Eric Krea2 Unload LoRA

Eric Krea2 Unload LoRA is a one-input, one-output utility with a single job: strip every LoRA adapter off the cached Krea 2 pipeline, right now, and clear the declared stack. If you're switching between style workflows in one ComfyUI session, this is the node that keeps a character LoRA from bleeding into your next render.

Why it exists

In this pack, LoRAs are applied to the cached pipeline - Eric Krea2 Apply LoRA fuses adapters onto the loaded Krea2Pipe and they stay there until something removes them. The pack's ephemeral LoRAs (the ones a sweep or a stack injection fuses per-cell) already clean up after themselves each run. This node is the explicit override for everything else: the LoRAs you applied in the panel and want gone before the next queue item. It calls unload_all_loras(pipe) and returns a fresh pipeline dict with lora_stack emptied, so you can keep threading the pipeline output onward in the graph.

It's the middle of the Apply / Unload / Diagnose trio the README advertises - apply to fuse, unload to strip, diagnose to figure out why a LoRA loaded as noise. If you're the kind of person who maintains a per-style workflow per tab, Unload LoRA sits at the top of each tab, right after the loader, and you never think about it again.

The one input that matters

The only input is pipeline (type KREA2_PIPELINE) - feed it the same pipeline object your Apply LoRA or the Component Loader handed down. The output is that pipeline, LoRA-free. That's the whole schema, and it's worth honoring: because it passes the pipeline through, you can drop it in the middle of a chain without breaking the graph. Put it before a Generate that needs the bare checkpoint, and the sampler sees a clean model.

Worth knowing for Krea 2 specifically: the community norm is train LoRAs on Raw, apply on Turbo - Turbo is the inference default and the distilled checkpoint most style work targets. Unloading is how you A/B a stack against the bare model in the same session: generate, unload, regenerate, compare, all without reloading 12B weights.

Install

It's part of the Krea2_ComfyUI_Advanced pack (Eric Hiss / EricRollei), self-contained with no dependency on other node packs:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Krea2_ComfyUI_Advanced

Restart ComfyUI, or find the pack in ComfyUI Manager. The pack needs the genuine Hugging Face Krea2Pipeline - install the diffusers build from source (--force-reinstall against the HF git repo) and have Krea 2 weights in diffusers layout, per the README.

Gotchas

The unload is immediate and un-doable - there's no "last applied" memory here. If you unload and realize you needed that LoRA, you're re-fusing it, not undoing. And it only clears what's on this pipeline dict: if a loader node is caching a separate pipeline instance, Unload LoRA won't touch it. If you're chasing VRAM rather than style contamination, you want Eric Krea2 Unload Models, which tears down the whole pipeline - this node frees the adapters, not the model.