LoRA Tag (Model Only) 💎 Just Nodes

LoRA Tag (Model Only) is for people who miss A1111's way of doing things. In the old web-UI world, you could write <lora:style_name:0.8> right into your prompt and the LoRA loaded itself, tag and trigger word traveling together. ComfyUI's standard answer is a separate LoraLoader node, which works but splits your LoRA from the text that goes with it. This node reads the <lora:name:strength> tags out of any text you give it and applies the LoRA directly to the model. The tag stays in the text for the CLIP encoder, and the model gets the LoRA weights. Both halves of the workflow agree, in one place.

You'll see this pattern if you hang around the community long enough - "I keep my loras and trigger words together on the same line, <lora:something:1>, trigger_word" is a genuine workflow people build for batch generation, and dedicated tag-loader nodes exist for it. This is Just Nodes' take.

The mechanism, grounded in the code: a regex pulls every <...> tag out of the text, each one is parsed for lora, the filename, and an optional strength, and the filename is matched against your loras folder by prefix - so <lora:epiNoiseoffset:1> finds the first file in your folder whose name starts with epiNoiseoffset. It then loads the LoRA weights and applies them via ComfyUI's load_lora_for_models at the tag's strength (default 0.0 if you omit it, which is worth knowing). Applied LoRA files are cached, so repeated runs don't re-read the weights off disk. Model Only is doing real work in the name: it touches the MODEL only, not the CLIP text encoder - the standard LoraLoader applies to both. If a LoRA shipped with separate text-encoder weights, that half is ignored here.

Inputs are just model (any MODEL) and text (multiline). One output: the modified MODEL, straight into your sampler.

Installing it

Same as every Just Nodes node: ComfyUI Manager → search "Just Nodes" → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/Arroz-11/ComfyUI-Just-Nodes.git. No models beyond the LoRAs you already have, no extra dependencies, MIT licensed.

Gotchas

• Unknown or mistyped tags are silently skipped (with a console warning), not an error - you can ship a prompt with a tag that doesn't resolve and the run proceeds without it.
• Filename matching is by prefix, first match wins. Two LoRAs starting with the same name can resolve to the same file. Rename your files if you have foo.safetensors and foo_v2.safetensors and expect them to differ.
• Model only, not CLIP. LoRAs that bundle text-encoder changes won't apply them here.
• If the tag has no strength, it loads at 0.0 - effectively nothing. Write the strength explicitly.