Camie Tagger

Feed this node an image and it hands you back a string of Danbooru-style tags describing what's in it. That's an interrogator - the reverse of prompting. You use it for two things: figuring out what tags to write to reproduce a look you like, and captioning a folder of images for LoRA training in exactly the tag format anime models expect.

Camie is one of three tagger models this pack supports (the others being the well-known WD tagger from SmilingWolf and the CL tagger). Camie specifically is Camais03's tagger, a slightly niche but real alternative to WD - it comes up in community discussion, just not as often. The node is the ComfyUI wrapper; the actual work is done by an ONNX model you have to download separately.

How it works

Under the hood it's a vision model trained on Danbooru-tagged images, run through ONNX Runtime. It scores every tag it knows against your image and keeps the ones above your confidence thresholds. Because it learned the same vocabulary Illustrious and NoobAI were trained on, the tags it produces drop straight back into a prompt for those models - that round-trip is the whole reason to use a Danbooru tagger over a generic caption model. It'll happily take a batch of images too; multi-image input just separates each image's tags onto its own line.

The inputs that matter

• image - what you're tagging.
• model_name - the ONNX model to use. If this dropdown only shows None, you haven't installed the model yet (see below). This is the single most common confusion with this node.
• general - the confidence threshold for general tags (default 0.49). Lower it to catch more, subtler tags; raise it to keep only the confident ones. This is your main quality/quantity dial.
• min_confidence - a floor (default 0.01) below which tags are dropped entirely.
• categories - which tag types to emit and in what order, e.g. rating,artist,general,character,copyright,meta,year. Trim it to just general if you only want descriptive tags and not the character/artist guesses.
• exclude_tags - a blocklist. Great for stripping tags you never want, like a rating or a watermark tag.
• session_method - CPU, GPU, or their Release variants. GPU is faster; the Release versions free the model from memory after tagging, which matters if you're tight on VRAM and don't want the tagger squatting on it between runs.

Output is a single tags string.

Install

ComfyUI Manager → search ComfyUI_Mira → Install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/mirabarukaso/ComfyUI_Mira.git and restart. If it errors on load, pip install -r requirements.txt in the pack folder - that also pulls in onnxruntime, which the taggers need.

The model download is the part people miss. The taggers do not ship weights. You download the Camie v2 ONNX model and its metadata JSON, drop them in ComfyUI/models/onnx/camie_tagger/, and - this matters - rename them to the exact expected filenames:

ComfyUI/models/onnx/camie_tagger/
  ├── camie-tagger-v2.onnx
  └── camie-tagger-v2-metadata.json

Get the folder or the names wrong and model_name stays stuck on None.

Common issues

The number-one problem is model_name showing only None - that's always the download. The files aren't there, aren't named right, or aren't in models/onnx/camie_tagger/. Fix the path and restart. Second: if you want the GPU session but tagging still runs slow, you likely have CPU-only onnxruntime installed rather than onnxruntime-gpu; that's an environment thing, not a node bug. Third, on VRAM: if you're chaining this before a heavy sampler and hitting OOM, switch session_method to a Release variant so the tagger lets go of memory when it's done.