VRGDG Musubi-Tuner Installer

Musubi-Tuner is the kohya-ss video LoRA trainer, and in 2026 it's the de facto way to train LTX-2.3 LoRAs locally - the kohya line's block-swapping and low-VRAM floor make it the default for video. But installing it, and pulling the right model family, is exactly the kind of fiddly multi-step chore that eats an evening. VRGDG_MusubiTunerInstaller exists to turn that chore into two buttons.

Here's the important thing about how it works: the node itself is passive. When you queue it, it just reports status - the actual work happens from the buttons on the node (install tuner, download models, or both). That's intentional and spelled out in the source: the installer only runs from the buttons so it can't trigger accidentally in the middle of a workflow. You set it up, click, wait, and the outputs get filled with paths.

The inputs that matter

• target_root - where Musubi-Tuner should live. Point this at a folder, not a file. The node creates Musubi-tuner inside it (or VRGDG_Musubi-tuner if that name's taken).
• model_family - what the model-download buttons fetch: LTX 2.3 (default) or Z-Image Turbo.

Outputs: install_root, status, checkpoint_path, assets_root (the model assets folder), and report_path - all STRING paths you can wire into downstream training nodes like VRGDG_LTXAudioVideoLoraTrainChunk.

How to use it

1. Drop the node in, set target_root to where you want the trainer.
2. Click "install musubi-tuner," wait for it to clone.
3. Click "download models" for your chosen family.
4. Queue once if needed to populate outputs, then wire checkpoint_path / assets_root into your training graph.

Be patient - downloading an LTX-2.3 checkpoint is multiple gigabytes, and this node does not magically make that faster. It makes it one button instead of five terminal commands, which is the whole pitch.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

Or search vrgamedev in ComfyUI Manager. It's part of the pack's LTXLoraTrain.py training module, which imports things like transformers and accelerate - the pack's requirements cover it.

Common issues

• "Status hasn't changed / nothing happened": you queued the node instead of clicking its buttons. The installer deliberately does nothing on queue. Click the button.
• target_root resolved to a weird path: the tooltip warns to point at the Musubi root folder, not an arbitrary file path. The node normalizes the path and will create the tuner subfolder itself.
• Model download stalls: it's a big download with no progress bar in some setups; give it time and watch the console.

Honest take: if you're only ever going to use VRGameDevGirl's training nodes, this installer is the easiest path to a working musubi-tuner setup, full stop. If you plan to do serious trainer tinkering you'll still want the underlying repo - but for "train an LTX LoRA from the workflow I downloaded," the two buttons win.