Nunchaku FLUX LoRA Stack

Chaining fifteen separate LoRA loader nodes is a graph nobody wants to look at. NunchakuFluxLoraStack collapses all of it into one node: one MODEL in, one MODEL out, and up to 15 LoRA slots in between, each with its own strength. It's the same job as NunchakuFluxLoraLoader, just done in bulk and without the wiring spaghetti.

The node's own description says it plainly - it's "equivalent to chaining multiple LoRA nodes but more convenient for managing many LoRAs." If you run one or two LoRAs, the single loader is fine. The moment you're juggling a realism LoRA plus a style LoRA plus a detail LoRA plus a character LoRA, this is the one you reach for.

How it works

It's a Nunchaku-native LoRA applier, so it belongs on the 4-bit path: put it after NunchakuFluxDiTLoader and it folds each active LoRA into the quantized model. That "folds in" bit is the point - on GGUF, stacking LoRAs murders your speed because every layer gets dequantized and requantized each step; Nunchaku applies them without that per-step tax, which is exactly why cramming several into one model is even practical here.

Each slot is a name/strength pair. Set a slot's name to None and it's skipped, so you don't have to fill all fifteen - just use the top few and leave the rest empty. It walks the slots in order and applies whichever ones point at a real LoRA.

The inputs and outputs that matter

• model (MODEL) - from your Nunchaku Flux loader. Required.
• lora_name_1 … lora_name_15 - each a dropdown of your Flux LoRAs, defaulting to None. Fill the ones you want; leave the rest on None.
• lora_strength_1 … lora_strength_15 - the matching strength for each (default 1.0). Standard Flux LoRA tuning: 1.0 full, ease toward 0.6–0.8 if something dominates.

Output is a single MODEL with everything applied, ready for your KSampler.

One honest caveat that isn't Nunchaku-specific: stacking many LoRAs is a quality gamble on any model. Two strong LoRAs can fight each other and muddy the output. Start with a couple at moderate strength and add from there rather than maxing out all fifteen and wondering why it looks off.

How to install it

Ships with the ComfyUI-nunchaku pack. ComfyUI Manager → search "ComfyUI-nunchaku" → install, or

cd ComfyUI/custom_nodes
git clone https://github.com/mit-han-lab/ComfyUI-nunchaku

then restart. As with everything in this pack, you also need the backend wheel (via NunchakuWheelInstaller) and a quantized Flux model from nunchaku-ai. This node only does anything downstream of a Nunchaku Flux loader.

Common issues & troubleshooting

Nothing happens. Most often an unfilled stack - every slot still on None. Point at least one slot at a real LoRA. Also confirm the upstream model is a Nunchaku 4-bit Flux, not a stock checkpoint.

One LoRA in the stack errors. Isolate it: set the others to None and test that one alone. Usually it's a LoRA trained for a different Flux variant, or a broken file.

Output looks over-processed. Too many LoRAs, or strengths too high. Drop the weakest-value LoRAs to None and pull the rest down; a stack of four at 0.7 often beats eight at 1.0.

Slots feel redundant with the single loader. They are, functionally. Use the stack for many LoRAs, the single NunchakuFluxLoraLoader when you only need one and want the graph obvious.