razzant/ouroboroshub/video-studio

Video Studio v1.2.0

Hollywood-grade photorealistic video generator with Gemini multimodal AV QC.

Key Features

• Effort Levels: low (fast, 1 candidate), regular (2 parallel candidates), max (3 candidates + Gemini AV QC + Director pass)
• Parallel Candidate Generation: Character refs, keyframes, and video clips generated in parallel batches; best candidate selected by VLM
• Gemini 2.5 Pro AV QC: At effort=max, full video+audio analysis via Gemini (visual, audio, AV sync, identity, motion)
• Director QC Pass: At effort=max, cross-scene coherence review by Gemini; flags scenes for regeneration
• Voice/Dialogue: When generate_audio=true, Seedance natively generates voices for dialogue lines
• Progressive Learning: Lessons from VLM rejections persist across jobs and inject into future prompts

Pipeline (effort=max)

1. Scenario — Cinematic screenwriter LLM creates storyboard with lens specs, color temp, camera notes
2. Character Refs — 3 candidates in parallel per character → VLM picks best
3. Location Art — Parallel generation
4. Keyframes — Sequential with inter-scene context chain; 3 candidates → VLM picks best
5. Music — AI-composed soundtrack via Lyria 3 Pro
6. Video — Sequential scenes with continuity chain; Gemini AV QC per scene; Director cross-scene pass
7. Assembly — ffmpeg concat with audio mix

Models Used (via OpenRouter)