Reference Tagging Options

Before Instant Reference LoRA trains anything, it has to know what your reference images are. That job goes to a WD14 tagger, which writes a caption for each image, and this node is how you steer that captioning. It's a bundle: plug one in, tweak the handful of fields, and pipe the tagging_options output into the training node's optional tagging_options socket.

Captioning is not a formality - it's half the training outcome. What you caption becomes what stays variable; what you leave out becomes what the model locks onto as "the identity." Get this wrong and your 50-step reference LoRA will happily learn the background instead of the character.

The inputs that matter

• general_threshold (0.35) and character_threshold (0.85) - the confidence cutoffs for the WD14 tagger. Lower general_threshold = more tags get kept, a noisier caption; raise it and captions get sparse. The default split - strict on characters, loose on general tags - is a sensible starting point. The default character_threshold of 0.85 is notably strict, so if a character in your reference set isn't getting named, lowering it is the first thing to try.
• prepend_tags / append_tags - tags forced onto the front or back of every caption. prepend_tags is where your trigger word goes. If you're training a character, a rare token here ("ocn7") is the classic move: caption it consistently, and the LoRA learns that token maps to your character.
• exclude_tags / replace_tags - filter or rewrite the tagger's output. Exclude is good for stripping a tag that's polluting every image (like a watermark tag); replace is a find-and-replace for when the tagger gets one consistent thing wrong.
• remove_underscore (true) - WD14 tags come out as long_hair; this turns them into long hair. Leave it on.

Outputs

One: tagging_options, the bundle that wires into the training node's tagging_options input.

Install

ComfyUI Manager → search "Instant Reference" → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/cstria0106/comfyui-instant-reference

Notes

This node only configures - the actual captioning happens during the training run, which is when the tagger model (SmilingWolf's WD14 convnext) gets downloaded if it isn't cached. So don't expect the tagging to happen when you run this node; it happens when the LoRA node runs. And since the captions are part of the cache key, changing any of these fields invalidates the cached LoRA and triggers a retrain. That's by design - you changed the dataset, so you get a new result - but it means dithering over thresholds costs you training runs. Set them, run, and iterate in batches rather than one tweak at a time.