ArcReel is an open-source, self-hosted workspace for producing short AI videos: AI comic dramas and novel adaptations, narrated and commentary shorts, and advertising and product clips. It takes a novel, a finished script, or product material and turns it into characters, scenes, props, episodes, storyboards, video clips, and a final cut, with characters and settings kept consistent across shots.
Production runs as a reviewable pipeline that can stop and resume: content analysis and project planning, an asset library of characters, scenes, and props, episode and structured script breakdown, storyboard images, video clips and narration tracks, then either FFmpeg composition or export to a Jianying (CapCut China) draft for manual editing. Each stage can be orchestrated by an agent or reviewed, adjusted, and regenerated by you in the workspace. A main agent decides which step a project has reached and hands tasks to focused subagents, which keep long inputs such as a full novel in their own context and pass back only a summary.
Consistency starts from the asset library. Before any shot is generated, a subagent scans the source text for characters, scenes, and props, and character design images become the reference for every later image and video. Individual assets can be regenerated, and earlier versions rolled back.
One backend protocol covers text, image, and video generation plus speech output, with 11 built-in generation providers and more than 100 preset models, including Gemini, OpenAI, Grok, Qwen, and MiniMax, alongside custom endpoints such as self-hosted Ollama and vLLM. Generation is rate-limited per provider with separate image and video channels, failed batch items retry without re-billing successful ones, and costs and actual usage are shown before and after generation.
Features
- Source to final cut: novels, finished scripts, or product material become assets, scripts, storyboards, clips, and a finished video
- Asset library: characters, scenes, and props tracked across the whole production for visual consistency
- Agent orchestration: a main agent dispatches stages to focused subagents that keep long source text out of the main context
- Human review: confirm key stages, regenerate single assets, and roll back to earlier versions
- Three creation types: storytelling, drama series, and advertising
- Model matrix: 11 generation providers and more than 100 preset models for text, image, video, and speech
- Custom endpoints: add self-hosted models through Ollama or vLLM, with automatic model discovery
- Cost tracking: estimated and actual spend shown before and after generation
- Resilient generation: per-provider rate limits, separate image and video channels, resumable tasks, and retries for failed items only
- Jianying export: per-episode draft packages for further editing of subtitles, voice, pacing, and transitions
- Storage options: SQLite by default, with PostgreSQL for production deployments
