HSWQ SDXL LoRA Stack V3

The thing everyone wants once they have more than two LoRAs is a single node that applies several at once, with per-LoRA toggles, instead of a daisy chain of LoraLoader nodes that's a pain to rewire. rgthree's Power Lora Loader is the famous answer to that and it's great. This node is the same idea tuned for SDXL, shipping inside the HSWQ pack - which means if you're already running HSWQ's loaders, it's the zero-extra-install option.

What it does

HSWQSDXLLoraStackV3 applies multiple LoRAs to an SDXL model and its clip in a single pass, with a dynamically generated UI. The lora_count widget (default 3, range 1–10) controls how many slots the node exposes; each slot gets three inputs:

• enabled_N - toggle that slot on/off.
• lora_name_N - the LoRA file. Select None to skip a slot entirely.
• lora_strength_N - strength from −100 to 100, default 1.0. Negative strengths are legitimately useful for un-learning a style.

There's also toggle_all to flip every slot at once and a debug switch that prints per-LoRA application logs. The LoRA application itself goes through ComfyUI's own comfy.sd.load_lora_for_models, so this works with normal SDXL LoRAs targeting the UNet and/or CLIP - no custom LoRA format, no magic. Outputs are the modified MODEL and CLIP, ready to feed your sampler.

Why you'd reach for it

If you're inside the HSWQ SDXL world - quantized checkpoint, this pack's loader - a multi-LoRA node that lives in the same pack and is wired to tolerate the quantized paths is convenient in a way a generic stacker isn't. The README doesn't even document this node, which tells you it's a quiet utility rather than a headline feature, but the source is straightforward: dynamic inputs, per-slot enable/strength, load_lora_for_models under the hood, and an IS_CHANGED hash so the node re-runs when any slot's parameters change.

A couple of practical notes. First, with lora_count at the max you get 10 slots, each with three inputs - that's 30 widgets plus the base set, so a crowded workflow can get visually noisy. Set lora_count to the number you actually use rather than the max. Second, the optional lora_name_N inputs are exactly that - a slot with lora_name left at None is skipped cleanly, which is the intended way to keep fewer than the configured count. And per the LoRA-to-quantized-model lesson from the broader ecosystem, stacking many LoRAs on quantized weights is slower than on fp16 because each layer has to be dequantized, patched, and requantized - this node doesn't change that math.

Installing it

It ships in the HSWQ pack, so you already know the drill:

cd ComfyUI/custom_nodes
git clone https://github.com/ussoewwin/ComfyUI-HSWQ-Loader-and-Tools

Restart, and it appears under the HSWQ category. The pack's requirements install a lot of machinery (diffusers, insightface, onnxruntime…) that this particular node doesn't need - if you installed the pack for the loaders anyway, none of that matters. If you came only for a LoRA stacker and aren't running HSWQ models, the honest advice is to install rgthree or any other well-supported stacker instead; you'd be pulling a heavy quantization pack just for a convenience node.