VRGDG Krea 2 Musubi Installer

VRGDG Krea 2 Musubi Installer is the "set up the training environment" node for the pack's Krea 2 LoRA story. It installs a native Krea 2-ready Musubi-Tuner into a folder you choose, verifies it actually works, and can download the training model assets (the Krea 2 DiT, the Qwen-Image VAE, and the Qwen3-VL text encoder) into your models folder. In one run you go from "no training stack" to "musubi installed, models in place, paths known."

Why does this need to exist? Because "train a Krea 2 LoRA" is not a one-command affair. Krea 2 (the 12B open-weights model Krea released mid-2026) needs a musubi-tuner build with Krea 2 support - the standard kohya-ss/musubi-tuner plus its krea2_train_network.py, krea2_cache_latents.py, and the lora_krea2 network module. Getting that environment right by hand is exactly the kind of dependency hell that eats an evening. The node automates the install, runs a self-check against the Krea 2 modules, and writes a report file so you can see what happened.

The inputs

• target_root - the parent folder for the musubi install (default A:/MUSUBI). This is a Windows-style path because the author runs Windows; point it at wherever you keep your training stack.
• models_root - where the training models get downloaded (default A:/MUSUBI/models).

The outputs

• install_root - where musubi-tuner landed.
• status - a summary string.
• raw_dit_path, turbo_dit_path, vae_path, text_encoder_path - the resolved paths to the Krea 2 RAW/Turbo DiT, Qwen-Image VAE, and Qwen3-VL text encoder. These wire straight into the train-chunk node's inputs.
• models_root - the models folder, echoed back.
• report_path - the install log, for when you need to know exactly what happened.

Real requirements

Two things to know before you click. Python must be 3.10, 3.11, or 3.12 - the node refuses to install on other versions, because musubi's Krea 2 support depends on that range. And the model downloads are big: the Krea 2 raw DiT alone is a serious multi-GB download, on top of the VAE and the 4B text encoder. Budget bandwidth and disk before you run it. Krea 2 is also worth understanding as a weight, not just a model: it's the open checkpoint from a company that's explicit that the license revenue funds Krea 3, so check the license terms if you plan anything commercial.

Installing it

The pack: Manager → search vrgamedev, or git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git into custom_nodes/, restart, hard-refresh. The pack's heavy pip requirements (transformers, opencv, voxcpm, llama-cpp-python) apply as usual.

Common issues

• "Requires Python 3.10, 3.11, or 3.12." The environment ComfyUI runs in has the wrong Python. This is the most common blocker - set up a 3.11/3.12 venv for the pack.
• Download stalls or fails. Resume and re-run; the report file shows where it stopped. Slow connections should give it time.
• install_root exists but status says verify failed. The Krea 2 module imports failed - check the report. Usually a partial install or a Python mismatch.

The honest take: this is a setup node for a genuinely experimental training path. It automates the annoying 80%, but training Krea 2 LoRAs is still bleeding-edge territory - expect iteration, not first-try success.