I came across this repo while looking for a less chaotic way to manage multi-step AI image and video experiments.
The interesting part isn't just that it supports multiple models. It's the workflow:
- Block a shot with a built-in 3D Director, including characters, poses, camera angles, and simple geometry
- Feed that composition into an image generation node
- Reuse the result as a reference, first frame, or last frame for video generation
- Keep every intermediate result connected on the same infinite canvas
- Save the graph in the browser or export it as JSON
It currently supports image models such as GPT Image, Nano Banana, and Seedream, plus video models including Kling, Seedance, and Veo.
Important caveat: the application itself is MIT-licensed and self-hostable, but the models do not run locally. Generation and media delivery use the paid SJolt API. The API key and project graph are stored in the browser.
GitHub:
https://github.com/sjolt-ai/open-canvas-tv
Live demo:
For people doing image-to-video work: would the built-in 3D blocking replace part of your Blender/ComfyUI workflow, or is the cloud API dependency a deal-breaker?
Not affiliated with the project — just thought the workflow was interesting.