JLC LoRA Loader - Multi-Model / CLIP

If you've ever stacked LoRAs in ComfyUI the native way, you know the drill: three or four LoraLoader nodes in a chain, each with its own strength widget, and by the time you want to tweak the second one you've lost the plot. This node is the answer to that. JLC LoRA Loader - Multi-Model / CLIP gives you up to ten LoRA slots in one node, each with its own file selector, MODEL strength, and CLIP strength - and it hides the empty rows so your graph doesn't look like a wall of widgets.

It ships in jlc-comfyui-nodes by J. L. Córdova (GitHub: Damkohler), a pack that mostly orbits ControlNet composition but carries this whole dynamic-LoRA family as supporting cast. The Loader is the plain MODEL+CLIP variant - you'd pick it when a workflow wants LoRA patches on both the diffusion model and the text encoder, which is most character/style LoRAs. If you never touch CLIP strength, the MODEL-only siblings in the same pack skip text-encoder patching entirely.

How it works

The trick is that the node always predeclares all ten slot groups, then a little frontend script hides rows above slot_count. Under the hood:

• slot_count (1–10, default 1) is authoritative - the backend only inspects slots 1 through slot_count. A hidden slot with a LoRA loaded and nonzero strengths is simply ignored.
• Hidden slot values stay serialized in the workflow JSON, so you can drop slot_count to tidy the canvas and crank it back up later without re-picking your LoRAs.
• Slots with "None" as the LoRA, or both strengths at 0.0, are skipped.
• Active LoRAs apply sequentially in slot order. With two style LoRAs that fight each other, order matters - swap them and the result changes.
• A per-node state cache keeps each LoRA's state dict in memory for the session, so repeated runs don't reread the same files.

The inputs that matter

Wire your checkpoint's model and clip in from the loader or another LoRA node, then set slot_count. For each visible slot you get three widgets:

• lora_01 - dropdown of your models/loras folder, "None" by default
• strength_model_01 - how hard the LoRA hits the diffusion model (0–1 is the sane range, though the widget allows −10 to 10)
• strength_clip_01 - same idea for the text encoder

Two gotchas that bite everyone: strengths default to 0.0, so selecting a LoRA does nothing until you raise them, and the negative range exists on purpose - negative LoRA strengths are a real (if advanced) technique for subtracting a style, not a UI bug.

Outputs are model and clip, which feed straight into your KSampler's model and the CLIP Text Encode for your prompt. That's the whole wiring.

Installing it

ComfyUI Manager (search JLC ComfyUI Nodes, it's on the registry), or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Damkohler/jlc-comfyui-nodes.git

Then restart ComfyUI. Update with git pull inside the folder. Good news: the LoRA loaders need no model downloads and no extra Python dependencies - no requirements.txt, nothing. (The optional comfyui_controlnet_aux install the README mentions is only for the pack's ControlNet Aux wrapper, not this node.)

Where it fits

Honest framing: rgthree's Power Lora Loader is the community default for this job, and if you already run it, this node isn't a must-install. It's worth reaching for inside a JLC-heavy workflow, or when you want independent MODEL and CLIP strengths per slot - something the plain stack approach makes fiddly. Pair it with the pack's other dynamic loaders when you want to start slicing LoRA influence across the network's blocks; this one is the clean baseline. If your model doesn't benefit from text-encoder patching, grab the MODEL-only variant instead and skip the CLIP wiring entirely.