DuoUmiWildcards

DuoUmiWild — ComfyUI Wildcard Node

A ComfyUI custom node for building dynamic prompts from .txt wildcard files and structured .yaml libraries. Supports random line selection, range picks, curly-brace randomization, and YAML tag-based selection with prefix/suffix injection. Includes a second node for aspect-ratio management.

---

Installation

Method 1 — Git (recommended)

cd ComfyUI/custom_nodes/
git clone https://github.com/UndefinedUo/ComfyUI_DuoUmiWild.git

Method 2 — Manual Download the ZIP, extract, and place the DuoUmiWild folder inside ComfyUI/custom_nodes/.

Restart ComfyUI. The nodes appear under the DuoUmiWild category.

---

Quick Start

1. Right-click the canvas → Add Node → DuoUmiWild → Wildcard Prompt
2. Connect processed_text → your CLIP Text Encode positive input
3. In the text field, enter a prompt with wildcards:

   masterpiece, best quality, __Prompt/score__, __Prompt/safety__, 1girl,
   __characters/Series Girls/Vocaloid FEM__, __Art/Artist Names__, <[Pose]>
   

4. Hit generate. Change the seed to get different wildcard selections.

Tip: Connect the seed input to your KSampler seed to keep wildcards and image generation in sync. Use a separate seed value to vary wildcards independently.

---

Syntax Reference

.txt Wildcards — __filename__

Picks one random line from wildcards/filename.txt:

__subject__
__Body/FE-Hair__
__characters/Series Girls/Vocaloid FEM__

Files can be referenced by bare name (node searches all subfolders) or by relative path for precision.

Lines starting with # are treated as comments and skipped. Inline comments after # are also stripped.

Range Selection — __N$$filename__ / __N-M$$filename__

Pick multiple lines from a file:

__2$$colors__          # exactly 2
__1-3$$mood__          # between 1 and 3
__-4$$tags__           # up to 4
__2-$$lighting__       # at least 2

Selected lines are joined with commas.

Curly Brace Randomization — {a|b|c}

Pick one option at random:

{happy|sad|neutral} expression
wearing a {red|blue|green} dress

Pick multiple options with range syntax:

{2$$red|blue|green|yellow}     # exactly 2
{1-3$$happy|sad|angry|calm}    # 1 to 3
{0-1$$dark skin}               # 50% chance

YAML Tag Selection — <[Tag]>

Pick a random YAML entry that matches a tag:

<[Pose]>                   # any entry tagged "Pose"
<[Upright Pose][Solo]>     # entries with BOTH tags (AND)
<[Hat|Crown|Headband]>     # entries with ANY tag (OR)
<[Outfit][Fancy|Formal]>   # AND + OR combined

See YAML_GUIDE.md for how to build .yaml files.

Combining Syntax

All three syntax types work together and are expanded recursively (up to 20 passes):

masterpiece, __Prompt/score__, 1girl, {a|b|c|d|e|f|g}-size, <[Pose]>, __Art/Artist Names__

Wildcard files can also contain other wildcards — the node will keep expanding until the prompt is fully resolved.

---

Prompting — Pony / Danbooru-Trained Models

This node is built around models trained on Danbooru-style tags. Follow these conventions for best results.

Tag Order

[quality/meta/year/safety]  [1girl/1boy]  [character]  [series]  [artist]  [general tags]

Within each section, order is arbitrary.

Quality Tags

Human score tags (no underscores):

masterpiece, best quality, good quality, normal quality, low quality, worst quality

Pony aesthetic score tags (underscores only for these):

score_9, score_8, score_7, score_6, score_5, score_4, score_3, score_2, score_1

Both sets can be used together. Use __Prompt/score__ and __Prompt/quality__ to randomize these.

Year / Period Tags

Steer the model toward artwork from a specific era:

year 2025, newest
year 2022, recent
year 2018, mid
year 2013, early
year 2007, old

Use __Prompt/year_period__ for combined pairs, or __Prompt/year__ / __Prompt/period__ separately.

Safety Tags

safe    sensitive    nsfw    explicit

Use __Prompt/safety__ to randomize, or hardcode the one you want.

