monbooru

README

Monbooru

Your own self-hosted and lightweight booru.
Designed for organizing your local media collection, including AI-generated images (Stable Diffusion, ComfyUI, A1111/Forge). Works fully offline and with no telemetry. Supports ONNX models for local auto-tagging of your collection (WD14, JoyTag...).

---

Features

• Tag-based gallery with folder tree, favorites, saved searches, related-image suggestions
• A watcher that picks up new, moved, and deleted files in the gallery within a few seconds; and support for multi-file browser upload
• Stable Diffusion metadata extraction from A1111/Forge and ComfyUI (prompts, models, seeds, full workflow)
• Optional auto-tagging with local ONNX models (WD14, JoyTag, or any compatible model), CPU or GPU (CUDA)
• Search with wildcards, OR, exclusions, plus filters on folder, date, size, dimensions, category, generation recipe...
• Batch operations: bulk delete, bulk move, bulk auto-tag, delete-all-search-results, tag-all-search-results / tag-selected (add or remove tags in one go)
• Create multiple galleries with their own filesystem and switch between them
• Per-gallery export and import: full backup (.db, .json, .zip with images), or a light format (.json or .zip with images) with only images<->tags map. See docs/MIGRATING.md for migrating from another supported booru-style app.
• REST API for third-party integrations (e.g. adding images to the gallery from an external app)
• Daily maintenance schedule: scheduled sync, auto-tag, recompute counts, merge general tags, and vacuum databases...
• Fully offline, no telemetry

---

Quick start (Docker)

Edit the volume paths in docker/docker-compose.yml, then docker compose up -d. The app is available at http://localhost:8080.

---

Search syntax

Tags separated by spaces means AND. Everything else stacks on top:

Syntax	Effect
cat dog	has both tags
cat OR dog	either one
-blonde_hair	exclude
blue* / *hair*	wildcards
fav:true	favorites only
source:a1111 / source:comfyui / source:none	by metadata source
folder:2024/january	images in this folder or any subfolder
folder:"my set 1"	quote paths that contain spaces
folderonly:2024/january	only images directly in this folder, no subfolders
width:>=1920 height:<768	dimensions
date:2024-03-15 date:>2024-01-01 date:2024-01..2024-06	dates
cat:character	any tag in that category
character:cat	tag "cat" in the character category
missing:true	files gone from disk
animated:true / animated:false	gif/mp4/webm
tagged:true / tagged:false	images with or without tags
autotagged:true / autotagged:false	images with or without auto-tags
generated:abcd1234abcd	same generation recipe (hash shown on the image page)

Autocomplete is combination-aware: the count next to each suggestion is for the full query, and suggestions that would return zero results are hidden.

Sort options: newest, file size, random shuffle. Any query can be saved from the sidebar for one-click access later.

---

Tags

Five built-in categories (General, Character, Artist, Copyright, Meta), each with its own color. Custom categories with custom colors are also supported.

• Add tags manually or let the auto-tagger do it
• The same name can live in multiple categories (character:cat vs artist:cat)
• Merge tags, rename, move between categories
• Category-aware autocomplete: typing artist: limits suggestions to artist tags

---

Auto-tagger

You can use autotaggers to tag your images automatically.
Place a model in the models/ volume. Each tagger lives in its own subfolder with two files:

• model.onnx, the weights
• tags.csv in WD14 format (tag_id,name,category_id), or tags.txt with one label per line (all labels are assigned to general)

If neither default filename is present, monbooru picks the lone .onnx plus the lone .csv/.txt found in the folder, so models distributed with nonstandard names (e.g. top_tags.txt) work without extra config.

Restart the container after adding a model, then enable it in Settings → Auto-Tagger. Multiple taggers can run at the same time, and each tag records which tagger produced it. Auto-tags can be removed per tagger or all at once without touching manual tags.

