LoRA Stacker V2

Despite the pack's Nunchaku-heavy branding, this particular node has nothing to do with quantization - it's a general-purpose LoRA stacker that works on any checkpoint using ComfyUI's standard MODEL + CLIP pair: SDXL, plain FLUX, WAN 2.2, whatever you're running. If the other Nunchaku-specific loaders in this pack don't apply to your setup, this is probably the one you actually want.

The mechanism is the least interesting and most reliable part: it calls ComfyUI's own load_lora_for_models under the hood, the same function core nodes use, so there's no custom patching logic to second-guess - it behaves exactly like chaining ten stock LoRA loaders, just in one node with a compact UI instead of ten nodes cluttering your graph.

Two required inputs, model and clip, straight from your checkpoint loader - this node doesn't care which loader produced them, unlike the FLUX-specific nodes elsewhere in this pack that insist on a Nunchaku DiT loader. From there, ten optional slot pairs: lora_name_1 through lora_name_10 pick the file (leave a slot on None to skip it, no error thrown), and lora_wt_1 through lora_wt_10 set that slot's strength, defaulting to 1.0 with a -100 to 100 range. Note there's a single weight per slot here - model and CLIP strength move together, they're not split like the advanced mode on this pack's legacy FLUX loader. If you need independent model/CLIP weighting, that's not what this node is for.

A couple of things worth knowing about LoRA behavior generally, since they apply here as much as anywhere: interactions between stacked LoRAs are genuinely unpredictable, so when something looks wrong, isolate - test with one slot active at a time on a fixed seed before blaming the stack. If you hit black images or NaN-looking garbage output with multiple LoRAs loaded, try reordering so the larger/heavier LoRAs sit in earlier slots. And the default weight of 1.0 that feels "safe" is really SDXL-era folk wisdom - it doesn't generalize; plenty of LoRAs on newer architectures (character LoKrs on Z-Image, for instance) are trained expecting weight 1.0-1.5, and dropping it can cost you likeness rather than just softening the effect. Check what the LoRA's own page says instead of reaching for a blanket 0.7 by habit.

Outputs are MODEL and CLIP, both patched, ready to feed into your sampler and text-encoding chain respectively - wire both, since leaving CLIP unconnected downstream means any LoRA that does touch text conditioning silently does nothing on that side.

Installing it: 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, and look under loaders.

Where people get stuck: base-model mismatch is the number one cause of "the LoRA did nothing" - an SDXL LoRA doesn't work on FLUX and vice versa, and even within the SDXL family, Pony/Illustrious/vanilla-SDXL LoRAs are often not cross-compatible despite sharing enough architecture to show up as selectable. Second most common: a missing trigger word - plenty of LoRAs need a specific phrase in your prompt to activate, and it's easy to load the file, set a strength, and wonder why nothing changed when the actual issue is the prompt.