Nunchaku FLUX LoRA Loader

If you loaded a Nunchaku 4-bit Flux model and then reached for your usual LoRA loader, stop - it won't behave. NunchakuFluxLoraLoader is the LoRA node built for Nunchaku's quantized Flux, and you want it instead of the stock one because the quantized model isn't a normal set of weights you can just patch.

The short version: a Nunchaku Flux model lives in 4-bit inside a custom engine. Applying a LoRA to that means Nunchaku has to fold the LoRA into its own representation, which is a different operation from ComfyUI's default "add the delta to the weights." Use the matching loader and it just works; use the wrong one and you get errors or a LoRA that does nothing.

How it works

You chain it right after NunchakuFluxDiTLoader: model in, model out, LoRA applied in between. Under the hood Nunchaku loads a standard Flux LoRA (the ordinary .safetensors you'd download from CivitAI) and applies it to the 4-bit model. Worth knowing why this is even a selling point: on GGUF, stacking LoRAs is the thing that tanks your speed, because every layer has to be dequantized, patched, then requantized each step. Nunchaku's approach folds the LoRA in without that per-step penalty, so you keep the fast generation you came for. Multi-LoRA support landed early in the pack's life and has been a headline feature since.

You can link several of these back to back to stack LoRAs, exactly like ComfyUI's native loaders. If you're running more than a couple, the NunchakuFluxLoraStack node in this same pack does the same job in one tidier box.

The inputs and outputs that matter

Three inputs, all straightforward:

• model (MODEL) - the output of your Nunchaku Flux loader (or the previous LoRA loader in a chain).
• lora_name - the LoRA file to apply. A normal Flux LoRA; nothing special about the file itself.
• lora_strength (default 1.0) - how hard it bites. 1.0 is full strength; drop toward 0.6–0.8 if a style LoRA is overcooking, push past 1 if it's too subtle. The range is enormous (-100 to 100) but you'll live between about 0 and 1.5.

Output is a MODEL with the LoRA baked in - feed it onward to the next LoRA loader or straight to your KSampler.

How to install it

It's part of the ComfyUI-nunchaku pack, so you get it when you install the pack. ComfyUI Manager → search "ComfyUI-nunchaku" → install, or

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

then restart. Remember the pack also needs its backend wheel (install it with NunchakuWheelInstaller) and a quantized Flux model from the nunchaku-ai HuggingFace org - this LoRA node only makes sense downstream of a Nunchaku Flux loader.

Common issues & troubleshooting

"My LoRA does nothing." Check you fed this node a Nunchaku model, not a stock Flux - and that you're using this loader, not ComfyUI's default LoRA loader, which doesn't understand the 4-bit model.

Errors on apply. Usually a mismatch: the LoRA was trained for a different Flux variant than the one you quantized against, or an old wheel that predates a LoRA-format change. Update the engine and confirm the LoRA is a genuine FLUX.1 LoRA.

Style is too strong or washed out. That's lora_strength, same as any LoRA - tune it. Flux LoRAs often want a touch under 1.0.

Stacking a lot of LoRAs. Switch to NunchakuFluxLoraStack - same behavior, up to 15 slots, far less spaghetti.