Local Lora Gallery (Model Only)

This is the sibling of the main Local Lora Gallery node, and if you mostly run newer models, it might be the one you actually want day-to-day. Same card gallery, same stacking, same tags and filters - but it takes only a model, never touches a CLIP, and under the hood it applies LoRAs with ComfyUI's LoraLoaderModelOnly instead of the full LoraLoader.

Why you'd reach for it

LoRAs traditionally had two halves: a model patch and a text-encoder patch. The classic LoraLoader applies both. But a lot of what people run now doesn't need the CLIP half at all. On the newer architectures the text encoder is a big LLM that isn't part of LoRA stacking - Flux-family LoRAs, for instance, patch the diffusion model and leave the encoder alone. If you've already got conditioned text from elsewhere, or the model in front of you has no CLIP to speak of, dragging around a CLIP port you'll never fill is just clutter.

So the use cases:

• Your LoRAs are model-only (the common case on Flux and newer bases).
• You're conditioning your prompt separately and just want to bolt LoRAs onto the model.
• You're working with a model that has no CLIP input at all.

How it works

Identical machinery to the main node: the JS gallery reads your models/loras folder, shows preview cards (same-filename image or video next to each LoRA), and serializes your selection into a hidden selection_data widget that rides along in the workflow. On queue, load_loras walks the stack in order and applies each enabled, non-zero-strength LoRA model-only. It also does the same Nunchaku auto-detection as its sibling - a Nunchaku-wrapped model gets the accelerated loader, anything else gets the standard model-only path.

Inputs and outputs

One required input: model. Two outputs:

• MODEL - straight into your sampler or whatever's next.
• trigger_words (STRING) - the concatenated trigger words of every enabled LoRA that has them. Handy to feed into your prompt node.

The trigger_words output is the same nicety as the full node: sync a card with Civitai (the ☁️ button) and the node remembers its trigger words, then pipes them out for you.

The tradeoff

The model-only path is strictly lighter, but it skips any text-encoder patch a LoRA might carry. Old SDXL-era LoRAs are the ones that sometimes put real weight on the CLIP side - if you're applying those, use the full Local Lora Gallery node with clip connected instead. If you're on a modern base, you'll probably never miss it.

Installing it

Same as the rest of the pack - no extra Python dependencies, it wraps ComfyUI's own loader.

cd ComfyUI/custom_nodes
git clone https://github.com/Firetheft/ComfyUI_Local_Lora_Gallery.git

Then restart ComfyUI. Or install the whole pack by searching "Local Lora Gallery" in ComfyUI Manager. The Nunchaku plugin is optional and only matters if you run accelerated models.

Gotchas

• No previews until you add them. Drop a same-named image next to the LoRA, or hit the sync button once (needs network) and it'll fetch one from Civitai.
• Don't expect the CLIP side to move. If you hook this into an SDXL workflow and wonder why a LoRA's text-encoder effect is missing, that's the model-only tradeoff biting you - switch to the full node.
• Save your workflow or the stack selection can vanish; it lives in the hidden selection_data property.