Self-hostable AIGC video engine with a node-based infinite canvas — script → storyboard → voice → finished video (source available)
DramaClaw is an open, self-hostable AIGC video engine — you feed it a script and it runs the whole chain: character extraction, beat planning, storyboards + first frames, voice-over, and final assembly. It's built for short drama, but the same pipeline handles ads, product videos, and otome games. Disclosure: I'm one of the people building it.
Two things this community will probably care about most:
1. A node-based infinite canvas. On top of the linear pipeline there's a freeform canvas where you branch, compare, and explore — three storyboards for the same beat, a forked character look — then wire the winner into the next stage. If you live in ComfyUI-style graphs it'll feel familiar, except the nodes are story beats, characters, shots, and scenes instead of samplers. The pipeline gives you a reproducible main line; the canvas gives you room to actually play.
2. Consistency, which is where most AI video falls apart. Character identity is carried forward through locked reference assets in a library, not prompt re-description (which drifts badly after a few shots). And there's a director/world layer that pins scene + camera so a location stays coherent when you cut around it, instead of re-rolling the room every shot.
The rest of why it might be worth a look:
- Model-neutral. Text/image/video/voice all go through one OpenAI-compatible gateway — frontier closed models for quality, open weights for control. No lock-in.
- Self-hosted.
docker compose upand it runs on your own boxes. Your scripts, your models, your servers. - Every step is decomposable — independent async tasks with their own UI, so you can run in order, skip, or resume from a checkpoint on long jobs.
- License: Elastic 2.0 — free to use, modify, and self-host; you just can't resell it as a hosted service.
Real videos our team cut on this exact pipeline (trailer + samples): [trailer link]
Repo: https://github.com/dramaclaw/dramaclaw
Happy to go deep on the canvas, the model gateway, or the consistency approach in the comments — feedback from people actually running this stuff is what I'm after.