LoRA Loader (Tags)

LoRA Loader (Tags) does the ordinary LoRA-loading job and then one extra thing: it reads the training tags out of the LoRA file and hands them to you as a comma-separated string. Load the LoRA, apply it to your model and CLIP, and also get back the tags it was trained with - ready to paste straight into your prompt, or into any node that takes a string.

That sounds like a small convenience, but it kills a genuinely annoying workflow problem. Character and style LoRAs from the anime side of the ecosystem usually ship with their trigger words and tag list baked into the file metadata. Normally you hunt for those on the CivitAI page or in the description. This node just... reads them off the file. rgthree's Power Lora Loader popularized exactly this idea ("right-click and it gives you the trigger words"), and this is the same trick in a leaner, no-UI package.

How it works

The node loads the safetensors metadata header (kohya-format trainers write a ss_tag_frequency entry listing every tag the model saw during training and how often). It aggregates those across the training buckets, sorts by frequency, and returns the top N as a string. That's why the "tags" are the training tags, not necessarily the author's curated trigger words - though for most tag-trained anime LoRAs, the top training tags are exactly what you want in the prompt.

The inputs that matter

• model and clip - the base model and text encoder the LoRA gets applied to, same as any loader.
• lora_name - dropdown of your ComfyUI/models/loras folder.
• strength_model and strength_clip - LoRA strength, default 1.0, and they go negative if you want to subtract the LoRA. Set both to 0 and the node short-circuits, returning your model untouched with an empty string.
• tag_limit (default 10) - how many tags to emit. Raise it if the top 10 feel thin; the frequency ranking means the first few are the most important anyway.

The outputs

• MODEL - model with the LoRA applied.
• CLIP - CLIP with the LoRA applied.
• STRING - the tags, comma-separated. Wire it into a CLIP Text Encode prompt, a text-concat node, or a queue/prompt utility.

Install and gotchas

ComfyUI Manager (search "DareMerge") or:

cd ComfyUI/custom_nodes
git clone https://github.com/54rt1n/ComfyUI-DareMerge

Restart ComfyUI. Dependencies are matplotlib, numpy, torch, pillow - no model downloads; the LoRA files are your own.

The gotcha is honesty about where tags come from: if a LoRA wasn't trained with a kohya-style trainer that wrote ss_tag_frequency, the node returns an empty string. That's not a bug - the file simply has no tags in it. Also worth knowing: the tag list is training data, so it includes tags the author might not have meant as trigger words (background terms, quality tags, and so on). Treat it as a strong suggestion for prompting, not a guaranteed activation phrase, and check the top few against your actual output.