A ComfyUI extension that tags images with booru tags for use in prompts. It bundles the commonly used open tagger models — WD, Pixai, Camie, CL Tagger and AnimeTimm — 23 taggers in total.

Features

Model choice — 23 taggers across WD, Pixai, Camie, CL Tagger and AnimeTimm; AnimeTimm models use 12,476-tag vocabularies. Use the one you prefer.
Loads once — a model is loaded the first time it is used and kept in memory, so later images do not wait for another load.
Multiple images per run — the node accepts a batch of images and returns tags for all of them.
Tag groups — results are split into general, rating and character (character/copyright/artist) tags, with options for underscores, sorting, tag exclusions and per-model thresholds.
Automatic downloads — missing model files are downloaded on first use; gated models require a HuggingFace token (accepted once on the model page).

For AnimeTimm models, use_best_threshold defaults to enabled and applies each tag's model-provided best_threshold as a minimum. Disable it to use only the node's threshold and character_threshold inputs, which is useful when you prefer higher recall. The defaults are the model author's category-level recommendations: eva02 (0.39 / 0.61), caformer (0.39 / 0.47), swinv2 (0.41 / 0.59), and ConvNeXtV2 Huge (0.38 / 0.51) for general / character tags.

Outputs

| Output | Description | |---|---| | tags | Combined character + general tags | | general_tags | Descriptive tags (attributes, clothing, composition, etc.) | | rating | Top rating tag (safe / sensitive / questionable / explicit) | | character_tags | Character, copyright, and artist tags |

Models

| Model | Parameters | Tags | Input Size | License | Gated | |---|---|---|---|---|---| | WD Series (eva02, vit, swinv2, etc.) | varies | varies | 448² | MIT | No | | Pixai Tagger v0.9 | — | 13,461 | 448² | Apache-2.0 | No | | Camie Tagger v2 | — | 70,527 | 512² | ? | No | | CL Tagger v1 (1.00 / 1.01 / 1.02) | — | 42,163 | 448² | Apache-2.0 | No | | CL Tagger v2 (2.00 / 2.01a) | — | 106,536 / 108,036 | 384² | Custom | Yes | | AnimeTimm swinv2_base | 99.7M | 12,476 | 256² | GPL-3.0 | Yes | | AnimeTimm caformer_b36 | 134.0M | 12,476 | 384² | GPL-3.0 | Yes | | AnimeTimm eva02_large | 316.8M | 12,476 | 448² | GPL-3.0 | Yes | | AnimeTimm ConvNeXtV2 Huge (community ONNX) | 692.6M | 12,476 | 512² | GPL-3.0 | Yes |

ConvNeXtV2 Huge uses the community ONNX conversion from itterative, with official AnimeTimm metadata and preprocessing.

Gated models require a HuggingFace token. Accept the license on the model page, then either run huggingface-cli login or set the HF_TOKEN environment variable before first download.

Credits:

Models created by:

WD Taggers: SmilingWolf
Pixai Tagger: pixai-labs
Camie Tagger: Camais03
CL Tagger v1 / v2: cella110n
AnimeTimm: DeepGHS / narugo1992

Installation

Install from the Comfy Registry (ComfyUI-Manager or comfy node install booru-tagger), or clone this repo into the custom_nodes folder.
Install dependency (onnxruntime or onnxruntime-gpu for CUDA acceleration). When installing via the registry this is handled automatically.
For gated models (CL Tagger v2, AnimeTimm): run huggingface-cli login once, or set the HF_TOKEN environment variable with your HuggingFace token.

Configuration

Edit models.json to customize defaults:

settings.ortProviders — ONNX Runtime execution providers. Providers that the installed onnxruntime build does not ship are dropped automatically (with a warning).
settings.preprocess — "tensor" (default) resizes/pads/normalises on the GPU straight from ComfyUI's image tensor; "pil" restores the original CPU/PIL path (same tags, ~13 ms slower per image).
threshold / character_threshold — Per-model default thresholds.
HF_ENDPOINT — Mirror/proxy URL for HuggingFace downloads.