18.1K ·GitHub
What is this?
Toonflow is an open-source AI workbench for producing short dramas and motion comics: scriptwriting, storyboarding, character and scene assets, and image and video generation chained into one pipeline on an infinite canvas.
Toonflow is an open-source AI workbench for producing short dramas and motion comics: scriptwriting, storyboarding, character and scene assets, and image and video generation chained into one pipeline on an infinite canvas.Is this a desktop app or something I run on a server?Both. It ships as Electron installers for Windows, Linux, and macOS, and it can also be self-hosted via Docker or on a Node.js cloud server (2GB+ RAM), with the web UI served on port 10588 and data stored in SQLite.
Both. It ships as Electron installers for Windows, Linux, and macOS, and it can also be self-hosted via Docker or on a Node.js cloud server (2GB+ RAM), with the web UI served on port 10588 and data stored in SQLite.How do I get my first project going?Install a Release build and log in with admin / admin123, then configure your text, image, and video model providers in the settings center. Create a project and import your source material for chapter event extraction, let ScriptAgent draft the script, and ProductionAgent assembles storyboards and video on the canvas.
Install a Release build and log in with admin / admin123, then configure your text, image, and video model providers in the settings center. Create a project and import your source material for chapter event extraction, let ScriptAgent draft the script, and ProductionAgent assembles storyboards and video on the canvas.Which AI services do I need to plug in?You need an LLM API endpoint, a Sora or Doubao video API, and a Nano Banana Pro image-generation API. It integrates via the Vercel AI SDK with OpenAI, Anthropic, Google, DeepSeek, Zhipu, MiniMax, Qwen, and xAI — and you can code custom provider logic in TypeScript right in the settings, effective immediately with no restart.
You need an LLM API endpoint, a Sora or Doubao video API, and a Nano Banana Pro image-generation API. It integrates via the Vercel AI SDK with OpenAI, Anthropic, Google, DeepSeek, Zhipu, MiniMax, Qwen, and xAI — and you can code custom provider logic in TypeScript right in the settings, effective immediately with no restart.Anything tricky about installing it?On macOS you must allow the app under Settings - Privacy & Security, or certificate issues will keep it from opening. Docker's online deployment is not finished yet, so container users currently build the image locally from source.
On macOS you must allow the app under Settings - Privacy & Security, or certificate issues will keep it from opening. Docker's online deployment is not finished yet, so container users currently build the image locally from source.What do the AI agents actually do?They cooperate in three layers — decision, execution, and supervision — handling task breakdown, content generation, quality review, and revisions; ScriptAgent drafts the story and script while ProductionAgent builds storyboards and video on the canvas. Both agents' core prompts are Markdown skill files you can edit online.
They cooperate in three layers — decision, execution, and supervision — handling task breakdown, content generation, quality review, and revisions; ScriptAgent drafts the story and script while ProductionAgent builds storyboards and video on the canvas. Both agents' core prompts are Markdown skill files you can edit online.How does it turn a novel into a script?It auto-extracts events from the source chapters into a structured event graph, and script adaptation pulls precise context from that graph — which, per the README, reduces information loss in long texts.
It auto-extracts events from the source chapters into a structured event graph, and script adaptation pulls precise context from that graph — which, per the README, reduces information loss in long texts.Do the agents remember earlier sessions?Yes — cross-session memory combines short-term messages, long-term summaries, and semantic recall, running on local ONNX vector retrieval so multi-round creation stays consistent.
Yes — cross-session memory combines short-term messages, long-term summaries, and semantic recall, running on local ONNX vector retrieval so multi-round creation stays consistent.What does a real production run cost?The authors' demo took about 2 hours and roughly ¥130 in API spend: about ¥120 for full video generation (Seedance 2.0), ¥10 for the LLM (Claude Opus 4.6), and under ¥1 for images (GPT Image 2), yielding a 2-minute episode.
The authors' demo took about 2 hours and roughly ¥130 in API spend: about ¥120 for full video generation (Seedance 2.0), ¥10 for the LLM (Claude Opus 4.6), and under ¥1 for images (GPT Image 2), yielding a 2-minute episode.What are similar projects?dramaclaw, huobao-drama, Open-Generative-AI, shuohao-skills, moyin-creator, CodeVideoModel, printfilm, ViMax
dramaclaw, huobao-drama, Open-Generative-AI, shuohao-skills, moyin-creator, CodeVideoModel, printfilm, ViMaxWhich topics is it in?infinite-canvas, storyboarding, short-drama, image-to-video, text-to-video, text-to-image, vector-search, onnx, agent-memory, workflow-automation, video-generation, i18n
infinite-canvas, storyboarding, short-drama, image-to-video, text-to-video, text-to-image, vector-search, onnx, agent-memory, workflow-automation, video-generation, i18n