SDNQ LoRA Stacker V2

This is a narrower tool than it looks, and the narrowness is the point: it's a LoRA stacker built specifically for models loaded through SDNQ, a quantization format that's still genuinely new in the ecosystem. If that name means nothing to you, don't feel behind - SDNQ (an experimental uint4-style quantization) only recently started showing up in the community, most visibly as an option for Z-Image Base, and it sits alongside FP8, GGUF and NVFP4 as one of the newer ways to squeeze large models onto smaller GPUs without the quality collapse naive 4-bit quantization usually causes. It's not yet as established as GGUF's Q-ladder, so expect rougher edges and a smaller ecosystem of compatible tooling around it - this node is one of the few pieces of that tooling that exists.

This node's own README is explicit about the dependency: it's "designed for use with comfyui-sdnq-splited" (also by this pack's author), and the required model input has to come specifically from "SDNQ Model Loader (DiffusionPipeline)." Feed it a model from anywhere else - a stock checkpoint loader, a Nunchaku FLUX loader, a GGUF loader - and it won't have the right kind of state dict to patch correctly. If you're not already running SDNQ models, this node has nothing to do for you; grab LoraStackerV2_10 or V3 instead for a general MODEL+CLIP pipeline.

Once you are on SDNQ, the interface will look immediately familiar if you've used any other loader in this pack: up to ten optional slots, lora_name_1 through lora_name_10 for the filename (default None, meaning skip) and lora_wt_1 through lora_wt_10 for strength (default 1.0, range -100 to 100, step 0.001). Set only the slots you're actually using; unfilled ones are simply ignored, no validation errors. The single output is MODEL - note there's no CLIP output here, same pattern as this pack's Nunchaku FLUX loaders, so whatever text-encoding path you're running stays untouched by this node.

Installing it: this node itself is bundled with the whole pack - through ComfyUI Manager, search "ComfyUI-NunchakuFluxLoraStacker," or:

cd ComfyUI/custom_nodes
git clone https://github.com/ussoewwin/ComfyUI-NunchakuFluxLoraStacker.git

Restart ComfyUI. But that alone doesn't get you a working SDNQ pipeline - you also need comfyui-sdnq-splited installed and an SDNQ-quantized model loaded through its own loader before this node has anything valid to patch. Check that pack's own instructions for getting the model side set up; this node is downstream of it.

Being honest about troubleshooting here: SDNQ is new enough, and this integration specific enough, that there isn't an established community troubleshooting pattern to draw on the way there is for GGUF or fp8 quirks. The one thing worth flagging plainly: because this is early-stage quantization tooling, LoRA support in this space (SDNQ generally, not just this node) has been described as still being finished as of mid-2026 - if a LoRA loads without error but produces no visible effect, that's as likely to be a maturity gap in the SDNQ+LoRA combination itself as it is a misconfiguration on your end. Worth testing the same LoRA on a non-quantized or GGUF version of the same base model to confirm the LoRA file itself is good before chasing this node as the culprit.