Japan Attire

The problem this node actually solves is the one you've hit: type "woman in a kimono" into a base model and you get a bathrobe with a belt. The model knows the word but not the garment - the layers, the obi, the way the collar sits. The Japan Attire node writes the description for you, in prose, with the real garment names spelled out and their details attached. It's a prompt generator, not an image node: it hands you a finished STRING, and you wire that string into your positive prompt.

What it is

It's one of the country nodes in ComfyUI Globetrotter Nodes, a pack by manifestations that builds culturally-specific character prompts from JSON data files. Everything lives in the repo's data/ directory - attire by body part, poses, cultural context, prompt templates. Nothing about kimonos is hardcoded in Python; the node class itself is generated at load time by a factory function that reads data/countries.json and builds a node for each country.

How it works

When you hit Queue, the node assembles a natural-language prompt in a fixed priority order: subject description → appearance → emotion → attire descriptions → pose → setting → atmosphere → composition → quality keywords. Two details matter. First, the attire dropdowns are gender-filtered at load time - pick Male and female-only items quietly vanish. Second, every attire item carries a full descriptive sentence, so the output reads like "wearing Kimono: a full-length traditional Japanese robe with wide sleeves and a wrapped front" instead of a bare tag. That's exactly what a prose-preferring model needs.

The inputs that matter

• gender and age drive both the subject line and which attire appears. Defaults are Female / young adult.
• Upper_Body is where the money is: Kimono, Yukata, Haori.
• Waist: Obi, Obiage, Obijime. Head: Kanzashi, Hachimaki, Tsunokakushi. Feet: Tabi, Geta, Zori. Legs: Hakama.
• randomization (off / light / moderate / full) plus seed. The author's own tooltip says it plainly: randomization controls how much is varied, and "use same seed to get consistent outputs." Want the same character twice? Randomization to off, seed fixed.
• lora_trigger puts a character name or trigger word into the subject line - useful if you're stacking a character LoRA. custom_prompt appends anything else.
• experimental_llm_rewrite: leave it off unless you've installed transformers (details below).

One honest caveat about this specific node: Japan is one of the thinner countries in the pack. There's no appearance data, so region only offers none/random/Unspecified, and the setting list is generic (Elegant Garden, Serene Lakeside…). You do get Japan-specific pose options. So treat it as an attire-and-subject generator, not a full scene builder.

Output

One output: prompt (STRING). Connect it to a CLIP Text Encode (Prompt) node's text input as your positive prompt, then on to KSampler as usual.

Install

cd /path/to/ComfyUI/custom_nodes
git clone https://github.com/manifestations/comfyui-globetrotter
cd comfyui-globetrotter
pip install -r requirements.txt   # optional - only needed for LLM rewrite

Restart ComfyUI and you'll see "Globetrotter: Loaded N node classes." in the console. Or skip all that and use ComfyUI Manager → search "Globetrotter". No model files to download for this node - the attire data ships inside the repo.

Troubleshooting

• LLM rewrite does nothing: transformers isn't installed, and the node silently skips rather than erroring. When it does run, it downloads distilgpt2 from Hugging Face on first use (a couple hundred MB) and can produce repetitive output - the code runs a repetition filter to compensate.
• Randomization "full" ignores your seed by design - it deliberately uses system randomness for maximum variety. Pin "off" or "light" for reproducible batches.
• A clothing item you expect is missing: that's the gender filter doing its job. It also skips mismatched combinations at generation time.
• This is a young, small pack with essentially no community write-ups yet. When something breaks, the console is your friend - the JSON loader prints "Could not load …" warnings there.