JoyCaption Lora

This is the smallest node in the pack and it does one job: pick a JoyCaption LoRA from a folder and hand it to a loader. It lists directories inside models/LLavaloras, and its single output (lora, type JOYCAPTIONLORA) plugs into the optional lora input on either the JoyCaption Loader or the JoyCaption Download And Load node. The loader then applies it on top of the base model, and your caption node uses the fine-tuned version.

What would you use it for? The same thing LoRAs are for anywhere else: changing behavior cheaply instead of loading a whole different model. There's a small ecosystem of JoyCaption LoRAs that retune the captioner - style-specific or tag-focused captioners, for example, or one tuned to emit captions formatted a particular way for your training pipeline. Wire one in and the captions you get out shift toward whatever that LoRA was trained on, while the base model underneath stays put.

Worth knowing before you rely on it

Two honest caveats. First, the author's own git history describes the LoRA support as "untested but should work" - it's the newest, least-exercised corner of the pack. The mechanism is sound (it loads the adapter through PEFT's PeftModel), but if you hit a load error or gibberish captions with a LoRA attached, that's the first place to look, and the LoRA you found might be built for a different JoyCaption version. Second, the folder rule from the loader applies here too: the dropdown lists directories in LLavaloras, so each LoRA needs its own subfolder, not loose files.

Install

Same pack, same steps as the rest of these nodes - ComfyUI Manager → search wo_joycaption_comfyui, or git clone https://github.com/without-ordinary/wo_joycaption_comfyui into custom_nodes and restart. PEFT comes in via the pack's dependencies. If you've never used it, the practical path is: download a JoyCaption LoRA, put it in models/LLavaloras/<name>/, drop this node in, and feed its output into the loader's lora socket alongside your joycaption_model into the main node. It's a thin utility, but it's what turns the captioner from a fixed tool into something you can tune.