← sources

I made this 90-second multi-shot video with Claude Code + ComfyUI. I open-sourced the production system behind it.

source: reddit (reddit:comfyui)published 2026-09-26 13:37ingested 2026-09-27 02:40

retained text

I've been building a tool called **konte** for making multi-shot AI videos, and this 90-second piece is what I've been using to dogfood it.

Generating a single clip is pretty easy these days. What kept annoying me was everything that happens once you have more than a few of them. With 10+ shots and several takes each, I kept losing track of which take was the good one. And if I regenerated the character image, I had to remember which shots were built on the old one.

So I started treating the whole thing more like a software project. In konte, shots are defined in a TypeScript DSL, and generated assets have stable addresses like `video:shot.05.motion`. Rerolls become variants instead of overwriting previous work, accepted takes are explicit, and downstream work becomes stale when something it depends on changes.

The slightly unusual part is that I don't really operate this system myself. Claude Code or Codex does. I tell it what I want to change, and it edits the production files, drives ComfyUI, generates the assets, keeps track of the state, and brings the result back for review. I pick takes and leave comments in a browser review UI, then it goes back and changes only the relevant parts. The creative calls are still mine; the agent handles the production machinery.

Everything was generated locally with open-weight models in ComfyUI: Krea 2 Turbo for reference images, MiniMax H3 for storyboard, video and dialogue, Qwen-Image-Edit 2511 for storyboard edits, Stable Audio 3 Medium for music and SFX.

Right now you need Claude Code or Codex and a reachable ComfyUI instance with ComfyUI-Manager (local or remote). konte can provision the custom nodes and model weights an adapter needs, and you can import your existing ComfyUI workflows as adapters.

It's very much an **early technical preview**. The DSL will change, the review UI is still pretty bare, and this 90-second piece is the longest thing I've made with it so far. But it did make this piece end to end, so I think other people trying to break it would be more useful now than me adding more things in isolation.

Repo (MIT): [https://github.com/shiwano/konte](https://github.com/shiwano/konte)

I also published the actual project that produced the video, including the shot list, every take, which ones I picked, and the review history. The models and ComfyUI workflows used are in there too:

[https://github.com/shiwano/konte-readme-hero-example](https://github.com/shiwano/konte-readme-hero-example)

It's a pre-release snapshot, so it's there to look at rather than build on.

If you're already doing multi-shot work in ComfyUI, I'd be very interested to hear where this falls apart for you.

document fetched 2026-09-27 00:38 · canonical https://reddit.com/r/comfyui/comments/1wqqrej/i_made_this_90second_multishot_video_with_claude