Some models that have been tested :

• Anime: https://huggingface.co/SmilingWolf/wd-swinv2-tagger-v3
• Realistic: https://huggingface.co/fancyfeast/joytag

Example setup (WD14 SwinV2):

MODEL_DIR=/data/monbooru/models
mkdir -p "$MODEL_DIR/wd-swinv2"

wget -O "$MODEL_DIR/wd-swinv2/model.onnx" \
  https://huggingface.co/SmilingWolf/wd-swinv2-tagger-v3/resolve/main/model.onnx

wget -O "$MODEL_DIR/wd-swinv2/tags.csv" \
  https://huggingface.co/SmilingWolf/wd-swinv2-tagger-v3/resolve/main/selected_tags.csv

If no tagger is enabled, auto-tagging is simply disabled. The rest of the app works normally.

GPU (CUDA)

The default image is CPU-only (~210 MB). For GPU inference, switch to the -cuda image (~2.3 GB), pass the GPU into the container the usual way, then enable Settings → Auto-Tagger → Use GPU (CUDA) (or set MONBOORU_TAGGER_USE_CUDA=true).

The current mode (GPU or CPU) is shown as a badge in the Auto-Tagger section. The -cuda image also runs on CPU when the toggle is off, so switching between the two does not require a rebuild. Worker count is configurable in Settings (default 8); increase it on GPU if preprocessing becomes the bottleneck.

---

Environment variables

All of these override the TOML config. Pattern: MONBOORU_{SECTION}_{KEY}.

Variable	Overrides	Type
MONBOORU_SERVER_BIND_ADDRESS	server.bind_address	string
MONBOORU_SERVER_BASE_URL	server.base_url	string
MONBOORU_PATHS_DATA_PATH	paths.data_path	string
MONBOORU_PATHS_MODEL_PATH	paths.model_path	string
MONBOORU_GALLERY_WATCH_ENABLED	gallery.watch_enabled	bool
MONBOORU_GALLERY_MAX_FILE_SIZE_MB	gallery.max_file_size_mb	int
MONBOORU_TAGGER_USE_CUDA	tagger.use_cuda	bool
MONBOORU_AUTH_ENABLE_PASSWORD	auth.enable_password	bool
MONBOORU_AUTH_PASSWORD_HASH	auth.password_hash	string
MONBOORU_AUTH_SESSION_LIFETIME_DAYS	auth.session_lifetime_days	int
MONBOORU_AUTH_API_TOKEN	auth.api_token	string
MONBOORU_UI_PAGE_SIZE	ui.page_size	int
MONBOORU_LOG_LEVEL	log.level	warn / info / debug

Per-tagger settings (enable flags, confidence thresholds, worker count) live in the Settings UI, not in env vars.

---

Keyboard shortcuts

Key	Where	Action
s	Anywhere	Focus search input
t	Image detail	Focus tag input
f	Image detail	Toggle favorite
Delete	Image detail	Delete image (advances to the next in the current search)
Space	Image detail	Play / pause the video
h / l	Gallery	Previous / next page
Arrows	Gallery	Navigate the grid
Enter	Gallery	Open focused image
← / →	Image detail	Previous / next image
Escape	Anywhere	Close dialog, blur input, or go back

---

REST API

Disabled by default. Generate a token in Settings → Authentication to enable it.

• HTML reference: /api/v1/docs (also linked in the footer)
• OpenAPI spec: /api/v1/openapi.json

Covers search, tag add/remove, upload, delete.

---

Building without Docker

# CPU only, no auto-tagger
go build -o monbooru ./cmd/monbooru

# With auto-tagger (requires the ONNX Runtime shared library on the system)
CGO_ENABLED=1 go build -tags tagger -o monbooru ./cmd/monbooru

./monbooru -config /path/to/monbooru.toml

---

Inotify limit (Docker)

If the watcher reports an inotify limit error (logged as no space left on device from inotify_add_watch), raise fs.inotify.max_user_watches on the host (not inside the container) and restart. If you see too many open files, raise fs.inotify.max_user_instances. Alternatively, disable the watcher in Settings and use the manual Sync button when adding new files.

---

Warning

Intended for local network use. This project is not designed for direct exposure to the public internet.