Overview
OpenCut v3 turns still images into 1080×1920 vertical short videos through a 20-stage AI-assisted pipeline. Domain configuration (education, travel, knowledge_paid, custom) drives prompt packs and storyboards so each output feels native to its niche.
TTS + BGM matching + auto-captions + Remotion rendering, with CLI and SaaS dual-track distribution.
Pain → Solution → Value
Short-form video production is stitched together from five different tools — script, voice, audio–visual alignment, face-aware cuts, BGM. OpenCut v3 collapses that pipeline into a single pass: image assets in, finished 1080×1920 vertical video out, with the right domain config to keep quality consistent across niches.
What changes for the team
A wider view of the same loop. Each card zooms into the corresponding column above.
Short-form production is stitched, not made
- Script writing, voice synthesis, audio–visual alignment, face-aware cuts, BGM — five separate tools.
- Operators stitch outputs manually; the round-trip is slow and error-prone.
- Quality drifts every time the domain changes — education vs. travel vs. knowledge_paid.
Drop images in, ship 1080×1920 out
- Drop in a folder of still images + a domain config — OpenCut v3 produces a finished vertical video in one pass.
- Voiceover, subtitles, BGM already composed — no manual stitching, no per-tool context switching.
- Domain presets lock tone, pacing, and visual style for every niche.
Ship in hours, quality stays native per niche
- Content teams ship short-form videos in hours instead of days.
- Domain presets lock tone, pacing, and visual style so every output feels native.
- Zero manual stitching means zero dropped frames between cuts.
Product architecture
A five-stage pipeline turns image assets and a domain config into a finished 1080×1920 short video. Each stage has a single, well-defined responsibility; arrows mark data flow from raw assets to final render.
Script
An LLM ingests the domain config (education / travel / knowledge_paid / custom) plus the image set, then writes a narration script and a shot-by-shot storyboard tuned to that niche.
LLM · domain configVoice
edge-tts synthesizes the narration into a clean voiceover track with the chosen language and voice profile, ready for downstream alignment.
edge-ttsAlign
PyTorch + Transformers score each still image against the audio track and pick the shot that best matches the current beat, producing a tightly cut visual timeline.
PyTorch · TransformersDetect
OpenCV runs face detection on each frame to find faces, then places camera cuts and framing so the subject stays centered and well-composed.
OpenCVCompose
The final renderer lays BGM underneath, burns auto-generated subtitles over the voiceover, and writes the finished 1080×1920 vertical video to disk — ready to publish.
BGM + subtitlesCapabilities
By the numbers
- 20-stage pipeline
- 4 domains
- 1080×1920 vertical
- TTS + auto-captions
- BGM matching
- CLI + SaaS