storycut-backend

StoryCut — AI B-Roll Finder + Automatic Video Editor

A complete, production-style automatic video editing web application. You provide the story (script) and the timing (voice-over); the app finds/selects visuals, trims and synchronizes them, builds subtitles and text animations, and hands you a full professional timeline for final review — then exports a real video file.

Quick start

cd ~/workspace/broll-editor
./start.sh
# open http://127.0.0.1:8099

start.sh creates/uses the local Python venv, installs Flask, and starts the server. All data (projects, uploads, settings, cache) lives in storage/.

The two automatic engines (one project, one timeline)

1. Automatic B-Roll Footage Finder — script + voice-over (+ optional own media) → searches Pexels / Pixabay (free API keys, yours) for every scene.
2. Auto Edit My Footage — uploads your videos/images, scores each file against every scene's visual intent, and extracts the best sub-clip (scene-change aware trimming, never destructive).
3. Use Both — per-scene source strategy: B-roll only, user only, user-first, B-roll-first, or balanced AI selection.

How the automation works (honest description)

• Script understanding — full-script read first: sentences, story map (dates, numbers, named entities, keywords, sections). Later references keep narrative context.
• Voice-over analysis — ffprobe duration + ffmpeg silencedetect speech-energy intervals. Sentences are distributed across detected speech intervals proportionally to word count, preserving real pauses. This is an energy-based alignment, not word-level ASR — the UI says so, and word timestamps are interpolated. (Install Whisper for true word alignment; the pipeline will prefer it if a whisper binary/API is configured — not bundled.)
• Visual checkpoints — narration is inspected roughly every 2–3 words, but cuts are made at natural phrase/sentence boundaries targeting 3.5–12 s shots (documentary pacing, not rapid-fire cuts).
• Visual intent → queries — content keywords → concise search queries (+ synonym alternates), never full sentences.
• Two-stage B-roll — stage 1 searches metadata/thumbnails (cached 24 h); stage 2 downloads full quality only for the chosen candidate.
• Ranking — keyword overlap, resolution, orientation, duration fit, uniqueness (repeat footage penalized). Best candidate wins, not the first result.
• Confidence + review queue — every automatic decision is scored high/medium/low; low-confidence scenes land in a Needs Review queue.
• Progressive — scenes appear on the timeline as they finish; you don't wait for the whole project.

Editor

• 6 tracks (Main Video, B-Roll/Overlay, Text/Graphics, Subtitles, Voice-over, Music/SFX), zoom, snap, thumbnails, real waveform, split / trim / move / duplicate / delete / multi-select / lock, undo-redo.
• Real preview — canvas compositor playing actual clips with cross-dissolves, Ken Burns on images, animated text overlays, styled subtitles, voice-over + auto-ducked music, fullscreen, inspect mode.
• Script sync bar (click a sentence to seek), Find-B-roll-from-phrase, per-clip Replace B-roll (candidate grid), Regenerate visual for one segment (locked segments are never touched).
• Subtitle editor + presets; text overlay editor; transitions; speed, opacity, filters (grayscale/sepia/brightness/contrast/saturation), 9:16 smart-reframe focal point.

Export

• Server render (FFmpeg) — true 720p/1080p/1440p/4K in 16:9 or 9:16, hardware encoder auto-detected and verified (falls back to libx264), xfade dissolves, drawtext overlays, ASS subtitles, sidechain-ducked audio.
• Fast capture — records the live preview in-browser (WebM), exactly what you see and hear.

B-roll sources & API keys

Settings → B-Roll Sources. Pexels and Pixabay need free API keys (links in the app). Keys are stored server-side in storage/settings/ and only ever shown masked. Provider priority, enable/disable, Test Connection, and license/attribution notes per provider. Every B-roll clip keeps provider, asset ID, source URL, creator, license, retrieval date and the query that found it; a credits list can be assembled from the timeline.

Mixkit / Coverr / Videvo / Dareful are listed with honest status notes — no verified public API is fabricated. Manual downloads from them work through "Auto Edit My Footage".

Project layout

backend/          Flask API (analysis, providers, export, storage)
  providers/      modular adapters: base, pexels, pixabay, registry
frontend/         vanilla JS SPA (no build step)
  js/             api, store, wizard, pipeline, timeline, preview,
                  panels, exporter, app
storage/          uploads, projects, cache, exports, settings, metrics

Limitations (documented, not hidden)

• Voice timing is energy-based alignment; true word-level ASR needs Whisper (not bundled — would add a large model download).
• 9:16 reframing uses a focal-point heuristic (adjustable per clip), not ML face detection.
• Server export renders the main video track; the overlay track is a preview-layer (its clips can be moved to main before export).
• B-roll search requires your free Pexels/Pixabay keys; without them the app still fully works with your own footage.
• Long 4K renders are CPU-bound on machines without a working HW encoder.

Testing

./start.sh, then: script+voice only, script+voice+videos, images, 16:9 and 9:16, provider failure (no keys), missing visual (gap slug), replace/regenerate, locked segments, export. Backend pieces are covered by the smoke tests in this build's development log.