# Stem Studio, my submission for #OpenSourceHiggsfield — retained record

> Retained reddit record from reddit:comfyui.

Canonical: https://highsignal.app/data/records/5193404932007183

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:comfyui)  published 2026-10-30 19:32 ingested 2026-10-31 02:39

[original source ↗](https://www.reddit.com/r/comfyui/comments/1xjoztu/stem_studio_my_submission_for_opensourcehiggsfield/)

## retained text



```
I built **Stem Studio** for a problem I kept running into with AI video: a scene can come back with music even when the prompt asks for silence. I wanted a way to separate the soundtrack into editable layers, recover useful sounds, and decide what stays in the final video.

Stem Studio runs on your own ComfyUI instance. You can use it as a standalone Windows desktop app or open the node pack directly inside ComfyUI. It supports dialogue, music, and effects separation with multiple models, targeted sound extraction, a waveform timeline for editing and comparing results, and exports of individual stems or a video with a new soundtrack. The processing workflows run through ComfyUI.

A candid note about the video: I cut this short demo together quickly from a much longer walkthrough and managed to hit the competition’s submit button **one minute before the deadline**. I know the edit is a little sloppy. A much more thorough high quality deep dive, including the workflows and design decisions, is coming to my YouTube channel soon.

https://reddit.com/link/1xjoztu/video/3ba06y6uenyh1/player

* **ComfyUI node pack:** [https://github.com/uniqornate/Comfy-Stem-Studio-Node-Pack](https://github.com/uniqornate/Comfy-Stem-Studio-Node-Pack)
* **Windows desktop app:** [https://github.com/uniqornate/Comfy-Stem-Studio](https://github.com/uniqornate/Comfy-Stem-Studio)
* **YouTube:** [https://www.youtube.com/@uniqornate](https://www.youtube.com/@uniqornate)

I’d love feedback from anyone working with AI video audio, especially on which sounds are hardest to recover cleanly.

*Videos generations for demo video:* ***Minimax H3***  
*TTS:* ***Omnivoice*** *(cloned in VoiceStudio desktop app)*

\#OpenSourceHiggsfield
```



document fetched 2026-10-31 00:44 · canonical https://reddit.com/r/comfyui/comments/1xjoztu/stem_studio_my_submission_for_opensourcehiggsfield
