Lora Stack apply

Lora Stack apply is the boring, important friend of Lora Stack: the stack node just collects your LoRA choices into a LORA_STACK object, and this node is the one that actually loads the files and patches your model and CLIP with them. If you grabbed the stack node and connected its output straight into a stock LoraLoader expecting it to work, you already know why this article exists - LORA_STACK is this pack's own type, and only this node knows what to do with it.

Mechanically it's not exotic. It takes a MODEL (required) and a CLIP (optional), reads the stack, and for each entry loads the LoRA file from your loras folder and calls ComfyUI's standard LoRA application - the same comfy.sd.load_lora_for_models machinery the stock LoraLoader uses. Outputs are the patched MODEL and CLIP, plus a lora_names string.

The lora_names output is a nice bonus

You might not expect the third output, and it's worth wiring somewhere. It emits a plain-text summary of what was applied in the standard <lora:name:model_weight:clip_weight> format - one string per LoRA, space-separated. That's exactly the format that shows up inside CivitAI metadata and prompt dumps, so it's a handy free way to record "these LoRAs were in this image" without re-typing them. In model_only mode the format drops the clip weight (<lora:name:model_weight>), since there isn't one.

The inputs that matter

• model - required, from your checkpoint/model loader.
• lora_stack - required, from Lora Stack [Eclipse].
• clip - optional, but if you want CLIP-patched LoRAs (most of them), connect it. The node is polite about it: a LoRA with a CLIP weight connected to a workflow with no CLIP just logs a warning and applies to the model only, instead of crashing.

The node also knows about Nunchaku quantized models (the fast FLUX/Qwen path this pack documents separately) - it detects those wrappers and routes them through the appropriate application path, so a stack that works on a normal FLUX model also works on the quantized one without you changing anything.

Install

Part of ComfyUI_Eclipse (r-vage). ComfyUI Manager, search ComfyUI_Eclipse, or:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_Eclipse

Restart ComfyUI. No model downloads here - it reads your existing loras folder. If you're pulling a shared workflow that references [Eclipse] nodes and it won't load, check the v4.0 legacy-node removal and run the pack's Workflow Migration Tool.

Common issues

"LoRA applied but image is identical" - check that you're actually using the MODEL output downstream, and that the LoRA name in the stack still points at a real file (the stack node deliberately skips combo validation so stale names load but fail here). "Works with a stock LoraLoader but not here" - usually means the stack has entries with "None" as the name, which apply is told to skip; or you forgot the CLIP connection and the LoRA's CLIP half silently didn't apply. When in doubt, look at lora_names - if the entry you expect isn't in the string, it wasn't applied, and now you know exactly which one.