Simple Tagger for Stable Diffusion WebUI Forge Classic

A minimal image-tagging tab for Forge Classic portable — ONNX only, no tensorflow.

Interrogate any image and get Danbooru-style tags, tunable thresholds, and one-click copy — without pulling in tensorflow/jax (which break Forge via protobuf/transformers conflicts).

---

Screenshots

---

Features

• ONNX-only inference (onnxruntime, CPU or CUDA) — WD taggers (CSV labels) and Camie tagger v2 (JSON metadata).
• Curated model list — v3 family plus MoAT v2, no duplicate generations.
• Auto-discovery — any extra *.onnx + *.csv (or *-metadata.json for Camie) found in the search folders appears in the model dropdown as (auto-local).
• Settings tab section — UI language, custom models folder, Rescan and Download from HuggingFace buttons.
• i18n — every *.json in locales/ becomes a selectable UI language (ships with en + uk, English default). Language applies live, no restart.
• Persistent settings — settings.json next to the extension is the single source of truth; components opt out of Forge ui-config.json (do_not_save_to_config) so values survive restarts.
• Result gallery — saves interrogated images under output/tagger_output/simple-tagger, skips pixel-identical duplicates, and lets you click a thumbnail to reload it.

---

Installation (Forge Classic portable only)

1. Copy (or git clone) this folder into your extensions directory: <forge-classic>\extensions\simple-tagger
2. Restart Forge. install.py pulls in onnxruntime + huggingface_hub automatically on first start.
3. Open the Simple Tagger tab.

Not tested with A1111 — Forge Classic only.

---

Models

Dropdown entry	Type	Source repo
WD ViT v3 (local)	WD	SmilingWolf/wd-vit-tagger-v3
Camie v2 (local)	Camie	local weights only
WD EVA02 Large v3	WD	SmilingWolf/wd-eva02-large-tagger-v3
WD SwinV2 v3	WD	SmilingWolf/wd-swinv2-tagger-v3
WD ConvNext v3	WD	SmilingWolf/wd-convnext-tagger-v3
WD MoAT v2	WD	SmilingWolf/wd-v1-4-moat-tagger-v2

Models download on first use. WD weights land in <models_path>/tagger_models/<dirname>/.

Where the extension looks for weights

In order, first hit wins:

1. Custom folder from Settings (flat <basename>.onnx + <basename>.csv, or <basename>.onnx + <basename>-metadata.json for Camie) — this is how you reuse weights from other installs (e.g. a ComfyUI models/ folder): just point the setting at it.
2. <models_path>/tagger_models — the shared Forge storage, incl. the canonical <dirname>/model.onnx + selected_tags.csv HF cache layout.
3. models/ inside this extension (good for a fully self-contained install).
4. HuggingFace — auto-download on first use, or on demand via the Download selected model button.

Use Rescan models after adding/removing weight files.

---

Settings

Stored in settings.json (created on first run, git-ignored). Copy settings.example.json → settings.json to pre-seed values:

{
  "model": "WD ViT v3 (local)",
  "gen_thresh": 0.35,
  "char_thresh": 0.85,
  "exclude": "",
  "save": true,
  "lang": "en",
  "custom_dir": "",
  "verbose": false
}

Console output is quiet by design: one ready: line at UI start, plus warnings/errors only. Set "verbose": true in settings.json if you need the old full-dict logging on every load/save.

Adding a language

Drop a new <code>.json into locales/ mirroring en.json keys ("_name" holds the display name). It appears in the language dropdown automatically — no code changes needed.

---

Repository layout

forge-simple-tagger/
├── scripts/simple_tagger.py   # the extension (Forge loads scripts/*.py)
├── locales/en.json, uk.json   # UI dictionaries (*.json = a language)
├── assets/                    # README screenshots + section icons
├── models/.gitkeep            # optional self-contained weights dir
├── install.py                 # Forge auto-installs onnxruntime + huggingface_hub
├── settings.example.json      # seed values for settings.json
├── .gitignore                 # ignores settings.json, models/, __pycache__
├── LICENSE                    # MIT
└── README.md

---

License

MIT — see LICENSE.