LoRA Unloader (INT-Crush)

The INT-Crush LoRA Unloader is the cleanup node of the pack, and it's about as unglamorous as ComfyUI nodes get: one input, one output, zero dials. It strips the INT-Crush LoRA residual buffers off a loaded model without forcing a reload. If you've never needed it, this paragraph is the whole article. If you've ever tried to reuse one loaded model for a LoRA pass and a clean pass, you know exactly why it exists.

Input: model (the MODEL output from an INT-Crush LoRA loader). Output: model, with the LoRA adapters cleared.

Why it has to exist

When the INT-Crush LoRA loader attaches an adapter, it writes residual buffers onto the quantized model's Linear modules - that's how it applies LoRA without corrupting the Hadamard-rotated weights (standard LoRA patching can't touch these models at all). Once those buffers are attached, the only ways to get them off are reloading the model from disk or running this node. If you're generating a batch of images from one loaded model - some with a LoRA, some clean - the unloader lets you swap between states without paying the reload tax each time. It also handles the ComfyUI cache edge case: it clones the patcher, bumps the patch UUID, and registers an ON_LOAD callback that clears adapters if the model gets reloaded from cache with inherited state. So the model you get out is genuinely clean, not just pretending to be.

When you actually need it

Honestly, niche. In a single linear pass you never need it - just don't wire a LoRA loader in and there's nothing to unload. It earns its keep in two situations: multi-pass or batch workflows that reuse a patcher across different LoRA states, and side-by-side comparisons where you want the exact same quantized model with and without a LoRA and don't want to wait for a reload between them.

One thing to know so you don't misuse it: it only clears INT-Crush adapters - the _intcrush_adapter buffers that IntCrushLoRALoader attaches. It's a companion to that node, not a general-purpose LoRA unloader, and it won't touch anything a standard ComfyUI LoRA node did.

Install

Same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/ThunderFun/ComfyUI-INT-Crush

Restart ComfyUI, or grab it via ComfyUI Manager by searching "ComfyUI-INT-Crush". There's no separate install for this node - it ships with all five nodes in the pack.

If you find yourself putting an unloader between every LoRA loader in a workflow, step back: it's a signal your workflow is fighting the model cache, and a fresh load or a restructured graph is often cleaner than a chain of unloaders. But for the occasional "clean run, please," this is exactly the node you want.