# I built an open-source local film studio for 8 GB cards (LTX 2.5 and MiniMax H3), with long takes joined in latent space — retained record

> Retained reddit record from reddit:StableDiffusion.

Canonical: https://highsignal.app/data/records/5f233428ee6153a2

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


source: [reddit](https://highsignal.app/data/reddit) (reddit:StableDiffusion)  published 2026-10-25 20:34 ingested 2026-10-26 02:38

[original source ↗](https://www.reddit.com/r/StableDiffusion/comments/1xfd1tv/i_built_an_opensource_local_film_studio_for_8_gb/)

## retained text



```
I have been working on this for a while now, and i just made the repo public. Kinoloom is a web app that runs on your Windows PC and using your cast it takes a idea and turns it in to a movie. You write the story with any LLM, generate the cast and the studio renders the shots, checks them and cuts them together. Everything runs locally on top of ComfyUI. My PC has a RTX 3070 8GB and 64GB ram, and its the only system it has been tested on.  
  
**About Minimax H3** : Its fully supported, but every clip in the showcase is generated by LTX 2.5. H3's license only allows use outside of EU, UK, South Korea and US, and they limit showing the output. I live in the EU, so i dont publish H3 clips. If you're outside those regions, you bring your own H3 weights; the installer sets up the rest.  
  
The video is a 16 second LTX 2.5 take, rendered in four segments with joins at 4,8 and 12 seconds. Normally each piece starts from the last frame and you see a jump at every join, Kinoloom passes the end of each piece into the next one as latents, before anything is turned in to pixels and decodes the whole take in one go at the end. This 16 second clip took about 28 minutes on my 3070.  
  
**The rest of it**  
  
\- Qwen image 2.1 makes the cast and the first frame of each shot, using the cast as references. a still can take about 38-78 seconds on my hardware once the model is warm.  
\- Besides LTX 2-5 and H3 there is also Wan 2.2 ( legacy, kept for old projects ) and five image models.  
  
**What you need**

  
\- Windows ( I have only tested 11), and Nvidia card with 8GB or more, 64GB RAM, because part of each big model sits in RAM.  
\- Disk space for the models you pick. You dont have to install everything, you can pick which weights you want to install  
\- INSTALL.bat sets up everything except the model weights.  
\- A LTX 2.5 5 second clip takes about 6 minutes once warm.  
  
Its a one person project and it has only run on my PC, so if you try it on other hardware i'd lke to hear how it went and hear your feedback!  
  
[https://github.com/harlixay7/Kinoloom](https://github.com/harlixay7/Kinoloom)  
  
The code is MIT (the exceptions are listed in THIRD\_PARTY\_NOTICES.md), and the models keep their own licences.
```



document fetched 2026-10-26 00:30 · canonical https://reddit.com/r/StableDiffusion/comments/1xfd1tv/i_built_an_opensource_local_film_studio_for_8_gb
