Miowtion Studio is a self-hosted web UI and job queue for MiniMax-H3, the open video model that generates video with sound (ambience, voices, music). It runs on a single consumer GPU and turns H3 into a small production tool: write scenes, queue them, and get finished clips or fully assembled episodes.

> This is a tool, not a model. It does not include MiniMax-H3 weights or the Miowtion toolkit. Download those from their official sources and respect their licenses.

Features

- Studio: 3–70 s videos in 16:9, 9:16, 1:1 or 4:3. Longer videos are chained clip by clip from the last frame, so they play like one continuous take.

- Start / end images: begin on your picture, end on your picture, or interpolate between both.

- Multi-character scenes: up to 6 named characters with their own look and voice, who is in each scene, spoken or sung lines per character.

- Presets: save and reuse full setups (style, cast, scenes, music).

- Fixed soundtrack: lay your own song over the finished video, with the clip's ambience ducked while the music plays and brought back for a final spoken line.

- Episodes: storyboard of shots and title cards, generated as separate jobs and auto-assembled with transitions (fade, flash, wipe, dissolve…), per-shot music cues ("start the track at 3:25 here") and an ambience-ducking slider.

- "Split into scenes": describe the whole video once and a local LLM (llama.cpp, any GGUF, vision models can see your start/end images) writes one scene per clip for you. It can also run on another machine on your network (e.g. a Mac) through a small agent.

- Queue that respects your hardware: priority, pause/resume, cancel, batching of short clips into one model load, survives restarts.

- CPU prompt pre-encoding: while the GPU renders one job, the next text prompt is encoded on the CPU with a RAM watchdog, so the GPU never waits.

- Live system meters: CPU, RAM, GPU usage, VRAM and temperature in the header.

- Batch CLI: put many ideas in a Markdown file → scenes are written → jobs are queued. Review the scenes first with --solo-escenas.

Tested setup

- RTX 4090 (24 GB), 64 GB RAM, Ubuntu 22.04

- MiniMax-H3 + Turbo LoRA (8 steps), Veda sparse attention or full attention for max quality

- Typical speed: a 14 s clip in ~10 min; a 70 s chained video in ~1–2.5 h depending on quality

Notes

- The model is offloaded block by block from RAM, so 64 GB RAM is the practical minimum; more RAM means faster runs.

- Write scene descriptions in English; dialogue can be in any language the model supports.

- Built with Flask, SQLite and ffmpeg. No cloud services, everything runs locally.

Quick reminder:

There's a README.MD file inside the miowtion-studio folder with the full guide, but here's the short version:

1. Miowtion Studio is the web UI + queue. It needs the Miowtion toolkit and the MiniMax-H3 weights installed first (they're not included in the zip, get them from their official sources). Make sure a plain Miowtion generation works on its own before installing the studio.

2. Then, on the same machine, unzip and run:

sudo MIOWTION_ROOT=/path/to/Miowtion bash install.sh

3. It installs a systemd service and prints the URL (port 7860). Open it in your browser and you're ready.

To help you faster if anything goes wrong, could you tell me:

- Your OS (it's built for Linux / Ubuntu 22.04, Windows only via WSL2, untested)

- Your GPU and how much VRAM it has

- How much system RAM you have (64 GB is the practical minimum, since the model is streamed from RAM)

- Whether Miowtion + the MiniMax-H3 weights are already installed and a test generation works

- If install.sh already ran: the last lines it printed

Note: the interface is in Spanish for now (English is on the roadmap). Happy to help with anything!

Roadmap (ideas for future versions)

Nothing here has a date. Tell me in the comments what you'd use most and that's what goes first.

- English UI: full interface translation with a language switch (the current UI is in Spanish).

- Easier setup: an installer that checks Miowtion, the H3 weights, ffmpeg and llama.cpp, and tells you exactly what is missing.

- Faster runs on 128 GB RAM machines: keep the video model and the prompt encoder loaded between clips (~25% faster per clip).

- Multi-GPU / multi-machine queue: send jobs to several boxes from one studio.

- More presets and episode templates: anime opening, music video, vertical promo for Reels/Stories, short dialogue scene.

- Scene editor upgrades: regenerate a single scene of a chained video, reorder clips, preview each scene's prompt before queuing.

- Smarter "split into scenes": style-aware prompts, a "rewrite this scene" button, and character consistency checks.

- Audio tools: beat-synced cuts for episodes, per-shot volume, sound-effect layers.

- Export presets: one-click 9:16 / 1:1 versions with blurred background, subtitles burned in from the dialogue lines.