browser’s H.264/AAC encoders at exactly 30 fps. If unsupported, download the timeline and use the documented local FFmpeg renderer. Laya cannot replace a video encoder.
Worked exports
Download improved 2b · 45s · Download improved 1b · 120sFull replacement narration: Silent 2b picture edit · Silent 1b picture edit.
All 18 model demos
Eye and Face need beta access; Who needs its supported Apple SDK. These are not browser inference passes. Title runs through its local Mac engine. Schemer and Toxic use the official model engines adapted for browser execution.
All model checks · Exact media and silence verification · Browser voice-over splice test · Injected failure test with real Laya · Actual full pipeline run · Each model run alone from the graph · Voice-over cue sheet