Artist Tags

Prefix with @ and use lowercase. The @ is required — the effect is weak without it:

@artgerm    @wlop    @kantoku    @nardack

__Art/Artist Names__ handles this automatically. The node converts any Artist:Name token in the output to @name (lowercase), so existing artist files stay compatible.

Prompt Weighting

Weights work but need higher values than SDXL. Example:

(chibi:2)    (detailed background:1.5)

Natural Language

You can mix tags and natural language freely. For natural language, aim for at least 2 sentences and follow standard English capitalization for character and series names:

masterpiece, best quality, @wlop. An anime girl with long silver hair stands at the edge of a cliff at sunset, wearing a white flowing dress.

Full Example

year 2024, newest, masterpiece, best quality, score_8, score_7, safe,
1girl, hatsune miku, vocaloid, @koi_(koitosoitosa),
long hair, twintails, teal hair, detached sleeves, bare shoulders,
smile, looking at viewer, outdoors, cherry blossoms

---

Wildcard Folder Structure

wildcards/
├── characters/
│   └── Series Girls/          300+ anime/game series (one file per series)
├── Body/                      Hair, eyes, expressions, body type, breast size
│   └── breast_sizes.yaml
├── Clothing and Gear/         Full outfits, individual pieces, themed collections
│   ├── Clothing Mods/
│   ├── Individual Clothing Pieces/
│   └── Themed Outfits/
├── Art/                       Artist names and art style keywords
│   ├── Artist Names.txt
│   ├── Art Keywords.txt
│   └── award-photo.yaml
├── Poses-Locations/           Poses, locations, YAML pose libraries
│   ├── GirlPoses.txt
│   ├── Poses.yaml             YAML pose library with full tag support
│   ├── sex-poses.yaml
│   ├── Location.txt
│   ├── indoor.txt
│   └── outdoor.txt  ...
├── Prompt/                    Model quality / prompt meta tags
│   ├── score.txt              score_9, score_8, score_7 …
│   ├── quality.txt            masterpiece, best quality …
│   ├── safety.txt             safe, sensitive, nsfw, explicit
│   ├── year.txt               year 2025, year 2024 …
│   ├── period.txt             newest, recent, mid, early, old
│   └── year_period.txt        combined year + period pairs
├── Meta/                      Visual modifiers: camera angle, colors, decorations
├──
└── [root]                     General: quality.txt, style.txt, subject.txt,
                               lighting.txt, mood.txt, colors.txt, where.txt …

Reference rules:

• Bare name: __GirlPoses__ — node finds the file anywhere in the tree
• Relative path: __Poses-Locations/GirlPoses__ — precise, preferred for ambiguous names
• Both .txt and .yaml files are loaded automatically from all subfolders

---

Node Reference

Wildcard Prompt

| Parameter | Type | Description | |---|---|---| | text | STRING (multiline) | Prompt with __wildcards__, <[tags]>, and {options} | | seed | INT | Seed for reproducible selections | | autorefresh | Yes / No | Yes: reload files each run (slow, live editing). No: cache (fast, default) |

Output: processed_text — fully expanded prompt string, also shown as copyable text in the node UI.

Latent Ratio Selector

Creates an empty latent tensor at a preset aspect ratio. See RATIO_SELECTOR_GUIDE.md for full documentation.

---

Troubleshooting

Wildcard not found

• Check the file exists in wildcards/ with a .txt extension
• Filenames are case-sensitive on Linux/Mac; match exactly
• Enable autorefresh if you just created the file without restarting

Same output every time

• Change the seed value, or connect to a random seed generator node

Artist names not getting @ prefix

• The node auto-converts Artist:Name tokens. If you're writing artist names manually, add @ yourself: @artgerm

YAML entry not selected

• Verify the tag spelling matches exactly (tags are case-insensitive)
• Check the YAML file is valid — look for ComfyUI console errors on load

Output has double commas / extra spaces

• This is cleaned up automatically. If it persists, check that wildcard lines don't start or end with commas

---

License

MIT — use and modify freely.

Built on the UmiAI wildcard concept for Stable Diffusion WebUI.