musubi

Musubi is the LoRA trainer the Wan community actually uses - it's the standard tool for training Wan LoRAs locally, and it's famously fiddly to stand up (Docker, CUDA versions, a mountain of moving parts). This node is the whole Musubi training recipe, running on Civitai's farm, controlled from a ComfyUI node. You get the Musubi parameter vocabulary without the setup: epochs, resolution, enable_bucket, unet_lr, lr_scheduler, network_dim/network_alpha, optimizer_type.

The framing matters: unlike kohya (SD/SDXL-centric), Musubi is the Wan-specific trainer. So this node is the answer to "I want a Wan LoRA" - for a character, a style, a motion tendency - without fighting a local Musubi install or renting a GPU with enough VRAM (Wan training is heavy; 5B eats a lot, 14B eats everything).

How it works

Hosted training, same as the pack's other training nodes: model (required, CIVITAI_AIR from a Civitai Model Selector - wire the Wan checkpoint you're training on), training_data (required, a URL), training_data_images_count (required, cost input). Then the Musubi surface, which is a trimmed kohya-style set:

• epochs - default 5, up to 20. One pass through the dataset.
• resolution - default 512; training images get scaled down to this.
• enable_bucket - default true; buckets images by size. Leave it on unless all your images are identical dimensions.
• unet_lr - 5e-05 default.
• network_dim / network_alpha - the rank/alpha pair; dim 1 is the default here which is very low capacity - for Wan, expect to raise dim into the 8–32 range and match alpha to it.
• lr_scheduler, lr_scheduler_num_cycles, optimizer_type, target_steps - the rest.

Outputs: moderation_status, epochs, sample_images_prompts, sample_input_images, stored_as_assets, eta, workflow_id, raw_json.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/civitai/civitai-comfy-nodes.git
pip install -r civitai-comfy-nodes/requirements.txt

Or ComfyUI Manager → "Civitai Comfy Nodes" → Install → restart. Authenticate (CIVITAI_API_TOKEN or sidebar); training bills in Buzz, and Wan training is not cheap.

Gotchas

The default network_dim of 1 is the thing that will bite you - that's effectively a toy-rank placeholder, and a Wan LoRA trained at dim 1 will look like nothing happened. Plan to raise it (and set network_alpha sensibly relative to it; the KB's rank/alpha guidance applies unchanged: alpha/dim ≈ how hard the rank applies). Also, Wan training wants a lot of data and a lot of steps, so eta will read like a small novel - that's normal, not a hang. And the usual cloud-training rules hold: dataset on a reachable URL, count your images honestly (it drives cost), and this being early preview means the recipe surface can shift. If you want a Wan LoRA and the Dockerfile was the blocker, this node is the point-and-click version - just remember it's Musubi underneath, with all of Musubi's data appetite.