video-studio

name	video_studio
description	Hollywood-grade photorealistic video production with Gemini multimodal AV QC, parallel best-of-N candidate generation, effort-based quality control (low/regular/max), voice/dialogue synthesis, and Director cross-scene review
version	1.3.0
type	extension
entry	plugin.py
runtime	python3
permissions	["net","route","widget","ws_handler","tool","read_settings","subprocess","fs"]
env_from_settings	["OPENROUTER_API_KEY"]
when_to_use	User wants to generate a photorealistic cinematic video, film scene, documentary, live-action style video, or realistic narrative with consistent characters, Gemini multimodal quality control, and Hollywood-grade production quality. Separate from anime_studio which generates 2D anime cartoons.
timeout_sec	300
dependencies	["Pillow"]
ui_tab	{"tab_id":"video_studio","title":"Video Studio","icon":"video","render":{"kind":"declarative","schema_version":1,"components":["[Truncated]","[Truncated]","[Truncated]"]}}

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)

Model	Purpose
google/gemini-3.1-flash-image-preview	Character refs, keyframes (default)
openai/gpt-5.4-image-2	Alternative via gpt-image-2
anthropic/claude-sonnet-4.6	Scenario + image VLM verification
google/gemini-2.5-pro	Video AV QC + Director QC (effort=max)
bytedance/seedance-2.0	Video generation (default)
google/lyria-3-pro-preview	Original soundtrack