Musubi Dataset Config (Wan)

If you're training a Wan 2.1 LoRA inside ComfyUI, this is where every run starts. WanDatasetConfig (display name "Musubi Dataset Config (Wan)") doesn't train anything and doesn't load any model - it looks at your folder of training images and their .txt captions and writes the .toml dataset file that every later node in the pack is going to read. Think of it as the paperwork node. Boring, but if you get it wrong, the whole pipeline fails before a single step of training runs.

The pack is jaimitoes/ComfyUI_Wan2_1_lora_trainer, a GUI wrapper around kohya-ss/musubi-tuner. Musubi is kohya's video-trainer lineage, and its docs (linked in the pack README) are the authority on what the settings here actually do. What this node does is take all of that command-line config-writing and shrink it into one form you can fill out in the graph.

How it works

Clicking it runs generate_config, which builds a dictionary with a general section (resolution, caption extension, batch size, bucket flags) plus a datasets list, then dumps it as TOML into ComfyUI/output/musubi_tuner_configs/ with a timestamp in the filename. The output path is your dataset_toml_path. Musubi's cache scripts and trainer all take a dataset config path as their entry point, so this single file is the handoff between your folder on disk and the training machinery.

The inputs that matter

• dataset1_image_directory - the folder of training images. Each image needs a matching caption file: cat.png + cat.txt. The README is explicit that this isn't video-only; plain images train fine too.
• dataset1_cache_directory - where the latents and text embeddings get cached later. The author's advice: use a fresh, empty folder per LoRA, so caches from different runs don't mix.
• general_resolution_width / general_resolution_height - default 512×512. This is the trap. The tooltip on the node itself says it: override is required to swap the 512 default. Video trains at 832×480 or 1280×720, not 512 square, so bump this before you cache anything.
• general_enable_bucket - on by default, keep it on. Bucketing groups images into aspect-ratio buckets so mixed-aspect datasets don't get squished.
• general_caption_extension - .txt by default. Leave it unless your captions live in a different format.
• dataset1_num_repeats - how many times each sample is seen per epoch. 1 is the sane start.

One genuine quirk before you go cross-eyed reading the tooltips: the author pasted the wrong descriptions onto a few fields (the dataset1_image_directory and dataset1_cache_directory tooltips are swapped, and dataset1_num_repeats inherits the cache-directory text). Trust the names, not the tooltips, on this node.

Outputs

• dataset_toml_path (STRING) → wire into the dataset_config_toml input of WanCacheLatents.
• trigger_out (any) → chain it into the next node's trigger_in so the pipeline runs in order.

Install

ComfyUI Manager → Install Custom Nodes → search ComfyUI_Wan2_1_lora_trainer, or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/jaimitoes/ComfyUI_Wan2_1_lora_trainer

then install requirements and restart:

..\..\..\python_embeded\python -m pip install -r requirements.txt

This pack is built for the Windows portable ComfyUI (Python 3.12, PyTorch 2.7.1) and the README expects you to launch ComfyUI from a custom .bat that loads Visual Studio Build Tools (vcvarsall.bat amd64) first, plus extract the bundled Include_and_libs.zip into your python_embeded. Don't skip those steps - the training node runs in a subprocess and will fail on paths otherwise.

Where people get burned

The 512 default resolution, mostly - it silently caps your dataset if you don't override it. Second: pointing every LoRA at one shared cache folder, which mixes cached latents between runs. Give each project its own cache directory and you'll dodge a confusing chunk of "why is my training garbage" posts.