A 6,800-star open-source AI short-drama tool that gets characters and costumes under control first

There are already plenty of generation models for AI short dramas; what really slows progress is management after generation. Dozens of shots are scattered everywhere, a character changes clothes in one shot, and then their face changes in the next, all relying on people to remember and manually match things. Jellyfish wants to take over this work.

The project comes from an individual developer in Beijing. The repo was only created in March and now has more than 6,800 stars and over 1,100 forks. The address is GitHub repo. Judging from the project structure, it does not bet everything on a single model; the focus is on the pipeline from script to storyboard, asset reuse, and generation tasks.

It first breaks chapter scripts into shots, and along the way extracts characters, scenes, props, costumes, and dialogue, saving them as shared assets for reuse across shots. During shot preparation, you can confirm or ignore candidate assets and dialogue, link existing characters and scenes, correct shot information, and use a unified status to judge whether the shot is ready.

Once ready, you enter the generation workspace, where keyframes, reference images, and video prompts are all managed in one place. Single shots can be pre-checked, and whole batches can be pre-checked before submitting generation. Text, image, and video tasks go into the same task center, with status, elapsed time, cancellation, and jumping back to the project chapter and shot all together. Model providers can come from multiple vendors, default models are set by category, and prompt templates and file media are also grouped under infrastructure.

The tech stack is React + Vite and FastAPI, launched directly with Docker Compose. For deployment, copy deploy/compose/.env.example to .env, then run docker compose --env-file deploy/compose/.env -f deploy/compose/docker-compose.yml up --build. The frontend defaults to 7788, the backend to 8000, and Swagger is at /docs. For ordinary users, this threshold is not really one-click finished video; it is more like a self-deployable short-drama workbench.

In the comments, someone asked whether characters will still match after switching model providers with this shared-asset system. It is a very practical question. The project makes characters, scenes, props, and costumes into cross-shot entities, which can reduce drift, but cross-provider consistency still depends on actual generation results. Others look at the ratio of 6,800 stars to 1,100 forks and conclude that quite a few people are deploying it themselves; after reusing assets across episodes, cost per minute is the real barrier.

These discussions are more interesting than the star count. The entry point for AI short dramas may not be a particular video model, but who can string scripts, assets, shots, and task management into a pipeline that can be run repeatedly. Jellyfish has open-sourced this pipeline; whether it is worth using, you will know after running the project once.