NO8D-LoRA stack

The stock way to apply a LoRA in ComfyUI is a chain: Load LoRA node after Load LoRA node, each one taking the previous model output, each one asking for a CLIP you may not even need. Run three LoRAs and your graph is a ladder. NO8D-LoRA stack collapses that into one node: it's a UNET loader with a LoRA list built in, so you pick the base model and then add, remove, enable, disable, and reorder LoRAs without growing the graph.

The notable design choice is right in the name - there's no CLIP input. This node only patches the UNET (the diffusion model), not the text encoder. That's the right call for the modern architectures: on Flux, SD3, Qwen, and Krea 2 the interesting LoRAs are trained on the diffusion model, and the CLIP patch is either irrelevant or actively unhelpful when you're using a separate text encoder. So the workflow is: load your UNET here, stack LoRAs, and encode your prompt wherever that base model's encoder lives.

What you set

• unet_name and weight_dtype - these are inherited wholesale from ComfyUI's native UNETLoader, and they follow upstream updates. Same model list, same fp8 options (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2).
• lora_picker - a dropdown of your installed LoRAs. This is the "currently adding" slot; pick one and it joins the stack.
• stack_json - the stack itself, as JSON. You won't type this by hand; the node's frontend edits it when you use the add/remove/reorder controls. New entries start at strength 1.0 with a 0–2 slider range.

Outputs are the two that matter: model (the patched model, wire it to your sampler) and trigger_words. That second one is quietly great - it merges the trigger words from every enabled LoRA into a single comma-separated string, so you don't have to remember to type ch9ractername, styl3name into your prompt. The node keeps them deduplicated and skips disabled or zero-weight entries.

How it works under the hood

The stack is applied in order, top to bottom, by loading each LoRA and patching the model once per entry. The caching is worth noting: if the same LoRA file appears twice, it's loaded from disk once and reused. There's also a smart IS_CHANGED implementation - the node only re-runs when something that affects the result actually changed (model choice, enabled entries, weights), so it won't gratuitously invalidate the graph downstream.

The honest caveats

Stacking LoRAs is still stacking LoRAs. Three at full strength will fight each other, and the node's per-entry toggles exist precisely so you can A/B which combination is actually helping. The usual rule applies: most LoRAs want to sit well below 1.0 when combined. And because this node loads the UNET itself, it replaces your model loader - if your workflow expects a checkpoint node with baked-in CLIP, you'll be restructuring that part of the graph.

Install

Same as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/no8d/ComfyUI-NO8D-controls.git
python -m pip install -r requirements.txt

ComfyUI Manager → search "NO8D-controls". No models to download, no heavy deps. If the LoRA dropdown comes up empty, check that your LoRAs are in ComfyUI/models/loras where every other loader finds them.