Open-source AI short-drama workbench: edit the storyboard first, then decide whether to generate

AI video workflows are scattered across five tools, and project state easily breaks. You write in one, keep character reference images in another, put storyboards in a spreadsheet, generate on yet another page, and finally switch to another tool for editing. Change one shot, and you have to hunt down the earlier setup again.

NovelVids, Chinese name Maoying Short Drama, is an open-source project introduced by Dan Kornas in a tweet. The author positions it directly: AI video workflows do not need to be scattered across five tools. It is aimed at people making illustrated stories and short dramas, putting story analysis, reusable assets, editable storyboards, and image and video generation into one project. The project uses the MIT license, can be used online, and can also be self-deployed.

What is more worth looking at is how it handles the intermediate layer. Many AI video tools make the generate button very big and project state very small. Change one shot, and character settings, prompts, reference images, and generation parameters are scattered across chat logs. NovelVids does the opposite: first let you edit the storyboard, assets, and parameters, then decide whether to generate.

It solves several specific problems.

Manuscript import supports TXT, Markdown, DOCX, and PDF, and you can check chapter boundaries before creating shots. Character, location, prop, and voice bindings are reused across chapters, reducing the need to re-describe the same person in every chapter. Storyboards are editable: shot descriptions, full prompts, reference materials, and generation parameters can all be changed, and then you can regenerate. An infinite canvas connects assets and video shots, and you can switch between storyboard and workflow views. A creative assistant can modify selected objects on request, show diffs, and undo supported edits.

Usage splits into two paths. Ordinary users can go directly to the Maoying commercial site, log in by scanning a WeChat QR code, choose an image or video model on the homepage, fill in the visual description, add reference materials, check the estimated cost, and generate a clip. They can also import a story from short-drama production and enter the asset and storyboard workspace. The online version provides model selection, reference materials, team collaboration, usage records, and a top-up entry; costs are calculated based on the prices shown on the page and actual usage.

People with a development background can self-deploy. You need Docker and Docker Compose, as well as an available model service. The basic commands are:

git clone https://github.com/Anning01/novelvids.git
cd novelvids
docker compose up -d --build

After startup, open http://localhost:8080. Before creating for the first time, add and enable a text model in the model configuration under settings, check the creative assistant and support for multi-turn tool calls, and enable the assistant in general settings. Then configure the image and video models and prices you want to use. If you only configure a text model, you can still first do story, asset, and storyboard planning.

There is an easy-to-misunderstand point here. Self-deployment does not equal offline inference. Current AI capabilities are called through the APIs you configure, and providers may charge. An open-source license does not make model bills disappear. The project also states clearly that explicit generation requests will call paid models, while only changing prompts will not automatically generate images or videos.

Data is stored locally by default. ./data stores the database and private exports, and ./media stores images, videos, and audio. Back up these two directories before updating. There is no login by default, which is suitable for personal local use. Before public deployment, configure AUTH_ENABLED=true, AUTH_PROVIDER=password, SUPER_ADMIN_USERNAME, and SUPER_ADMIN_PASSWORD through environment variables, and enable HTTPS. When you need Qiling QR code login, WeChat Pay, and remote remastering, configure the corresponding services according to the README.

The tech stack is not complicated. The frontend uses Vue 3, TypeScript, and Vue Flow; the backend uses FastAPI, Tortoise ORM, and Pydantic AI. Creation rules are in prompts/skills/; asset design templates, storyboard pacing rules, and 15 visual styles are provided with the source code. Rules can be changed, while model adaptation and task execution remain independent.

The project provides two finished examples. The commercial recreation example and the comic-drama production example can both be watched on Bilibili. The comic-drama example is labeled with a 3D xianxia visual style and a narration-driven storyboard strategy, generated automatically from story to finished video, with no manual changes to prompts or AI-generated prompts throughout. Actual results still depend on materials, models, and creative adjustments.

Watch the commercial recreation example on Bilibili

Watch the comic-drama production example on Bilibili

You can also enter the remastering workshop from a reference video you have the right to use, analyze shots, and build an editable creative project. This entry point is useful for people breaking down short dramas, but you must confirm material authorization yourself; the project license does not cover third-party novels, music, or voices.

Tutorials are on Bilibili: https://www.bilibili.com/video/BV12JW76dEkj/ . GitHub repository: https://github.com/Anning01/novelvids . Online commercial site: https://shortdramas.xiazq.com . Original tweet: https://x.com/DanKornas/status/2120666895213035553 .

One judgment: the value of this kind of tool is not an extra generate button, but making project state that can be modified, tracked, and reused. Generation models will change, and prices will change. If storyboards and assets can remain, the workflow has continuity. NovelVids has now put this into an open-source project; the remaining questions are whether asset consistency, deployment recovery, and finished-video quality checks can continue to be refined.