Eric Qwen-Edit Unload LoRA

Eric Qwen-Edit Unload LoRA strips every LoRA adapter off the Qwen-Edit pipeline and restores the base weights. One input (pipeline, a QWEN_EDIT_PIPELINE), one output (the same pipeline, LoRA-free). It's the sibling of Eric Qwen-Edit Apply LoRA, and together they give you a clean way to stack adapters - apply a couple, run an edit, unload them all, and the next edit starts from a stock model instead of whatever the previous experiment left behind.

This matters more with Qwen-Edit than it does with SDXL. The Edit line grew its own LoRA library - pose, angle, style, and character LoRAs trained against the edit model itself - and they're cheap to stack and just as cheap to forget you stacked. Qwen-Image-Edit-2511 even folds "integrated LoRA capabilities" into the base weights, so a LoRA that's still loaded from last week is quietly steering every edit you run now. The unload node is your "start clean" reset.

Why not just reload the pipeline?

Fair question. If you have keep_in_vram on (you should), reloading means re-reading the pipeline config and often re-firing the transformer - a long wait when you're iterating. Unloading the LoRA is a lightweight in-place operation: the adapter weights are peeled off the transformer and the base weights are restored, no pipeline teardown. That's also why the output is the same pipeline object, not a new one - you can wire it straight back into your edit node and keep going.

The practical workflow the author clearly had in mind:

1. Apply LoRA → Edit node → inspect result.
2. Try a different LoRA: Unload LoRA → Apply LoRA (new one) → run again.

That's the loop where this node earns its place - it's the "undo last experiment" button.

The input that matters

• pipeline - connect the LoRA-loaded pipeline. That's it. There's no "select which LoRA" here; it's an all-at-once reset, which is exactly what you want when the point is to guarantee a clean slate. If you only want to drop one of several LoRAs and keep the rest, this node isn't the tool - rebuild the chain (or check the per-stage weight controls on Eric Qwen-Image Apply LoRA for the generation side, which handles stacking differently).

The output pipeline is the same QWEN_EDIT_PIPELINE type, so it drops back into any downstream edit node (Image, Inpaint, Multi-Image, Style Transfer) with no conversion.

One thing to note if you're coming from the standard ComfyUI LoRA habits: this pack runs the diffusers loading path, not ComfyUI's native LoRA loader, so "unload LoRA" is a real diffusers operation - it calls unload_lora_weights on the pipeline and, per the README, restores the base model. You can't fake it by disconnecting a wire; the adapter stays fused into the transformer until this node (or a pipeline reload) runs. If you ever see an edit that's clearly being influenced by a LoRA you thought you'd removed, this is the node that was missing from your graph.

Install: ComfyUI Manager search "Eric Qwen Edit", or git clone https://github.com/EricRollei/Eric_Qwen_Edit_Experiments into custom_nodes/ and restart. It needs the edit model loaded upstream - it's a cleanup node, not a loader, so no extra downloads.