Lortnoc Loader 🍌

Say "lortnoc" backwards. That's not an accident - it's "control" spelled in reverse, and it's exactly the kind of dry joke that fits a node the pack's own README describes, in the author's own words, as "謎のControlNet": a mysterious ControlNet. This isn't lllyasviel's ControlNet architecture or one of the union checkpoints from Alibaba PAI or InstantX that dominate spatial conditioning today - it's a personal, experimental technique for image-conditioned model patching, documented mainly in a note.com blog post rather than a paper.

What it's doing structurally

Real ControlNet works by training a whole trainable copy of the model's encoder and feeding its output back in through zero-initialized convolutions. Lortnoc, going by its inputs, takes a lighter-weight approach: it patches your MODEL directly using a loaded file plus the actual conditioning image, with two independent strength dials rather than one. That two-dial shape - a strength for the trained weight patch itself, and a separate strength for how much the live input image drives the effect at inference - reads like a LoRA-shaped conditioning module rather than a full second encoder network, closer in spirit to the lightweight ControlNet-LLLite idea (a LoRA-like module injected at attention layers, rather than a duplicated encoder) than to a standard ControlNet union.

Where this really is a "mysterious" technique rather than an established one: there's essentially no public discussion of it, no known model zoo of Lortnoc files the way ControlNet has union checkpoints for every current architecture, and the one citation the README gives is the author's own blog post. If you want image-conditioned spatial control with broad model availability and documented behavior, standard ControlNet is the better-trodden path. This node is for people who enjoy the rabbit hole for its own sake, or who specifically want to try what the author built.

Inputs and outputs

Required:

• model - the MODEL to patch.
• image - the conditioning image driving the effect.
• file_name - a combo, picking a Lortnoc file from wherever this pack expects them to live locally.
• strength_lora - -20 to 20, default 1. The weight-patch strength.
• strength_hint - -20 to 20, default 1. How strongly the live image conditions the output.

Output is a single MODEL - patched, ready to feed into a normal KSampler the same way any patched model works.

Installing it

ComfyUI Manager, search "cgem156-ComfyUI". Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/laksjdjf/cgem156-ComfyUI

Restart ComfyUI afterward. Note the file_name field is a dropdown, not a free-text HF repo id like the Dart loaders - that means it's reading from a local folder, and you'll need an actual Lortnoc-format file placed there before the combo has anything to select. Given how little published material exists for this technique, sourcing that file is likely to mean training your own via whatever process the linked note.com article describes, rather than downloading a ready-made one.

Common issues

Both strength fields accept a striking -20 to 20 range - far wider than typical ControlNet conditioning scales (usually 0.5-1.0). Start much closer to 1 on both and adjust from there; the wide range invites overcorrection long before it invites useful control. Beyond that, there's genuinely no community troubleshooting trail here - you're working from the author's own note and whatever you can infer from experimenting, which is the honest state of this corner of the pack.