Icy LoraLoader

LoraLoader is the node you'll use more than almost any other once you get past generating with a plain checkpoint. It takes a diffusion model and its CLIP text encoder, applies a LoRA - a small adapter file that patches the model to draw a specific character, style, or concept - and hands you back both, modified. The "Icy" version does all of that identically; it just lives under the IcyHider Comfy Core category so its preview and the rest of the graph stay hidden until you hover.

And here's the thing about LoRA nodes in particular: the file name on that node is often the most identifiable thing on screen. The pack's whole pitch - privacy, clean screenshots, working in public - runs straight through this node, because the LoRA you've got loaded is frequently the one thing you don't want casually visible.

On startup, IcyHider subclasses every core ComfyUI node, keeping the exact same code and inputs and only changing the category. So this is core LoraLoader with a coat on. Everything you know about stacking LoRAs applies.

The inputs that matter

From the schema, five inputs, and three are the ones you'll actually touch:

• model - the MODEL to patch (wire from a CheckpointLoader or another LoRA node).
• clip - the CLIP encoder, patched in parallel.
• lora_name - dropdown of .safetensors files in your ComfyUI/models/loras folder. Drop files there and hit refresh to see them.
• strength_model - how hard the LoRA affects the diffusion model, default 1.0. This is the dial you'll actually turn. As the KB's concepts essay notes, default 1.0 is often too strong - 0.5–0.8 is a common starting point, and negatives invert the LoRA.
• strength_clip - how hard it affects the text encoder. Lower it if a LoRA is mangling your prompt understanding while the imagery looks right.

Outputs: MODEL and CLIP, both modified, ready to chain into more LoRA nodes or into a sampler. The author's tooltip spells out the use: multiple LoRA nodes can be linked together, which is how style + character stacks happen.

The gotcha that burns people

LoRA compatibility is not universal. An Illustrious LoRA won't work on a Flux base, and vice versa - a mismatched LoRA usually doesn't error, it just silently produces garbage or nothing. The KB's lora-training doc hammers this: "an Illustrious LoRA won't work on Flux, and vice versa." If your output suddenly looks off after adding a LoRA, the first question isn't strength - it's whether the LoRA matches your checkpoint family at all.

Install

Frontend-only pack, no Python dependencies, and no model downloads from the pack itself:

# ComfyUI Manager: search "IcyHider", Install, restart

# or:
cd ComfyUI/custom_nodes
git clone https://github.com/icekiub-ai/ComfyUI-IcyHider.git
# restart ComfyUI

Pack-wide: the hiding is DOM-based, so the newer Nodes 2.0 canvas frontend may not engage it, and it's screen-only - the LoRA file on disk is unaffected. Change Avalanche or hide mode and ComfyUI needs a reload to pick it up.