MusubiTunerWanGenerateVideo

Let's be clear about what this node is, because the name overpromises: it doesn't load Wan into ComfyUI's graph and it doesn't call any API. MusubiTunerWanGenerateVideo is a command builder. It takes your settings, turns them into a wan_generate_video.py invocation for kohya-ss/musubi-tuner, and runs that script as a subprocess in musubi-tuner's own virtualenv. The output never enters the graph - it's written to disk, and the node hands you back stdout, stderr, and a return code.

Why you'd reach for it

musubi-tuner is the community's primary tool for training Wan LoRAs - it's how most of the good character-LoRA packs on CivitAI got made, and it's where you end up when you outgrow ComfyUI-native training. But it's pure CLI with no UI. If you train there, you already have a full Wan 2.1 stack (DiT, VAE, T5, CLIP checkpoints plus a venv) sitting on disk, and this node is a way to run that same environment for generation without keeping a second terminal workflow alive.

Honest framing: if you don't already live in musubi-tuner, skip this. Native ComfyUI Wan (or Kijai's wrapper) is easier, and this node deliberately forgoes everything that makes ComfyUI generation nice - no preview in the graph, no upscaling into other nodes, video saved to a folder you have to go dig in. It exists for musubi-tuner people who want to generate from the same weights, LoRAs, and flag vocabulary they train with.

How it works

The node validates your inputs, builds the CLI args, locates your musubi-tuner install, and shells out. Under the hood it's picky in ways that tell you it's a wrapper:

• It needs musubi_tuner_path pointing at the musubi-tuner repo root - the code insists on src/musubi_tuner existing, and on a venv, .venv, or env directory with a Python binary inside. That venv's Python is what actually runs the generation.
• Video sizes are enforced per task. t2v-14B only accepts 720x1280, 1280x720, 480x832, or 832x480; the 1.3B tasks only the two 832/480 combos. Set 1024x576 and it errors before anything runs.
• i2v-* tasks require a clip path; *-FC (Fun Control) tasks require control_path. Empty them and it refuses.

The interesting inputs beyond task, prompt, and the three model paths (dit, vae, t5 - all plain file-path strings, no auto-downloading here):

• image_path + clip for i2v, plus end_image_path for first/last-frame (flf2v).
• lora_weight (comma-separated paths) with lora_multiplier - the reason most people install this.
• The memory knobs: fp8, fp8_scaled, fp8_fast, fp8_t5, and blocks_to_swap (0–39) for offloading transformer blocks to RAM. attn_mode picks torch, sdpa, xformers, sageattn, flash/flash2/flash3.
• output_type (video, images, latent, both, latent_images) and cfg_skip_mode for the CFG-skip tricks.

Its three outputs - stdout, stderr, return_code - are diagnostics, not media. If you're hunting for the clip, the node saves into ComfyUI/custom_nodes/ComfyUI-musubi-tuner/output/.

Installing it

The node itself is trivial - no Python deps of its own, since it reuses musubi-tuner's venv:

cd ComfyUI/custom_nodes
git clone https://github.com/keit0728/ComfyUI-musubi-tuner.git

Then restart ComfyUI. The real installation is the prerequisite the README states flatly: musubi-tuner must already be installed and working for Wan generation before you start. Get that running first; this node is only a front door to it.

Where people get burned

• "musubi-tuner path not found": point musubi_tuner_path at the repo root, not src/musubi_tuner. The node checks for that subdirectory itself.
• "Virtual environment not found": musubi-tuner setup creates a venv; if you skipped it and installed into system Python, the node can't find a Python to execute with.
• The size error: it's not a bug - those are the only resolutions musubi-tuner's script supports for that task. Drop to a listed size.
• "Command execution failed": that's musubi-tuner's script dying with a non-zero exit; the real reason is in stderr, and with show_progress on, the live log also streams into ComfyUI's console.

One more honest note: the author's last commit here was mid-2025, and it's a personal, low-star wrapper rather than a maintained ecosystem project. It works, but treat it as a handy shim for your existing musubi-tuner setup, not as the future of ComfyUI Wan.